Kunchakara Raja Sekhar
Shaiku Shahida Saheb- VIT-AP School of Business, VIT-AP University, Amaravati, AP, India
Introduction: Digital lending, also known as alternative lending, refers to fintech platforms that offer quick and easy loans through digital channels, bypassing many of the limitations of traditional banking. Since the mid-2000s, digital lending has become a major fintech innovation, with rapid growth in India driven by financial inclusion measures. However, the sector continues to face challenges, including fraud, transparency issues, and consumer dissatisfaction. The primary objective of this study was to understand how consumers perceive and assess India’s RBI-approved P2P digital lending apps by analyzing a large dataset of customer feedback to identify strengths, weaknesses, and overall satisfaction levels.
Methods: The study analyzed a final dataset of 15,408 user reviews collected from seven RBI-approved digital lending platforms: 5Paisa, Faircent, i2iFunding, LenDenClub, CashKumar, Lendbox, and IndiaMoneyMart derived from an initial 15,537 reviews. The cleaned data was then examined using natural language processing, topic modeling, and supervised machine learning and deep learning models to identify key themes and evaluate predictive performance.
Results: Topic modeling identified 11 recurring topics. Sentiment analysis revealed that 55% of evaluations were positive, 41% were negative, and 4% were neutral. Strengths included loan disbursement, withdrawals, and EMI payments, while weaknesses involved interface design, transparency around rejections, and login functionality. Comparative data revealed that IndiaMoneyMart and i2iFunding received the highest user satisfaction, while 5Paisa and Lendbox trailed due to recurring complaints about transparency, accessibility, and overall user experience. In terms of modeling, the deep learning model VGG16 and ensemble machine learning techniques (XGBoost, CatBoost, and LightGBM) consistently achieved the highest predictive accuracy (up to 0.88), outperforming simpler models such as decision trees and ResNets.
Discussion: The findings indicate that digital lending platforms support financial inclusion but require improvements in user interface and user experience, better transparency in loan decisions, and stronger customer support. Addressing these areas can help strengthen trust and promote long term adoption of digital lending services.
1 Introduction
Digital lending, also known as alternative lending, provides affordable, easily accessible loans through online platforms. Zopa, the first such platform, debuted in the UK in March 2005 (Pang et al., 2022). In the early 2010s, digital lending gained popularity alongside the explosive expansion of fintech. Between 2012 and 2020, digital lending platforms experienced significant growth, offering additional credit access options (Cevik, 2024). During this period, sophisticated financial services powered by technology emerged, swiftly revolutionizing the conventional banking system (Anifa et al., 2022). Currently, numerous platforms offer commercial lending services, such as Easy Credit, PPDai, Lending Club, Zopa, and Prosper (Chen et al., 2014). The digital lending industry has expanded considerably over the past few years (Sarungu, 2020). It is regarded as one of the most transformative fintech innovations, reshaping traditional banking and increasing access to credit (Modi and Kesarani, 2023). This rapid growth has been driven by the adoption of new technology and shifting customer demands, making borrowing and lending faster, easier, simpler, and more convenient through digital channels (Modi and Kesarani, 2023). Moreover, fintech, philanthropy, development, and the monetization of digital footprints drive the expansion of digital lending (Gabor and Brooks, 2017). These lending platforms significantly improve financial inclusion and access to capital, particularly in emerging nations (Zetzsche et al., 2017). In countries such as India, fintech has transformed the way consumers engage with financial institutions. Since 2014, India’s digital financial industry has witnessed rapid growth. Government-led initiatives such as demonetization have encouraged people to shift from cash to digital transactions (Ghosh and Hom Chaudhury, 2022). Furthermore, the implementation of the Pradhan Mantri Jan Dhan Yojana in August 2014 has significantly improved financial inclusion by simplifying the process of opening bank deposit accounts (Barik and Sharma, 2019).
However, fraud and identity theft remain significant challenges in digital lending, as dishonest actors exploit personal data for illicit activities and financial gain (Saunders and Zucker, 1999). These crimes have increased due to the expansion of internet services and digital transactions, particularly during and after the COVID-19 pandemic (Luong and Ngo, 2024). To mitigate these risks, additional security measures such as multi-factor authentication, secure payment gateways, and encryption are implemented (Ogunola et al., 2024). In India, the Reserve Bank of India (RBI) regulates digital lending platforms through its 2017 Master Directions (Khan et al., 2024), under which Faircent.com emerged as the country’s first peer-to-peer lending platform (Khatri, 2019). As fintech continues to evolve, technologies such as blockchain and smart contracts have become key tools for enabling secure, transparent, and sustainable digital financing solutions (Elias et al., 2024). It enhances transaction security, transparency, and trust, while smart contracts enable automated, secure enforcement of lending arrangements (Omowole et al., 2024). With loan origination, repayment, and collateral management, distributed ledger systems ensure immutable, transparent transactions, reducing fraud and enhancing trust (Rijanto, 2021). These technologies are particularly useful in the context of green finance and microfinance, where accountability and transparency are vital (Elias et al., 2024; Omowole et al., 2024). To handle the increasing complexity of lending applications and risks associated with P2P platforms, policymakers are using machine learning technologies to process vast amounts of data to inform regulatory and financial decisions (Xu et al., 2021). These technologies improve the security and efficiency of lending processes while also promoting environmental sustainability by incorporating ecological considerations into financial decision-making (Addy et al., 2024).
2 Literature review
The banking industry has been revolutionized by digital lending, which uses technology to streamline loan processes and increase credit availability (Mallinguh and Wasike, 2025). In recent years, the digital lending industry has grown rapidly (Sarungu, 2020). The growing popularity of online loans each year is driven by the expansion of the Internet and the rise of big data (Liu, 2025). Several commercial lending platforms have been made available, including Prosper, PPDai, Lending Club, Zopa, and Easy Credit (Chen et al., 2014). As per the Reserve Bank of India (RBI), 27 NBFC-P2P platforms have been registered under the 2023 Directions. These revolutionary force in the fintech industry, competing with traditional lending strategies. The integration of advanced technologies such as Big Data, AI, and Machine Learning (ML) has revolutionized credit assessment and risk management processes (Aldboush and Ferdous, 2023; Bazarbash, 2019). These technologies empower fintech firms to analyze vast datasets, including non-traditional sources (Bazarbash, 2019; Warin and Stojkov, 2021).
Digital lending apps have become extremely popular in recent years. Studies are now applying machine learning to study customer reviews and opinions, with these models helping to sort and interpret user feedback more effectively (Alawaji, 2025). While positive views emphasize its role in financial inclusion and efficiency, negative perspectives focus on ongoing challenges of risk, regulation, and disclosure (Maulida and Surbakti, 2024). However, borrowers’ perceived trust had a small impact on the incentive to use the P2P lending network (Gupta and Mahajan, 2023). According to reviews, borrowers were largely delighted with the platform’s services and speedy loan processing, but convenience of use, cost, and risk were less important. Both lenders and borrowers experienced problems (Gupta and Mahajan, 2023). Meanwhile, the study helps borrowers and lenders select appropriate applications and enables P2P platforms to assess their strengths and weaknesses. The study focuses on user-generated information, particularly online reviews, to examine which service features people evaluate and how these characteristics predict a consumer’s recommendation (Siering et al., 2018).
To achieve greater predictive power and flexibility, interpretable ML models autonomously identify internal structure and correlations, thereby challenging conventional statistical methods (Huang et al., 2004). Big data risk management systems have progressed from the fundamentals of machine learning approaches to more advanced deep learning techniques (Bao et al., 2024). Among the models used to predict loan default, logistic regression is among the most frequently used to estimate the probability of successful loan funding on peer-to-peer lending platforms (Puro et al., 2010). Another highly regarded method is the Support Vector Machine (SVM), a sophisticated machine learning technique based on statistical learning theory, known for delivering consistently high performance across a wide range of applications (Huang et al., 2004). Among classification and prediction tools in machine learning, decision trees are widely used for their simplicity and effectiveness. Studies have shown that machine learning models can effectively evaluate personal credit information and predict the likelihood of loan default. Of these models, the Deep Neural Network achieved the best accuracy of 0.94 (Liu, 2025). Similarly, a study using Naïve Bayes algorithms achieved 94% accuracy and identified several key factors influencing loan success, including interest rate, repayment time, loan description, credit grade, loan history, gender, and credit score (Vedala and Kumar, 2012). Meanwhile, another study employed multivariate logistic regression to predict both prepayment and default risks, two critical events associated with loan termination and creditor profit loss, achieving an overall model accuracy of 76.63%. Additionally, a separate study found that using LightGBM to forecast default risk on digital lending platforms could enhance lending clubs’ revenues, achieving a prediction accuracy of 68% (Ko et al., 2022).
The application of ML algorithms, including neural networks and ensemble models, has significantly improved the accuracy and efficiency of financial decision-making (Odei-appiah and Adjei, 2021). However, the widespread adoption of these advanced technologies also introduces several ethical, social, and regulatory challenges. These concerns include algorithmic bias, discrimination, lack of transparency, and potential violations of data privacy (Aldboush and Ferdous, 2023). Issues could result in digital financial exclusion, particularly through algorithmic redlining, in which automated systems deny credit based on proxy variables that correlate with race, income, or geographic location (Bazarbash, 2019). Additionally, the expansion of digital lending into riskier, less-regulated segments of the financial system has posed ongoing challenges for regulators and policymakers. Previous studies on P2P lending and crowdfunding have also explored how user sentiment and comments affect funding performance, interest rates, and default probabilities. According to certain studies, the default likelihood and cost of capital are only adversely impacted by favorable improvements in media and social media for P2P lending platforms (Wang et al., 2020). In conclusion, although digital lending powered by AI and ML offers substantial promise to enhance access to credit and operational efficiency, it also requires careful consideration of ethical, legal, and systemic risks. To ensure responsible innovation, maintain financial stability, and safeguard consumer rights, the development of robust regulatory frameworks and effective oversight mechanisms has been emphasized (Cevik, 2024) (Table 1).
