Quantifying and mapping uncertainty in urban sentiment prediction: a combined approach with entropy and SHAP explanations

Abstract

Introduction:

Mental health issues have been increasing globally, possibly linked to urbaniza¬tion and related lifestyles. There’s a growing awareness that different aspects of the urban environment can influence mental health, either enabling or restrict¬ing behaviors that affect well-being. Recent studies have increasingly used social media data and machine learning techniques to assess urban well-being. However, these approaches often lack interpretability and do not explicitly account for prediction uncertainty, limiting their reliability for spatial decision-making.

Methods:

This study uses sentiment labels derived from georeferenced posts on X (formerly Twitter) via the NRC Emotion Lexicon. Supervised machine learning models are trained to predict sentiment based on urban environmental variables and generate spatial predictions for Lisbon, with performance compared across K-Nearest Neighbour (KNN), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Neural Network (NN). To go beyond simple point predictions, uncertainty at each location is measured with Shannon Entropy, based on class probabilities. Post-hoc explanations using SHapley Additive exPlanations (SHAP) identify urban features most impacting sentiment predictions.

Results and discussion:

Uncertainty maps combined with SHAP outputs show where interpretations are more reliable (low uncertainty) and where caution is needed (high uncertainty). In Lisbon 2019, SHAP identified proximity to cycling networks, fitness facilities, and NDVI as key factors influencing sentiment. This framework enhances spatial understanding of sentiment-environment relationships and offers a transparent way to identify areas where predictions are more reliable.

1 Introduction

The city can be viewed as a dynamic interaction between humans and the environment (1, 2). This relationship influences daily experiences, behaviors, and choices, affecting both physical and mental health. While individual health is mainly influenced by social and economic factors, environmental factors and health behaviors are also significant. Behaviors are personal decisions, but environments can promote or trigger specific actions (3). For instance, access to parks encourages physical activity and is linked to lower stress and better mood, while pollution exposure can cause health issues like asthma and is associated with depression and anxiety (3–7). In this framework, subjective sentiments are key indicators of quality of life and mental health, and improving the built environment may foster better residents’ well-being (8).

Although sentiment analysis is important, capturing subjective sentiments in real time remains difficult. Many studies depend on questionnaires, which have limitations like limited quantitative data, coverage issues, data-collection challenges, and problems reproducing results (8). As an alternative, social media data are increasingly used to evaluate population sentiments, emotions, and signals related to mental health (9). Compared to questionnaires and interviews, social media data provide larger samples for sentiment analysis (8) and can aid decision-making aimed at creating healthier, more innovative, and sustainable cities (10). As a result, sentiment analysis has become a common method to explore how online-shared sentiments relate to urban environments (2, 11). However, despite rapid growth in research and data availability, there is still limited evidence from big data on how sentiments vary across space and time and what factors drive them (2). To address this gap, it is necessary to develop approaches that can link sentiment patterns to urban environmental characteristics in a spatially explicit way, while accounting for the complexity and variability of these relationships.

Empirical and traditional statistical techniques like principal component analysis, clustering, regression, and other linear methods have frequently been utilized (12). While these approaches can offer valuable insights for effective planning, management, and decision-making, they do have certain analytical limitations. Specifically, they often fail to fully capture nonlinear behaviors or account for spatial heterogeneity and autocorrelation effects (13–15). These issues are especially significant when modeling complex phenomena such as sentiment expression across diverse urban settings. Machine Learning (ML), a branch of artificial intelligence (AI), has been increasingly used to tackle these issues and is often reported to match or surpass traditional methods in predictive accuracy (16–18). ML can handle diverse data types, structures, and large volumes (big data) (19), and it tends to be less affected by variable scaling (12), making it easier to combine data from various sources to model complex nonlinear relationships that explain sentiment changes in urban settings. However, many ML algorithms remain treated as “black boxes,” complicating the interpretation of how predictions are generated (19). These benefits are especially important in urban sentiment research, where predictors are often numerous, varied, and spatially arranged.