Table 1. Model comparison with earlier research.
Previous research on peer-to-peer lending and crowdfunding has examined how user comments and sentiments affect factors such as fundraising success, interest rates, and default rates (Gupta and Mahajan, 2023). Khan et al. (2024) analyzed Google Play Store reviews to assess user perceptions of P2P lending systems. They found that customers prioritize speedy loan approvals, transparency, and responsive services as the key drivers of satisfaction. The majority of earlier P2P lending research was platform-specific, employed lexicon-based or simple machine learning techniques, and neither integrated deep learning nor conducted extensive comparisons across regulated apps (Niu et al., 2020). Text mining and sentiment analysis have also been utilized in several studies to explore how users interact with financial technologies. A text mining analysis revealed that reliability, usability, and security are the most critical factors in determining user satisfaction with P2P payment services. This strategy aligns with the current study’s emphasis on eliciting user perceptions to better understand how Indians perceive digital lending apps (Perea-Khalifi et al., 2024). There aren’t many studies that connect sentiment findings to behavioral theories, such as trust-risk frameworks or TAM. This leaves a research void for a thorough, theory-driven investigation that assesses user attitudes across several P2P lending platforms regulated by the RBI, utilizing topic modeling and sophisticated ML-DL models.
This study is the first to use extensive user-generated data to compare seven P2P lending apps in India that are regulated by the RBI. It combines hybrid ML–DL models, topic modeling, and sophisticated embeddings to increase sentiment prediction accuracy and offer empirical insights into satisfaction, usability, and trust. This study demonstrates how perceived simplicity, utility, and transparency impact the adoption of digital lending by integrating data-driven analysis with TAM, UTAUT, and trust-risk theories. This sets a standard for future fintech sentiment research and provides useful advice for improving platform performance and user trust.
2.1 Theoretical foundation
Theoretical frameworks such as the Technology Acceptance Model (TAM), the Unified Theory of Acceptance and Use of Technology (UTAUT), and Consumer Decision-Making Process theories (Chen and Zhou, 2016), along with the SERVQUAL Model and Trust Risk Theory, have been widely used to explain the adoption of digital financial services and lending platforms (Abu-taieh et al., 2022). While these models effectively describe user acceptance of new technologies, they often overlook post-adoption outcomes, such as borrower satisfaction, changes in financial behavior, and sustainable credit practices, which would provide stronger theoretical support (Singh, 2020). There is growing recognition of the need for studies that prioritize borrowers and examine the broader socio-economic and psychological impacts of digital borrowing, which remain underexplored in culturally diverse and economically stratified contexts (Wang et al., 2020). The Technology Acceptance Model (TAM) was proposed by Davis (1989) to explain how users adopt and continue using FinTech applications. It suggests that technology adoption is primarily influenced by two perceptions: perceived usefulness and perceived ease of use (Sylvie and Pascal, 2021). Similarly, the UTAUT framework is applied to interpret users’ sentiments toward FinTech apps, in which constructs such as performance expectancy and effort expectancy are reflected in users’ perceptions of app usefulness and ease of use (Venkatesh et al., 2019). In this context, trust and satisfaction further influence users’ behavioral intentions to continue using the platform. The SERVQUAL Model extends this understanding by emphasizing that service quality drives satisfaction and loyalty across five dimensions: reliability, responsiveness, assurance, empathy, and tangibles (Jain and Gupta, 2004). In online lending, users’ willingness to engage also depends on perceived trust, security, and reduced uncertainty, as explained by Trust-Risk theory. Prior studies have employed TAM to explain consumer behavior and their propensity to embrace technological improvements (Baron et al., 2006). In the context of digital lending and FinTech apps, these constructs expanded to include trust and transparency, which play a critical role in shaping user confidence and satisfaction. Users who perceive a platform as transparent and reliable are more likely to trust it, leading to greater satisfaction and continued use (Yadav, 2024). Thus, these theories collectively provide a strong foundation for linking user perceptions, trust, and satisfaction with technology adoption behavior observed in online reviews. This framework guides the interpretation of sentiment analysis results and explains how positive user experiences translate into greater acceptance of FinTech services (Table 2).
Table 2. Mapping of P2P App Review Topics to Theoretical Constructs.
Although this study employs machine learning and deep learning models for sentiment analysis, the theoretical interpretation is grounded in the UTAUT model, which highlights how factors such as trust, transparency, and perceived ease of use influence users’ intentions to adopt FinTech applications.
3 Objectives and methodology
3.1 Objectives
1. To analyze user-generated reviews of regulated P2P lending applications in India using text mining and sentiment analysis.
2. To conduct a comparative analysis of digital lending applications by applying machine and deep learning models, aiming to evaluate app performance, borrower satisfaction, and trustworthiness.
3. To provide a theory-driven interpretation of user attitudes and behavioral intentions toward digital lending platforms (Figure 1).
Figure 1. Proposed methodology.
3.2 Materials and methods
This study employs machine learning (ML), deep learning (DL), and natural language processing (NLP) techniques to analyze user feedback from digital lending apps approved by the RBI. Seven popular peer-to-peer lending platforms were examined using purposive selection between January 2020 and June 2024. The selection criteria included an active presence in the Google Play Store and an adequate volume of user reviews. Reviews were collected through a Python script built with the Google Play Scraper API. Extraction relied on specific application package IDs rather than keyword searches. The analyzed apps were 5Paisa (com.a5paisa.trade), Faircent (com.faircent.app), I2IFunding (com.i2ifunding.app), LendenClub (in.infra.lendenclub), CashKumar (com.cashkumar.loan), Lendbox (com.lendbox.app), and IndiaMoneyMart. These platforms were chosen for their accuracy in text mining and sentiment analysis, their relevance to the Indian digital lending sector, and their regulatory approval under the RBI’s NBFC-P2P framework. The collected fields included review ID, user name, score, review content, timestamp, and app metadata. Because the Google Play Store aggregates feedback across multiple releases, the dataset includes reviews from various app versions. Focusing on these platforms ensured data consistency and meaningful comparability between applications. A total of 15,537 reviews were scraped from the Google Play Store using a Python-based scraping technique. To prepare the raw textual data for analysis, preprocessing techniques such as noise reduction, lowercasing, tokenization, and stopword removal were applied.
Furthermore, topic modeling techniques, such as latent Dirichlet allocation (LDA) and word-frequency analysis, were applied to identify recurring themes. Multiple topic configurations (K = 5–20) were tested for the LDA model, with the final number of topics chosen based on the highest coherence score and consistent perplexity behavior. LDA preprocessing included lemmatization, removal of custom stop words, and bigram generation with a minimum document frequency threshold (min_df = 5). To improve subject separation, highly unusual and unduly frequent terms were filtered out. These stages ensured that the final subjects were stable, understandable, and data-driven rather than subjective. Feature extraction used traditional vectorization techniques, such as Bag-of-Words (BOW), TF-IDF, and hashing, as well as advanced word-embedding models, including Word2Vec, FastText, GloVe, and IndicBERT. After preprocessing, the cleaned dataset of 15,408 reviews was split into training and testing sets using an 80:20 ratio, yielding 12,326 and 3,082 samples, respectively (Figure 2).
Figure 2. End-to-end review data flow.
To perform classification tasks, several machine learning models were employed, including logistic regression, SVM, random forest, XGBoost, and decision trees CatBoost, AdaBoost, LightGBM; deep learning models such as ResNet, BiLSTM and VGG16 were also investigated. VGG16 and ResNet were not applied to raw text or text converted to graphics. Instead, they were employed in a transfer-learning setup using dense word embeddings generated by Bag-of-Words, TF-IDF, Word2Vec, FastText, GloVe, and IndicBERT. These embeddings produce structured numerical matrices that convolutional layers can process, a method backed by previous NLP research on embedding grids (Li et al., 2020). With limited labeled data, using pretrained VGG16 and ResNet models enables faster feature extraction and improved generalization. This method avoids converting words to images; instead, each review is represented as a reshaped embedding tensor suited for convolution. Including these architectures enables comparisons between CNN-based transfer learning and NLP-specific models, as well as an evaluation of whether pretrained convolutional networks offer useful hierarchical features for text classification (Wang and Li, 2022). Ultimately, the model’s effectiveness was assessed using recall, accuracy, precision, and F1 score, ensuring the effective categorization of user review sentiments and theme analysis.
3.3 Data description
Our dataset included Google Play Store reviews of seven Indian RBI-approved lending platforms: 5Paisa, Faircent, i2ifunding, LenDenClub, CashKumar, Lendbox, and IndiaMoenyMart. The Google Play Store was the largest online market for mobile apps with over 2.6 million free and premium apps available as of May 2025 (Pamplona, 2022). Accordingly, we gathered this data using a site-scraping technique. The data file includes the user’s name, time and date, thumbs-up, written comments, and a rating on a scale of 1 to 5. Based on this, we divided the rating into three categories: positive (three or more), negative (below two), and neutral (more than two) (Pamplona, 2022; Wang, 2015) (Table 3).
Table 3. Platform-wise compilation of collected user reviews.
3.4 Data pre-processing
Preprocessing is an essential stage in machine learning and deep learning workflows, especially when handling unstructured user review data. This study scraped reviews from the Google Play Store using Python tools, capturing details such as username, timestamp, star rating (1–5), number of likes, and written comments. The raw text was cleaned by removing non-ASCII characters, HTML tags, URLs, emojis, punctuation, special symbols (@, #, %, etc.), and digits. All text was then converted to lowercase for consistency. Tokenization was applied to split sentences into individual words (e.g., “Loan process is fast” → [“loan,” “process,” “is,” and “fast”]). Common stop words like “the,” “is,” and “in” were removed using a tool, NLTK Lemmatization, to downsize words to their most basic form (e.g., “running” → “run”). In this study, sentiment labels were determined directly from user star ratings rather than through manual annotation or lexicon-based sentiment approaches. This approach has been widely used in previous studies (Thelwall et al., 2010; Pagolu and Majhi, 2016) because the numerical star rating is an explicit indication of user sentiment provided at the time of review submission. It provides a more objective reflection of the user’s overall contentment or dissatisfaction than textual polarity, which can be unclear or influenced by linguistic variances. Using rating-based sentiment tagging reduces subjectivity and ensures consistency across the dataset, preventing misclassifications caused by sarcasm, mixed viewpoints, or casual language in the review text. Reviews were categorized as positive (>3), neutral (=3), or negative (<3) based on ratings. A final manual review ensured the cleaned data preserved the intent and quality of the original content. To ensure data quality and consistency, the dataset was preprocessed multiple times. Missing or null reviews were removed, content was standardized to lowercase, and undesirable components such as URLs, punctuation, numerals, emojis, and special characters were removed. Stopwords were filtered using the NLTK package, and words were reduced to their base form using lemmatization. To prevent redundancy and bias in model training, duplicate reviews identified by identical text content were removed using the Python drop_duplicates () method.
3.5 Operationalization of theoretical constructs
This study includes TAM, UTAUT, SERVQUAL, and the Trust-Risk paradigm into the analytical design by transforming their basic conceptions into measurable text features. Perceived usefulness, ease of use, trust, responsiveness, reliability, and perceived risk were predefined as aspects expected in user evaluations and operationalized using topic-model probabilities, keyword clusters, and sentiment-based expressions. These constructs were extracted using two methods: topic modeling, which offered numerical indicators of construct salience, and supervised keyword dictionaries, which capture explicit occurrences of theory-related ideas. The generated variables were utilized as input features in machine learning models to investigate how each component correlates with sentiment and ratings. Random forest, XGBoost, and deep learning networks are examples of multivariate classifiers that incorporate topic weights and lexicon counts. Model interpretability methods, such as feature importance and SHAP analysis, quantified each construct’s impact, allowing theory-driven relationships to be validated within the ML framework rather than interpreted retrospectively.
4 Results
4.1 Analysis of words
Word analysis is a fundamental technique in text mining and natural language processing, enabling the extraction of insights from unstructured text. By analyzing word frequency and distribution, it is possible to find significant terms, recurring motifs, and underlying issues (Feldman et al., 2007). This method exposes patterns and linguistic trends that would otherwise be overlooked in manual study (Liu, 2020). Furthermore, when used alongside visualization tools such as word clouds or frequency histograms, word analysis can help evaluate public opinion, customer feedback, or social media content (Weiss et al., 2005) (Figure 3).
Figure 3. Most commonly used words in the review dataset.
The top 20 terms used in user evaluations of the chosen apps are displayed in the bar chart. The top three terms on the list are “app” (6,800), “good” (3,800), and “loan” (3,200), indicating that users regularly leave comments on the application as a whole, offer compliments, and discuss loan-related aspects. Terms that convey a generally positive attitude and contentment with the app’s services include “nice” (1,400), “money” (1,300), “application” (1,200), and “easy” (730). Conversely, words such as “bad” (780), “fake” (1,000), “worst” (950), and “do not” (1,700) indicate user apprehension and unfavorable experiences. Additionally, references to “service” (900), “customer” (710), and “team” (700) reflect opinions about customer service and support interactions.
4.2 Topic modeling
Topic models are methods for identifying hidden themes in text, with Latent Dirichlet Allocation among the most widely used topic modeling approaches. Previous studies have demonstrated that topic models outperform more traditional clustering-based methods (Wei and Croft, 2006). Topic modeling is a useful approach for identifying textual groupings in huge collections (Liu, 2013). LDA, a widely used topic modeling algorithm, is effective in revealing latent semantic structures in text (Hofmann, 1999). Therefore, this study follows a research design that begins by preprocessing the raw textual data, then applies LDA to extract latent patterns, and finally integrates the LDA results with existing metadata for further analysis and visualization. Moreover, using the Gensim package in Python, we identified the following 11 topics (Table 4).
Table 4. Summary of extracted LDA topics with key terms and illustrative examples.
4.2.1 Topic frequency overview
The significance of the subjects fluctuated over the dataset. Document verification, payback concerns, and the overall application experience were the most frequently mentioned topics. Loan processing, loan rejections, login difficulties, and interface issues occurred regularly.
4.2.2 Topic quality and diagnostics
After analyzing various topic sizes, the final LDA model developed a coherent structure. The chosen model achieved a high coherence score, indicating well-defined and interpretable motifs. Review excerpts with the greatest topic likelihood were personally reviewed to ensure topic consistency, and noisy or overlapping clusters were reduced using preprocessing techniques such as bigram generation and frequency-based token filtering.
4.2.3 Platform-specific salience
Topic distribution varies across platforms. Loan denials, login issues, and customer service concerns were more common at Lendbox, CashKumar, and Faircent. Document verification, repayment, and withdrawal-related discussions were more prevalent on IndiaMoneyMart and i2iFunding. Topics such as interface and application experience were discussed across all platforms, though 5Paisa and LendenClub received the most attention. Platform-specific processes and operational practices impact user concerns and experiences, as demonstrated by these distinctions.
4.3 Evaluation of sentiment
A computational technique called sentiment analysis is used to identify and categorize textual opinions about a person, an event, or a product as neutral (=3), negative (<3), or positive (>3) (Alam and Yao, 2019). In this study, we applied sentiment analysis methods to classify users’ online text comments as either good or negative accounts of their experiences with lending apps. Sentiment labels were assigned based on star ratings, as they provide a clear and consistent representation of customer satisfaction. Text-rating comparisons revealed strong alignment, with positive terms appearing in 4-5-star reviews and complaints dominating in 1-2-star reviews. A human-labeled subset supported this pattern, and inter-rater reliability (Cohen’s kappa) confirmed consistent judgment among annotators (Al-Natour and Turetken, 2020). Rating-based labeling also matched the polarity created by VADER and transformer models, particularly for obvious positive and negative examples. Sensitivity tests utilizing stricter criteria (e.g., ≥4 for positive) yielded consistent results, indicating the reliability of rating-based sentiment classifications (Noori, 2021). To evaluate user feedback, we use sentiment analysis to assess reviews; it is superior to traditional techniques (Greaves et al., 2013). Platform comparisons are descriptive and based purely on the dataset’s observed distribution of feelings and reviews. Because the study is based on user-generated reviews, the differences across apps should not be considered statistically significant. These results reflect patterns observed in accessible reviews, and variances may be driven by factors such as review volume, app age, update history, or review-soliciting techniques.
4.3.1 Overall sentiment analysis of the combined data
The sentiment classification of user reviews and comments was carried out in this study utilizing a rule-based methodology based on the user-provided numerical ratings. Reviews that scored higher than three were classified as positive, reviews that scored lower than three as negative, and reviews that scored three or above were classified as neutral (Sherman, 2014; Moraes et al., 2013). Without requiring pre-trained models or external sentiment analysis tools, this approach enabled the creation of a systematic, interpretable framework for sentiment categorization (Figure 4).
Figure 4. Sentiment distribution of the merged review dataset.
The sentiment distribution for targeted apps shows a predominantly positive user experience, with 55.38% of reviews rated positive. However, a sizable proportion, 40.96%, reflects negative sentiment, indicating widespread dissatisfaction among users. Neutral feedback accounts for only 3.66%, indicating that consumers prefer to share strong opinions rather than moderate or uninterested reviews. This division in attitude underscores the need to address users’ concerns while preserving and enhancing features that have already been well received. Overall, the finding demonstrates the importance of sentiment analysis in capturing user perceptions and guiding improvements in app quality and satisfaction.
4.3.2 Overall sentiment analysis of applications
Better comprehension across all lending platforms, India Money Mart receives the most positive feedback, with 66.0% of reviews indicating positive sentiment. It is closely followed by LendClub (65.5%) and 5paisa (65.0%), indicating high user performance and a user-friendly, smooth lending process. On the other hand, apps such as Cash Kumar (40.0%) and Lendbox (34.9%) had the fewest positive reviews, indicating poor user satisfaction. In terms of negative sentiment, Lendbox had the highest proportion at 62.5%, followed by CashKumar at 57.4% and Faircent at 55.3%, which might be attributed to difficulties or unfavorable lending terms. Meanwhile, India Money Mart and 5paisa had comparatively few negative reviews at 30.7 and 28.8%, respectively, confirming their positive public image. Neutral sentiment was low across all platforms, with 5paisa at 6.2% and i2ifunding at 1.5%, indicating that consumers often expressed strong opinions. Overall, India Money Mart and 5paisa have the finest sentiment balance, whereas Lendbox has the most unfavorable customer experience and might benefit from strategic customization (Figure 5).