Machine learning methods have gained popularity because of their ability to identify patterns in complex, high-dimensional data (20). High-dimensional data are characterized by having a number of features per observation that is similar to or greater than the number of observations (21). These methods can model both categorical and continuous response variables. However, unlike traditional parametric methods such as multiple regression, they do not easily provide information about prediction error, like the standard error of prediction for a new data point (22).

Explainable Artificial Intelligence (xAI) has become a crucial research area, providing statistical and visualization tools that improve the interpretability of machine learning models (19, 23, 24). Model-agnostic methods are introduced as interpretation techniques to explain the underlying functions driving the entire behavior of ML models (19, 24, 25). The main benefit of a model-agnostic approach is its versatility, as it can interpret any type of black-box ML model, which is vital when ML outputs influence decision-making processes (25).

The model-agnostic approach offers explanations based on the varied behaviors of complex fitted models, providing insights at both global and local levels (12). Typically, global techniques like Permutation Feature Importance (PFI) and Partial Dependence Plot (PDP) describe the overall behavior of a machine learning model, usually as expected values derived from the data distribution. In contrast, local interpretation methods such as Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive ExPlanations (SHAP) focus on explaining individual predictions (19, 26). It is important to remember that local fidelity does not equate to global fidelity; features that are globally significant may not be relevant in specific local contexts, and vice versa (26). Depending on the analysis goal, different methods can be combined to interpret the same model either globally or locally (12).

However, interpretability alone does not resolve a key challenge in spatial prediction: uncertainty. When estimating a variable at a specific location, some degree of uncertainty is inevitable, and “prediction uncertainty” indicates our confidence level in the model’s output (27). While many studies have compared geostatistical and machine learning methods based on point prediction accuracy, fewer have evaluated the trustworthiness of their uncertainty estimates, especially in spatial contexts (22, 28, 29). For example, Coulston et al. introduced a method to estimate prediction uncertainty in Random Forest regression models within a spatial framework (22). Kirkwood et al. analyzed how well Ordinary Kriging and Quantile Regression Forests can produce reliable prediction uncertainties for geochemical mapping in southwest England (28). Similarly, Vaysse and Lagacherie compared these approaches for French digital soil mapping products (29).

At the same time, machine learning techniques have seen a significant rise in remote sensing and geospatial data development (22). For instance, Homer et al., (30) used regression trees to create a categorical land cover map for the United States, while Coulston et al. (31) employed random forests to produce a continuous map of percent tree canopy cover. Additionally, various algorithms such as linear models, generalized additive models, artificial neural networks, support vector machines, normal Bayes, and k-nearest neighbor have been explored and tested (32, 33).

Although several studies have combined machine-learning models with model-agnostic explainability methods to explore links between urban environments and public sentiment (34, 35), the explicit quantification and spatial evaluation of predictive uncertainty in these areas remain largely underexplored. Existing research includes using random forests with Permutation Feature Importance (PFI) to identify park features associated with positive emotions (36) and applying LightGBM with SHapley Additive exPlanations (SHAP) to analyze interactions between geospatial features and sentiments (37). Other approaches involve multilevel regression, gradient boosting decision trees (GBDT) (8), and support vector regression (38) to examine how various urban factors influence sentiment. Despite these efforts, uncertainty is seldom visualized alongside model outputs and explanations, restricting clarity regarding the reliability of spatial predictions and their interpretations.

This study bridges the gap by combining predictive uncertainty with explainable modeling within a spatial sentiment framework. We train multiple machine-learning models to forecast sentiment based on urban environmental factors and assess location-specific uncertainty using Shannon entropy from predicted class probabilities. Using different classifiers together has been shown to be effective, uncovering spatial uncertainty among media sources through an entropy analysis. From a Digital Humanities perspective, a lack of transparency about algorithms’ outputs can raise questions regarding the method’s validity, affecting the analysis (39).