Figure 5. Sentiment comparison across the seven lending apps.
4.3.3 Topic-based sentiment analysis
To learn more about user experience, we conducted sentiment analysis on topics retrieved via topic modeling. This strategy helps developers and prospective users make better decisions. The results show significant variation in sentiment across services. Document verification (87.8%), withdrawals (89.5%), CIBIL score (84%), and OTP and verification (84.0%) received very excellent reviews from users, indicating that post-approval procedures are generally efficient and seamless. On the other hand, significant negative sentiments were expressed regarding the app interface (79.6%) and login rejection (76.0%). All elements received small neutral feedback. To improve overall user satisfaction and app retention, developers should immediately address issues with user onboarding and accessibility (Figure 6).
Figure 6. Topic-wise sentiment analysis of the merged dataset.
4.3.4 Topic sentiment for individual apps
To better understand, each P2P lending app undergoes sentiment analysis on a specific topic. This study enables the classification of apps based on several criteria inferred from user sentiment. Furthermore, it helps identify areas for improvement by highlighting user-identified negative aspects.
4.3.4.1 Topic sentiment analysis for 5Paisa
The sentiment analysis of 5Paisa reviews reveals both advantages and disadvantages. Users were very satisfied with CIBIL and credit scores (89.7%), OTP and verification (84.3%), document verification (85.1%), and withdrawals (80.2%), suggesting consistent execution across core activities. However, the unfavorable opinion was high for the application interface (loan rejection: 62.0%; login troubles: 48.4%). The amount of neutral input was minimal. Overall, while the app excels in financial services, improving usability and login functionality would enhance the user experience (Figure 7).
Figure 7. Topic-wise sentiment analysis for 5paisa.
4.3.4.2 Topic sentiment analysis for Faircent
Faircent outperforms in essential lending activities, such as document verification (87.9%), withdrawals (91.8%), and OTP verification (83.1%), demonstrating high user satisfaction with its core financial services. However, the platform shows significant user-experience difficulties, including the app interface (87.7% negative sentiment), the application process (79.2%), the application experience (68.7%), and login issues (70.6%). While customer service received mixed feedback, most other areas elicited strong emotional responses, indicating substantial usability and technical problems. Overall, Faircent is dependable for transactions, but it needs to improve its user interface, onboarding process, and accessibility to increase overall user satisfaction (Figure 8).
Figure 8. Topic-wise sentiment analysis for Faircent.
4.3.4.3 Topic sentiment analysis for i2iFunding
i2iFunding receives strong, favorable ratings for its basic services, including document verification (93.9%), withdrawals (84.6%), CIBIL and credit score checks (84.6%), and repayment and EMI processes (83.5%), indicating high user satisfaction with its financial operations. However, the platform is criticized for its app interface (86.4% negative sentiment), loan rejection experiences (88.3%), application experience and customer service (73%), and login issues (59.3%). These negative ratings point to notable usability and transparency concerns. Overall, the feedback highlights i2ifunding’s expertise in financial operations while also indicating the need for significant changes in interface design, accessibility, and user onboarding experience (Figure 9).
Figure 9. Topic-wise sentiment analysis For i2iFunding.
4.3.4.4 Sentiment analysis for LenDenClub
According to a sentiment analysis of LenDenClub evaluations, consumers reported highly positive experiences with withdrawals (96.8%), document verification (90.9%), OTP verification (89.0%), and the loan process (88.2%). However, unfavorable opinions predominated across the app interface (72.2% negative), loan rejection (70.2%), login issues (67.3%), and overall app experience (62.8%), indicating user dissatisfaction. Although customers value the basic lending features, technological issues and early-stage procedures appear to require improvement, revealing both strengths and weaknesses in the overall user experience (Figure 10).
Figure 10. Topic-wise sentiment analysis for LenDenClub.
4.3.4.5 Topic sentiment analysis for CashKumar
The sentiment analysis of CashKumar reviews reveals highly unfavorable comments, particularly regarding loan rejection (88.6%), customer service (81.0%), the interface (78.7%), and the application experience (65.4%), indicating issues with usability, support, and loan approvals. Login difficulties (56.2%) were also criticized. On the positive side, users rated document verification (87.4%), withdrawals (83.3%), repayment (70.4%), and the loan process (51.4%) favorably, with mixed feedback on CIBIL and credit score checks. Neutral feedback was minimal, indicating a strong overall consensus among users. Although CashKumar performs well in loan disbursements, repayments, and withdrawals, it needs to improve its user interface, customer service, and transparency in loan approvals to enhance user satisfaction (Figure 11).
Figure 11. Topic-wise sentiment analysis for CaskKumar.
4.3.4.6 Sentiment analysis for Lendbox
Lendbox performs well across financial operations, including loan repayment and EMI (82.4%), loan processing (80.8%), document verification (80.8%), and credit score handling (60%). Nevertheless, the platform receives strong criticism for loan refusal (95.9% negative), customer service (90% negative), app experience (87.5% negative), login issues (86.9% negative), and the app user interface (82.8% negative). Opinions regarding OTP verification are mixed, with 54.5% positive and 45.5% negative sentiment. Although Lendbox performs reliably in core learning operations, its overall dependability and user satisfaction are limited by technological issues and an inadequate user interface (Figure 12).
Figure 12. Topic-wise sentiment analysis for Lendbox.
4.3.4.7 Sentiment analysis for IndiaMoneyMart
IndiaMoneyMart receives exceptionally positive feedback on its financial operations, including CIBIL score checks, OTP and verification, and withdrawals (all 100%), demonstrating strong user confidence in its core services. The loan process (97.3%), document verification (85.7%), repayment (71.9%), customer service, and overall app experience also receive positive feedback. However, users are dissatisfied with loan rejections (100% negative) and the app interface (71.9% negative), indicating clear areas for improvement. Overall, IndiaMoneyMart stands out as a dependable, high-performing loan platform, though its user interface and login functionality need improvement (Figure 13).
Figure 13. Topic-wise sentiment analysis for IndiaMoneyMart.
4.4 The extraction of features
Feature extraction is essential to the suggested methodology for evaluating user reviews of digital lending applications regulated by the RBI, as it transforms unstructured textual data into representations useful for ML and DL algorithms. In this study, we use traditional vectorization techniques and advanced word embedding approaches to capture both semantic and structural information.
Traditional methods, such as Bag-of-Words, TF-IDF, and hashing algorithms, are initially employed. BOW measures the frequency of words such as fault-related phrases, treating each text as a collection of words regardless of structure or word order (Zhao et al., 2022). TF-IDF improves on BOW by calculating the relevance of a word in a document relative to the entire corpus, resulting in a more informative vector representation (Ahuja et al., 2019). Hashing uses hash functions to convert words into fixed-length sparse vectors, which is computationally efficient. However, while these algorithms capture basic text structure, they sometimes fail to preserve contextual meaning, have excessive dimensionality, and have limited accuracy (Singh and Gupta, 2022). To overcome these drawbacks. Advanced word-embedding methods, such as word2Vec, GloVe, FastText, and IndicBERT, are used. Word2Vec learns high-quality word representations from context, enabling the model to interpret word relationships (Alshari et al., 2020). FastText goes beyond this by including subword information, which is especially valuable in morphologically rich languages (Yao et al., 2020). GloVe constructs word vectors using global co-occurrence statistics and performs classification tasks using backpropagation neural networks (Mahmood and Abdulazeez, 2019). IndicBERT, a transformer-based multinational model trained on various Indian languages, improves the system’s ability to understand regional language nuances. In morphologically rich languages, models such as GloVe construct word vectors using global co-occurrence statistics, providing a semantic foundation for performance classification (Sankalp et al., 2024). By integrating both conventional and current feature extraction techniques, the framework facilitates a more comprehensive and contextualized examination of user feelings, often compared with traditional methods and subjects in reviews of digital lending apps, ultimately enhancing model performance and insight generation.
4.4.1 ML algorithms
We are applying machine learning to customer reviews for improve analysis by enabling automation, increasing accuracy, and delivering valuable insights. Unlike rule-based techniques, ML models can correctly identify new, unread reviews as positive or negative. This enables the organization to monitor real-time sentiment, discover issues such as bugs or UI complaints, and assess app performance following changes. ML also powers advanced applications that recognize users’ tone and systems that identify unexpected increases in negative comments. Furthermore, it supports customer segmentation and turnover prediction, enabling app developers to optimize the user experience and make informed decisions based on user feedback patterns (Table 5).
Table 5. Manual vs. machine learning analysis.
4.4.2 DL algorithms
Without the need for handcrafted features, deep learning models automatically discover complex patterns and contextual correlations, making them ideal for analyzing text data such as user reviews. DL techniques like LSTM and BERT, unlike classical models, capture word order, semantics, and deeper meaning, thereby improving sentiment classification accuracy. In addition to supporting real-time analysis and scaling to large datasets, these models are ideal for gaining insight into user sentiment and improving decision-making in app review analytics.
4.5 Evaluation of prediction model effectiveness
To evaluate the classification models’ performance, we use F1 score, recall, accuracy, and precision as evaluation metrics. Accuracy indicates total correctness, whereas precision represents the proportion of actual positive predictions. Recall represents the model’s ability to recognize true positives, and the F1 score provides a comprehensive view of the model’s effectiveness by balancing accuracy and recall, particularly with imbalanced data. This evolution is conducted across various ML models using BOW, TF-IDF, hashing, and DL models such as Word2Vec, FastText, GloVe, and Indic-BERT embeddings (Table 6).
Table 6. Performance of ML and DL models using the BOW technique.
The table compares the performance of several machine learning and deep learning models with Bow across key parameters, including accuracy, precision, recall, and F1. CatBoost and XGBoost are the most accurate machine learning classifiers (0.87 and 0.86, respectively), followed by SVM, random forest, LightGBM, and logistic regression, all of which perform well across various criteria. AdaBoost also performs well, although at a significantly lower level. In deep learning models, VGG16 matches CatBoost’s accuracy (0.87), while BiLSTM performs similarly (0.87). Overall, both advanced ML models, such as CatBoost, and the deep learning model VGG16, achieve top results, though ResNet’s low scores underscore the need for careful model selection (Table 7).
Table 7. Performance of ML and DL Models using the TF-IDF technique.
TI-FIDF shows that the two machine learning classifiers with the highest accuracies (0.87 each) are CatBoost and LightGBM. They are closely followed by SVM, random forest, XGBoost, and logistic regression, all of which have strong, comparable performances (0.86). The scores of AdaBoost and the decision tree are marginally lower. VGG16 stands out in deep learning for its high accuracy and balanced metric values (all 0.87 except the F1 score, which is 0.85). BiLSTM also exhibits robust results (0.87), comparable to the best models. While ResNet performs poorly, the majority of deep learning models perform well. This underscores the importance of choosing models that are appropriate for the data (Table 8).
Table 8. Performance of ML and DL models using the hashing technique.
The table shows the performance of machine learning models that employ hashing algorithms. SVM achieved the best performance, with an accuracy of 0.86 and an F1 score of 0.85. XGBoost, CatBoost, and LightGBM were close behind, each obtaining an accuracy of 0.85 and an F1 score of 0.84. Random forest also performed well (0.84 accuracy, 0.83 F1 score). Logistic regression and AdaBoost achieved lower accuracies of 0.81 and 0.80, respectively. The decision tree achieved the lowest results (0.77 accuracy, 0.76 F1 score).
Among the deep learning models, VGG16 and BiLSTM performed best, achieving classification accuracies of 0.84 and F1 scores of 0.83 and 0.82, respectively. ResNet, however, performed poorly, with an accuracy of 0.61 and an F1 score of 0.59. Overall, the SVM, VGG16, and BiLSTM models outperformed the others (Table 9).
Table 9. Performance of ML and DL models using the Word2Vec technique.
According to the Word2Vec evaluation results, the majority of machine learning models, including logistic regression, Random forest, SVM, and CatBoost, achieved high classification accuracy (87%) and balanced precision, recall, and F1 scores. LightBGM closely followed with 0.86% accuracy, while XGBoost achieved the highest F1 score (0.86) for precision and recall. CatBoost performed exceptionally well in recall (0.87). AdaBoost and the decision tree showed comparatively lower performance. VGG16 and BiLSTM matched the top ML models, achieving 86% accuracy and an F1 score of 0.84, while ResNet trailed with 85% accuracy and an F1 score of 0.80. Overall, the most reliable algorithms were logistic regression, random forest, XGBoost, CatBoost, VGG16, and BiLSTM (Table 10).
Table 10. Performance of ML and DL models using the FastText technique.
The performance comparison between deep learning and machine learning with FastText models demonstrates that XGBoost, Random forest, and CatBoost outperform, with accuracies of 0.87%. XGBoost, in particular, shows strong performance with a precision of 0.85, a recall of 0.87, and an F1 score of 0.85. In comparison, the decision tree performs poorly, with an F1 score of 0.80 and an accuracy of 0.80. Deep learning models such as VGG16 achieved competitive results, with an accuracy of 0.85, while ResNet and BiLSTM each achieved an accuracy of 0.84 (Table 11).
Table 11. Performance of ML and DL models using the GloVe technique.
According to the results, XGBoost, SVM, and CatBoost outperformed the other machine learning models, achieving 87% accuracy and strong precision, recall, and F1 scores (0.83–0.87), demonstrating a solid balance between accurately detecting approvals and rejections. With 85–86% accuracy, LightGBM random forest and logistic regression also performed well; however, AdaBoost and the decision tree performed worse, indicating lower prediction reliability and somewhat lower accuracy. VGG16 outperformed all other models in deep learning and produced reliable, robust predictions for accuracy, precision, recall, and F1 score, with an accuracy of 0.87%. BiLSTM trailed closely with 0.86% accuracy, while ResNet’s poor performance (78% accuracy) indicated lower predictive capacity than the other models. Overall, the outcomes of the best deep learning and machine learning models were competitive (Table 12).
Table 12. Performance of ML and DL models using the Indic-BERT technique.
The findings indicate that Extreme Gradient Boosting was the top-performing machine learning model (74% accuracy) and LightGBM (73%), followed closely by CatBoost (71%). These models demonstrated balanced precision, recall, and F1 scores, indicating consistent predictions. Random forest and logistic regression produced acceptable results, while the decision tree and SVM performed worse. Deep learning performance was generally lower, with VGG16 obtaining 61% accuracy and BiLSTM at 56%. ResNet had the lowest overall performance, with a precision of 0.29 and an F1 score of 0.37, suggesting poor predictive ability. Gradient-boosting models (XGBoost, LightGBM, and CatBoost) beat both classical ML and deep learning approaches in this examination.
4.5.1 Total ML model accuracy
The accuracy comparison shows that the majority of machine learning models performed well across a range of embedding approaches, including Word2Vec, Indic-BERT, FastText, GloVe, Hashing, BOW (bag-of-words) performance metric, and TF-IDF. The robustness of XGBoost, SVM, CatBoost, random forest, logistic regression, and LightGBM to feature representation techniques is demonstrated by their top accuracies of 0.86–0.87 with little change among embeddings. The accuracy of AdaBoost ranged from 0.65 to 0.85 depending on the embedding, whereas the decision tree’s accuracy decreased significantly, especially with IndicBERT (0.56) and GloVe (0.76). Regardless of the embedding choice, sophisticated ensemble models and gradient boosting techniques often demonstrated stability and improved accuracy, whereas simpler models were more sensitive to the feature representation used (Figure 14).
Figure 14. Total ML model accuracy comparison across different embedding techniques.
The chart compares F1 scores from various machine learning models that use embeddings, including BOW, FastText, GloVe, hashing, Indic-BERT, TF-IDF, and Word2Vec. Logistic regression, random forest, and XGBoost achieved the most consistency, with scores ranging from 0.85 to 0.87. AdaBoost, CatBoost, and LightGBM all perform well, but the decision tree and SVM exhibit more fluctuation, with the decision tree dropping to 0.55 (Indic-BERT) and the SVM dropping below 0.56 in some embeddings (Indic-BERT). BOW, TF-IDF, and FastText all produce superior results, whereas Indic-BERT is less consistent. Overall, ensemble and boosting models perform well across embeddings, making them suitable for text classification applications (Figure 15).
Figure 15. F1 score comparison of ML models across embedding techniques.
4.5.2 Total Dl model accuracy
When comparing the accuracy of deep learning models, VGG16 and BiLSTM consistently outperform ResNet across the majority of embedding approaches. VGG16 achieved the highest accuracy (0.88) with GloVe, followed by BOW, TF-IDF, Word2Vec, and FastText (0.85–0.88), demonstrating strong adaptation across different feature representations. BiLSTM fared well with BOW and TF-IDF, reaching 0.87, but dropped significantly to 0.56 with Indic-BERT, showing sensitivity to certain embeddings. ResNet had the poorest overall performance, with accuracies ranging from 0.54 (Indic-Bert) to 0.86 (Word2Vec), suggesting limited tolerance for alternative embedding types. Overall, VGG16 emerged as the most consistent and accurate model, followed closely by BiLSTM, whereas ResNet struggled with many embedding strategies. Additionally, models such as VGG16, which have stronger feature-extraction capabilities, remain more stable across a range of text representations. These results demonstrate that embedding selection can significantly influence the effectiveness of deep learning (Figure 16).
Figure 16. Total DL model accuracy comparison across different embedding techniques.
The chart compares the F1 scores of deep learning models (BiLSTM, ResNet, and VGG16) utilizing embeddings such as BOW, FastText, GloVe, Hashing, Indic-BERT, TF-IDF, and Word2Vec. VGG16 performs well, notably with GloVe (0.87), and generally scores 0.83–0.85 with other embeddings, with the lowest result for Indic-BERT (0.58). BiLSTM performs well across most embeddings (0.82–0.85), except for Indic-BERT (0.53). ResNet has the greatest variety, with Word2Vec (0.84) performing best and Indic-BERT (0.38) performing worst. Overall, VGG16 and BiLSTM maintain consistent high performance, whereas ResNet’s effectiveness depends largely on the embedding used, underscoring the need to match the proper embedding to the model (Figure 17).
Figure 17. F1 score comparison of ML models across embedding techniques.
5 Discussion
The analysis of users’ reviews reveals substantial variation in public perception across the various money-lending platforms and shows clear patterns in satisfaction and operational efficiency. Overall, most platforms outperform in key financial services such as loan processing, withdrawals, repayment, OTP verification, and document verification, demonstrating customer trust in post-approval processes.
Among the platforms, IndiaMoneyMart has the most favorably rated platform, with consistently high positive satisfaction in core services such as CIBIL handling (100%), OTP (100%), and withdrawals (100%). Users also express positive responses toward remaining features. It suggests exceptional consistency in financial execution. However, this platform receives completely negative feedback on loan rejections (100%) and receives criticism for its poor app interface and weak responsiveness. i2iFunding also performs strongly, excelling in document verification (93.9%) and payment (83.5%), while withdrawals score (84.6%), which indicates reliability in transaction-related functions. Despite this, the platform faces strong criticism for its high rejection rate (88.3%) and weaknesses in app interface, user experience, customer service, and login issues. This indicates reliability in transaction-related functions. Similarly, LendBox also performs well in repayment (82.4%), document verification (80.8%), and the loan process, thereby confirming user trust in its credit handling. Nevertheless, users highlight major drawbacks, including high loan rejection rates (95.9%). LenDenClub and Faircent maintain high positive sentiment for withdrawals (96.8 and 91.8%). Simultaneously, both apps receive positive feedback for document verification (90.9 and 89.9%). However, it still suffers from poor user satisfaction with the application interface, user experience, customer support, and login experience. While CashKumar demonstrates moderate outcomes, it performs well in document verification (87.4%) and withdrawals (83.3%). It receives an average inclusive response (50%) for CIBIL and credit scoring and earns positive feedback on the loan process and OTP verification. Users remain dissatisfied with its loan rejections (88.6%). Finally, 5Paisa achieves the lowest overall satisfaction, despite good performance in CIBIL handling (89.7%), document verification (85.1%), and OTP verification (84.3%), highlighting consistent execution of its core financial tasks. However, it receives negative feedback for loan rejection and the app interface. All platforms received the fewest neutral ratings.
Overall, users are very satisfied with key financial processes, such as loan processing, withdrawals, repayments, and OTP verification, aligning with the Technology Acceptance Model (TAM) and UTAUT framework principles of perceived usefulness and performance expectancy. Platforms such as India Money Mart and i2iFunding excel in these areas. High ratings for loan disbursement, repayment, and withdrawals across most platforms are also consistent with the SERVQUAL dimensions of reliability and responsiveness, which highlight prompt, dependable service as critical drivers of customer satisfaction. Variables such as app interface, login experience, and navigation efficiency are related to perceived ease of usage (TAM) and effort expectancy. Those who have a positive experience with the loan acceptance, withdrawal, and repayment procedures are more likely to continue utilizing these services. Platforms that handle credit ratings with security and transparency align with the trust-risk paradigm. Loan denials and ambiguity erode users’ sense of control and trust, which in turn erodes their confidence that the system will successfully address their financial needs. Login issues and loan denials indicate deficiencies in perceived ease of use and effort expectancy, which are essential obstacles to TAM and UTAUT’s use of technology.
The platform-specific variances in sentiment and topic trends are descriptive contrasts obtained from the review data. There are no inferential statements about the superiority of one platform over another. Some variations across apps were minor and should not be interpreted as statistically significant. The distribution of reviews may be influenced by a variety of contextual factors, such as software maturity, release cycles, and timing of user interactions (Table 13).
Table 13. Performance assessment of money lending loan applications.
Topics were labeled for each platform based on the dominant sentiment in reviews. A topic was rated as good if the proportion of positive reviews exceeded the combined share of neutral and negative reviews, average if neutral reviews accounted for the majority, and below average if negative reviews dominated. When no review data was available for a specific topic on a platform, it was designated as NA (Table 14).
Table 14. Overall public perception of digital lending platforms.
Beyond modeling performance, the perception analysis of lending platforms reveals distinct differences in public opinion. Ensemble and gradient boosting classifiers such as XGBoost, SVM, CatBoost, random forest, logistic regression, and LightGBM consistently achieved highest accuracies, ranging from 0.86 to 0.87, indicating strong adaptability to diverse text representations, including BOW, TF-IDF, FastText, Word2Vec, and GloVe. In contrast, simpler models such as AdaBoost and Decision achieved lower accuracies (0.77–0.80), indicating limited generalization across different embedding techniques. Among deep learning approaches, VGG16 consistently performed best, achieving 0.88 accuracy with GloVe embeddings, followed closely by BiLSTM with 0.87 for BOW and TF-IDF. Both models maintained balanced precision and recall, with F1 scores between 0.84 and 0.87, while ResNet showed poor adaptability, dropping to an accuracy of 0.50 under Indic-BERT. These results highlighted that model architecture and embedding compatibility significantly influence predictive performance. According to the findings, the most dependable methods for sentiment and loan classification tasks in digital lending applications are sophisticated ensemble and deep learning techniques, which deliver the most accurate and stable results, with top accuracies of 0.86–0.88 and consistent F1 scores.
As presented in Table 15, the models’ primary results highlight clear patterns in user perception and behavioral patterns in digital lending. Boosting models such as CatBoost, LightGBM, and XGBoost achieved the highest accuracies (0.86–0.87) across all approaches, while the deep learning model VGG16 achieved the overall best accuracy (0.88 with GloVe). These findings demonstrate that hybrid ML-DL frameworks outperform standard models such as decision trees and AdaBoost when processing large-scale unstructured review data. These models’ strong, stable performance reflects their ability to capture emotional tone, contextual semantics, and topic variations in user feedback, which are critical for understanding behavioral constructs such as perceived trust, usefulness, and ease of use, as described in the TAM, UTAUT, and Trust-Risk frameworks. Predictive modeling accuracy translates into clearer behavioral insights. Apps with higher expected sentiment, such as India Money Mart and i2iFunding, correlate with greater user satisfaction and trust, consistent with the performance expectancy and reliability constructs of UTAUT and SERVQUAL, respectively. Platforms such as 5Paisa and Lendbox, which had higher levels of unfavorable sentiment, showed problems with perceived ease of use and service assurance. Overall, the combination of quantitative modeling and behavioral interpretation illustrates how strong ML-DL models not only improve prediction accuracy but also expand theoretical understanding of user adoption, trust, and satisfaction in the digital lending landscape. This combined evidence extends previous research by empirically demonstrating that greater model precision yields more reliable insights into borrower experience and perceived transparency, which are major drivers of long-term FinTech adoption. The inclusion of theory-derived variables enabled the ML models to identify which factors had the greatest influence on user outcomes. SHAP research revealed that utility- and trust-related material elicited positive sentiment, whereas responsiveness and dependability issues elicited negative sentiment. These findings complement TAM and the Trust-Risk paradigm by showing that usefulness and trust enhance positive perceptions, while service quality gaps heighten discontent. The evidence indicates that theory constructs have a meaningful impact on sentiment patterns.
Table 15. A concise summary table comparing model performance.
6 Conclusion
This study evaluated 15,408 Google Play Store reviews from seven RBI-approved Indian P2P lending apps (5Paisa, Faircent, i2iFunding, LenDenClub, CashKumar, Lendbox, and IndiaMoneyMart) using an integrated NLP, ML, and DL framework. The combined preprocessing, topic modeling, sentiment analysis, and predictive modeling workflow included data cleaning, LDA topic modeling (11 topics), sentiment analysis using VADER, and a broad suite of classical and deep-learning classifiers. Overall, sentiment was moderately favorable at 55%, with a focus on post-approval activities, including loan processing, withdrawals, EMI repayments, and OTP verification. Negative feedback (40.96%) was related to onboarding, interface, login issues, loan denials, and CIBIL. The comparative findings reveal various strengths and shortcomings among the lending systems. The results demonstrate that robust text categorization depends more on careful selection of models and architectures than on any single embedding technique. In terms of user impressions, the studies show clear differences among platforms. IndiaMoneyMart and i2iFunding receive the highest ratings, indicating consistent satisfaction with core lending services, including loan processing, verification, and withdrawals. Lendbox and LenDenClub both perform well in these areas but face frequent concerns about interface design, login issues, and transparency around rejections. Faircent and CashKumar have received mixed reviews, with strengths in withdrawals and service but issues with support and loan processing.
In contrast, 5Paisa receives the most unfavorable feedback since its powerful CIBIL and verification features are balanced by poor repayment, login, and user experiences. Among the models tested, boosting classifiers (XGBoost, CatBoost, and LightGBM) and the deep learning architecture VGG16 consistently provided the most reliable performance, boosting models’ high accuracies (0.86–0.87), and VGG16 had the study’s best single accuracy (0.88 with GloVe), confirming their suitability for text-based financial applications. India Money Mart is the most well-liked app, with consistently high ratings for CIBIL handling, loan processing, OTP verification, and withdrawals. Finally, 5Paisa has the lowest overall satisfaction: despite great results in CIBIL handling, document verification, and OTP verification, it receives overwhelmingly unfavorable feedback on repayment processes, login functionality, loan denials, and general application use. The importance of the TAM, UTAUT, SERVQUAL, and Trust-Risk frameworks in digital lending contexts was reinforced by empirical evidence of their impact on user sentiment, as evidenced by the incorporation of theory-derived features into ML models. Overall, these findings highlight the need to adopt robust analytical models while addressing usability, transparency, and customer service issues in digital lending ecosystems to increase user trust and satisfaction.
6.1 Practical implication
Managerial and policy implications follow immediately. Platforms should prioritize UI/UX redesign, strengthen authentication/login procedures, increase transparency about rejection criteria and credit-score consequences, and invest in proactive customer support for high-friction steps (onboarding, verification, and rejections). Large-scale review mining provides regulators and consumer-protection agencies with a low-latency signal to monitor market behavior and identify recurring harms (e.g., opaque rejections and aggressive collection), in addition to standard supervisory instruments.
6.2 Future directions
Future studies can be expanded to include all remaining RBI-approved P2P lending applications, thereby providing a more comprehensive assessment of the Indian digital lending ecosystem. Furthermore Compare RBI registered apps and non registerd apps. They include human-annotated, multilingual corpora, longitudinal tracking, fairness and interpretability audits of top models, and correlations between sentiment and operational KPIs such as approval times, defaults, and repayment outcomes. Combining review signals and structured telemetry can improve early warning and product-quality analytics in digital lending. The dataset can also be expanded to include evaluations from different platforms (iOS, web portals, social media, and consumer forums) to provide a more complete picture of user perceptions. Multilingual reviews and complex preparation approaches can help capture regional language nuances and subtle expressions that are often lost in ASCII-based filtering. In addition.
Data availability statement
The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/supplementary material.
Ethics statement
Ethical approval was not required for the study involving humans in accordance with the local legislation and institutional requirements. Written informed consent to participate in this study was not required from participants or their legal guardians/next of kin, in accordance with national legislation and institutional requirements.
Author contributions
KRS: Methodology, Conceptualization, Writing – original draft, Formal analysis, Software. SSS: Supervision, Validation, Investigation, Writing – original draft, Formal analysis.
Funding
The author(s) declared that financial support was received for this work and/or its publication. This research paper funded by VIT-AP University.
Acknowledgments
We sincerely thank Basaraboyina Yohoshiva for their insightful assistance throughout this study. Their guidance strengthened both the clarity and depth of the work, and their support was invaluable during the research process.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that Generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
Abu-taieh, E. M., Alhadid, I., Abu-tayeh, S., and Alkhawaldeh, R. S. Continued intention to use of M-banking in Jordan by integrating UTAUT, TPB, TAM and service quality with ML Basel, Switzerland: MDPI AG. (2022) doi: 10.3390/joitmc8030120
Addy, W. A., Ofodile, O. C., Adeoye, O. B., Oyewole, A. T., Okoye, C. C., Odeyemi, O., et al. (2024). Data-driven sustainability: how fintech innovations are supporting green finance. Eng. Sci. Technol. J. 5, 760–773. doi: 10.51594/estj.v5i3.871
Ahuja, R., Chug, A., Kohli, S., Gupta, S., and Ahuja, P. (2019). The impact of features extraction on the sentiment analysis. Procedia Comput. Sci. 152, 341–348. doi: 10.1016/j.procs.2019.05.008
Alam, S., and Yao, N. (2019). The impact of preprocessing steps on the accuracy of machine learning algorithms in sentiment analysis. Comput. Math. Organ. Theory 25, 319–335. doi: 10.1007/s10588-018-9266-8
Alawaji, R. (2025). Sentiment analysis of digital banking reviews using machine learning and large language models. Electronics 14:2125. doi: 10.3390/electronics14112125
Aldboush, H. H. H., and Ferdous, M. (2023). Building Trust in Fintech: an analysis of ethical and privacy considerations in the intersection of big data, AI, and customer trust. Int. J. Financ. Stud. 11:90. doi: 10.3390/ijfs11030090
Al-Natour, S., and Turetken, O. (2020). A comparative assessment of sentiment analysis and star ratings for consumer reviews. Int. J. Inf. Manag. 54:102132. doi: 10.1016/j.ijinfomgt.2020.102132
Alshari, E. M., Azman, A., Doraisamy, S., Mustapha, N., and Alksher, M. (2020). Senti2vec: an effective feature extraction technique for sentiment analysis based on word2vec. Malays. J. Comput. Sci. 33, 240–251. doi: 10.22452/mjcs.vol33no3.5
Amrie, S., Kurniawan, S., Windiatmaja, J. H., and Ruldeviyani, Y., Analysis of Google play store’s sentiment review on Indonesia’s P2P Fintech platform, 2022 IEEE Delhi sect. Conf. DELCON 2022, 1–5, (2022).
Anifa, M., Ramakrishnan, S., Joghee, S., Kabiraj, S., and Bishnoi, M. M. (2022). Fintech innovations in the financial service industry. J. Risk Financ. Manag. 15:287. doi: 10.3390/jrfm15070287
Bao, T., Ding, Y., Gopal, R., and Möhlmann, M. (2024). Throwing good money after bad: risk mitigation strategies in the P2P lending platforms. Inf. Syst. Front. 26, 1453–1473. doi: 10.1007/s10796-023-10423-4
Barik, R., and Sharma, P. (2019). Analyzing the progress and prospects of financial inclusion in India. J. Public Aff. 19:e1948. doi: 10.1002/pa.1948
Baron, S., Patterson, A., and Harris, K. (2006). Beyond technology acceptance: understanding consumer practice. Int. J. Serv. Ind. Manag. 17, 111–135. doi: 10.1108/09564230610656962,
Bazarbash, M. (2019). FinTech in financial inclusion: machine learning applications in assessing credit risk. IMF Work. Pap. 2019:1. doi: 10.5089/9781498314428.001
Cevik, S. (2024). The dark side of the moon? Fintech and financial stability. Int. Rev. Econ. 71, 421–433. doi: 10.1007/s12232-024-00449-8
Chen, D., Lai, F., and Lin, Z. (2014). A trust model for online peer-to-peer lending: a lender’s perspective. Inf. Technol. Manag. 15, 239–254. doi: 10.1007/s10799-014-0187-z
Chen, Q. L., and Zhou, Z. H. (2016). Unusual formations of superoxo heptaoxomolybdates from peroxo molybdates. Inorg. Chem. Commun. 67, 95–98. doi: 10.1016/j.inoche.2016.03.015
Davis, F. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS quarterly. 319–340.
Elias, O., Awotunde, O. J., Oladepo, O. I., Azuikpe, P. F., Samson, O. A., Oladele, O. R., et al. (2024). The evolution of green fintech: leveraging AI and IoT for sustainable financial services and smart contract implementation. World J. Adv. Res. Rev. 23, 2710–2723. doi: 10.30574/wjarr.2024.23.1.2272
Feldman, R., Sanger, J., and Mihalcea, R. Book reviews, Cambridge, England: Cambridge University Press. (2007)
Gabor, D., and Brooks, S. (2017). The digital revolution in financial inclusion: international development in the fintech era. New Polit. Econ. 22, 423–436. doi: 10.1080/13563467.2017.1259298
Ghosh, C., and Hom Chaudhury, R. (2022). Determinants of digital finance in India. Innov. Dev. 12, 343–362. doi: 10.1080/2157930X.2020.1850012
Greaves, F., Ramirez-Cano, D., Millett, C., Darzi, A., and Donaldson, L. (2013). Use of sentiment analysis for capturing patient experience from free-text comments posted online. J. Med. Internet Res. 15, e239–e239. doi: 10.2196/jmir.2721,
Gupta, S. S., and Mahajan, J. (2023). User sentiment analysis of Cashkumar peer-to-peer (P2P) lending platform: based on Google reviews, 110, 97–122. doi: 10.1108/S1569-37592023000110B006
Hofmann, T.. Probabilistic latent semantic indexing, Proc. 22nd Annu. Int. ACM SIGIR Conf. Res. Dev. Inf. Retrieval, SIGIR 1999, 50–57, (1999).
Huang, Z., Chen, H., Hsu, C. J., Chen, W. H., and Wu, S. (2004). Credit rating analysis with support vector machines and neural networks: A market comparative study. Decis. Support. Syst. 37, 543–558. doi: 10.1016/S0167-9236(03)00086-1
Jain, S. K., and Gupta, G. (2004). Measuring service quality: SERVQUAL vs. SERVPERF scales 29, 25–37. doi: 10.1177/0256090920040203
Khan, S., Singh, R., Baker, H. K., and Jain, G. (2024). Public perception of online P2P lending applications. J. Theor. Appl. Electron. Commer. Res. 19, 507–525. doi: 10.3390/jtaer19010027
Khatri, P. (2019). An overview of the peer to peer lending industry of India. Int. J. Bus. Manag. Invent. ISSN 8, 1–11. doi: 10.13140/RG.2.2.35247.25765
Ko, P. C., Lin, P. C., Do, H. T., and Huang, Y. F. (2022). P2P lending default prediction based on AI and statistical models. Entropy 24, 1–23. doi: 10.3390/e24060801,
Kumari, S. P., Ali, B. M., Mohanty, M., Dash, B. B., Rafiq, M., Mohanty, S. N., et al. (2025). Customer satisfaction in peer-to-peer lending platforms: a text mining and sentiment analysis approach. Results Control Optim. 20:100598. doi: 10.1016/j.rico.2025.100598
Li, Q., Li, P., Mao, K., and Lo, E. Y. (2020). Neurocomputing improving convolutional neural network for text classification by recursive data pruning. Neurocomputing, 414, 143–152. doi: 10.1016/j.neucom.2020.07.049
Liu, B. (2020). Sentiment analysis: mining opinions, sentiments, and emotions. 2nd Edn, Cambridge, United Kingdom: Cambridge University Press. 1–432.
Liu, G. (2025). Research on personal loan default assessment based on machine learning. ITM Web of Conferences, 1012, 1–14. doi: 10.1051/itmconf/20257001012
Liu, X. (2013). Full-text citation analysis: a new method to enhance. J. Am. Soc. Inf. Sci. Technol. 64, 1852–1863. doi: 10.1002/asi
Luong, H. T., and Ngo, H. M. (2024). Understanding the nature of the transnational scam-related fraud: challenges and solutions from Vietnam’s perspective. Laws 13, 1–15. doi: 10.3390/laws13060070,
Mahmood, M. R., and Abdulazeez, A. M. Different model for hand gesture recognition with a novel line feature extraction, International Conference on Advanced Science and Engineering (ICOASE2019), 2018, 52–57, (2019)
Mallinguh, E., and Wasike, C. (2025). An empirical analysis of loan repayment behavior and default rates on digital lending platforms: evidence from an emerging market. Qeios 7, 1–13. doi: 10.32388/dkjluj.2
Maulida, S., and Surbakti, H. (2024). Sentiment analysis of peer-to-peer (P2P) lending: a study of scientific publications. Bus. Suatain. 2:3793. doi: 10.58968/bs.v2i2.379
Modi, A., and Kesarani, V. (2023). Digital lending laws in India and beyond: scrutinizing the regulatory blind spot. Indian J. Econ. Finance 3, 1–7. doi: 10.54105/ijef.a2542.053123
Moraes, R., Valiati, J. F., and Gavião Neto, W. P. (2013). Document-level sentiment classification: an empirical comparison between SVM and ANN. Expert Syst. Appl. 40, 621–633. doi: 10.1016/j.eswa.2012.07.059
Niu, B., Ren, J., Zhao, A., and Li, X. (2020). Lender trust on the P2P lending: analysis based on sentiment analysis of comment text. Sustainability 12:3293. doi: 10.3390/SU12083293
Noori, B. (2021). Classification of customer reviews using machine learning algorithms. Appl. Artif. Intell. 35, 567–588. doi: 10.1080/08839514.2021.1922843
Odei-appiah, S., and Adjei, J. (2021). DigitalCommons @ Kennesaw state university Fintech use, digital divide and financial inclusion Fintech use, digital divide and financial inclusion, Bingley, United Kingdom: Emerald Publishing Limited. 12.
Ogunola, A. A., Sonubi, T., Toromade, R. O., Ajayi, O. O., and Maduakor, A. H. (2024). The intersection of digital safety and financial literacy: mitigating financial risks in the digital economy. Int. J. Sci. Res. Arch. 13, 673–691. doi: 10.30574/ijsra.2024.13.2.2183
Omowole, B. M., Urefe, O., Mokogwu, C., and Ewim, S. E. (2024). Integrating fintech and innovation in microfinance: transforming credit accessibility for small businesses. Int. J. Front. Res. Rev. 3, 90–100.
Pagolu, V. S., and Majhi, B., Sentiment analysis of twitter data for predicting stock market movements, 2016 International Conference on Signal Processing, Communication, Power and Embedded System, 1345–1350, 2016
Pamplona, D. D. M. (2022). Topic identification and classification of GooglePlay store reviews. Cham, Switzerland: Springer. 71, 37–49.
Pang, S., Deng, C., and Chen, S. (2022). System dynamics models of online lending platform based on Vensim simulation technology and analysis of interest rate evolution trend. Comput. Intell. Neurosci. 2022. doi: 10.1155/2022/9776138,
Perea-Khalifi, D., Irimia-Diéguez, A. I., and Palos-Sánchez, P. (2024). Exploring the determinants of the user experience in P2P payment systems in Spain: a text mining approach. Financ. Innov. 10:2. doi: 10.1186/s40854-023-00496-0
Pohan, N. W. A., Budi, I., and Suryono, R. R., Borrower sentiment on P2P lending in Indonesia based on Google playstore reviews, 172, Sriwijaya International Conference on Information Technology and its Applications (SICONIAN 2019), 17–23, (2020), doi: 10.2991/aisr.k.200424.003.
Puro, L., Teich, J. E., Wallenius, H., and Wallenius, J. (2010). Borrower decision aid for people-to-people lending. Decis. Support. Syst. 49, 52–60. doi: 10.1016/j.dss.2009.12.009
Rijanto, A. (2021). Blockchain technology adoption in supply chain finance. J. Theor. Appl. Electron. Commer. Res. 16, 3078–3098. doi: 10.3390/jtaer16070168
Sankalp, K. J., Jain, V., Bhaduri, S., Roy, T., and Chadha, A., Decoding the diversity: a review of the Indic AI research landscape, (2024). doi: 10.48550/arXiv.2406.09559
Sarungu, C. M., Digital lending high level system architecture in Indonesia, Int. Conf. Inf. Technol. Adv. Mech. Electr. Eng. ICITAMEE 2020, 159–164, 2020
Saunders, K. M., and Zucker, B. (1999). Counteracting identity fraud in the information age: the identity theft and assumption deterrence act. Int. Rev. Law Comput. Technol. 13, 183–192. doi: 10.1080/13600869955134
Siering, M., Deokar, A. V., and Janze, C. (2018). Disentangling consumer recommendations: explaining and predicting airline recommendations based on online reviews. Decis. Support. Syst. 107, 52–63. doi: 10.1016/j.dss.2018.01.002
Singh, A., and Gupta, S. (2022). Learning to hash: a comprehensive survey of deep learning-based hashing methods. Knowl. Inf. Syst. 64, 2565–2597. doi: 10.1007/s10115-022-01734-0
Singh, S. (2020). An integrated model combining the ECM and the UTAUT to explain users’ post-adoption behaviour towards mobile payment systems. Australas. J. Inf. Syst. 24, 1–27. doi: 10.3127/ajis.v24i0.2695
Sylvie, M., and Pascal, K., Mobile money: décryptage d Une succes story africaine perceived usefulness, perceived ease of use, and user acceptance of information technology, Minneapolis, Minnesota, USA: Management Information Systems Research Center, University of Minnesota. (2021).
Thelwall, M., Buckley, K., Paltoglou, G., and Cai, D. (2010). Sentiment strength detection in short informal text 1 61, 2544–2558. doi: 10.1002/asi.21416
Vedala, R., and Kumar, B. R., An application of naive Bayes classification for credit scoring in e-lending platform, Proc. 2012 Int. Conf. Data Sci. Eng. ICDSE 2012, 81–84, 2012, doi: 10.1109/ICDSE.2012.6282321.
Venkatesh, V., Morris, M. G., Davis, G. B., and Davis, F. D. (2019). User acceptance of information technology: toward A unified View1. Quarterly 27, 425–478. doi: 10.2307/30036540,
Wang, J. (2015). Predicting Yelp Star Ratings Based on Text Analysis of User Reviews. Stroudsburg, PA, USA: Association for Computational Linguistics (ACL).
Wang, H., and Li, X. (2022). Chinese news text classification based on convolutional neural network. J. Big Data 4:41. doi: 10.32604/jbd.2022.027717
Wang, M., Zhan, G., Lai, K. K., Zhang, L., and Meng, L. (2022). Posts and reviews in P2P online lending platforms: a sentiment analysis and cross-culture comparison. Behav. Inf. Technol. 41, 3591–3597. doi: 10.1080/0144929X.2021.2005679
Wang, X., Xu, Y. C., Lu, T., and Zhang, C. (2020). Why do borrowers default on online loans? An inquiry of their psychology mechanism. Internet Res. 30, 1203–1228. doi: 10.1108/INTR-05-2019-0183
Wang, Z., Jiang, C., Zhao, H., and Ding, Y. (2020). Mining semantic soft factors for credit risk evaluation in peer-to-peer lending. J. Manag. Inf. Syst. 37, 282–308. doi: 10.1080/07421222.2019.1705513
Warin, T., and Stojkov, A. (2021). Machine learning in finance: A metadata-based systematic review of the literature. J. Risk Financ. Manag. 14:302. doi: 10.3390/jrfm14070302
Weiss, S. M., Indurkhya, N., Zhang, T., and Damerau, F. J. (2005). Text mining: predictive methods for analyzing unstructured information, 1–237.
Wei, X., and Croft, W. B., LDA-based document models for ad-hoc retrieval, Proceedings of the Twenty-Ninth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 178–185, (2006)
Xu, J., Lu, Z., and Xie, Y. (2021). Loan default prediction of Chinese P2P market: a machine learning methodology. Sci. Rep. 11, 18759–18719. doi: 10.1038/s41598-021-98361-6,
Yadav, M. (2024). Factors influencing behavioral intentions to use digital lending: an extension of TAM model. Jindal J. Bus. Res. 13, 213–226. doi: 10.1177/22786821231211411
Yao, T., Zhai, Z., and Gao, B., Text classification model based on fastText, Proc. 2020 IEEE Int. Conf. Artif. Intell. Inf. Syst. ICAIIS 2020, 154–157, (2020).
Zetzsche, D. A., Buckley, R. P., Arner, D. W., and Barberis, J. N. (2017). Regulating a revolution: from regulatory sandboxes to smart regulation. SSRN Electron. J. 23:31. doi: 10.2139/ssrn.3018534
Keywords: digital lending, RBI-approved apps, fintech, user perception, sentiment analysis
Citation: Sekhar KR and Saheb SS (2026) User perceptions of RBI-approved P2P digital lending apps: an NLP, machine learning, and deep learning approach. Front. Artif. Intell. 8:1708080. doi: 10.3389/frai.2025.1708080
Edited by:
Arianna Agosto, University of Pavia, ItalyReviewed by:
Jianzheng Shi, Singapore University of Social Sciences, SingaporeDavid Perea-Khalifi, University of Malaga, Spain
Copyright © 2026 Sekhar and Saheb. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Shaiku Shahida Saheb, c2hhaGlkLnNrQHZpdGFwLmFjLmlu