To understand the algorithms in this work, we utilize explainable AI (xAI), also called agnostic models. We employ SHapley Additive exPlanations (SHAP) to elucidate positive sentiment predictions, pinpointing the urban features that most influence the model’s outputs. Additionally, uncertainty maps can be integrated with SHAP explanations to identify areas where spatial predictions and interpretations are more dependable (lower uncertainty) and where caution is needed (higher uncertainty).

The methods described in this paper do not explain how the sentiment classifiers operate or their behavior. Instead, they suggest what features the underlying models might be considering or utilizing for their predictions. This approach enhances trust in the results from a Digital Humanities perspective, since the data is used to draw conclusions about the phenomenon represented in the case study. By linking uncertainty with interpretability, this method enhances the reliability of spatial inferences regarding sentiment and environmental interactions in Lisbon.

2 Materials and methods

The methodological framework was developed to identify the urban environmental factors associated with sentiment in the city of Lisbon. It comprises six main phases: (1) collection and pre-processing of spatial data, (2) sentiment analysis, (3) diagnosis of multicollinearity among predictors, (4) development and validation of machine-learning models, (5) application of Shannon entropy to quantify prediction uncertainty, and (6) interpretation of model outputs using SHAP. The workflow is illustrated in Figure 1.

Figure 1

2.1 Sentiment analysis

Geographic and text-based analyses of social media data allow for the extraction of quantitative sentiment measures that are reliable and highly precise in space and time. Consequently, using social media to measure sentiment is now broadly acknowledged and accepted in academic circles (8). Among available platforms, Twitter (now X) is commonly selected due to its mixed-information content (42) and the relative ease of accessing and collecting data (43). X offers a large volume of near-real-time user-generated content, making it perfect for tracking public opinion across different topics (44–46). The choice of X instead of other platforms is also influenced by its wide public access to content and its easy-to-use interface, which supports the creation of data-extraction methods (47).

In the past, there were some concerns regarding the use of social media content in this context. These are mainly related to data quality, the post’s potential location inaccuracy, the representativeness of the population using social media, and possible biased behavior on social media (48).

Using social media data (49) demonstrated that there is a significant positive correlation between the seasonal patterns of visitors and hotel occupancy rates. Also, a study from (50) revealed a strong correlation between the number of Twitter users in each state and the 2010 US Census state populations.

Research utilizing social media has proven valuable for gaining deeper insights into city life, which is difficult to attain through other methods (51, 52). This approach offers advantages over traditional data sources, providing new opportunities beyond the limitations of costly, invasive cross-sectional surveys (53).

Sentiment Analysis seeks to automatically identify emotions and polarity in a text to classify its sentiment (54, 55). In this study, weuse a lexicon-based method with the NRC Emotion Lexicon (EmoLex), which links words and phrases to eight fundamental emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust), and two sentiment polarities (negative and positive). These associations are created through manual annotations gathered via crowdsourcing (56).

From the tweets published in Lisbon, only the georeferenced and public ones where users could provide exact GPS coordinates were selected. Initially, there were 16,791 georeferenced points from Twitter (now X) comments in Lisbon from 2019. Of these, only the comments made during daytime hours, between 9 a.m. and 7 p.m., summing 9,446 tweets, were included, as these hours represent peak urban space usage (57).

The approach to distinguish visitor tweets from local ones was based on the research conducted by (58). The authors measured the number of days between each user’s first and last uploaded comments. If this difference exceeded the average visit duration in Lisbon (2.1 nights), the tweets were classified as local. If it were shorter, the tweets would be categorized as belonging to visitors. For this analysis, we used the comments posted by local users in Portuguese.

Using the Portuguese version of EmoLex, each tweet was categorized according to specific emotions and sentiment polarities, creating a results table that was imported into a GIS system for visualization and spatial analysis of emotions and sentiments across the city. Out of these tweets, 1,919 matched the lexicon and were assigned a sentiment polarity label.

Lexicon-based approaches can offer distinct advantages over machine learning (ML) methods in sentiment analysis, particularly in specific contexts or languages like Portuguese from Portugal. Lexicon-based methods do not require extensive labeled training data, which is often scarce for less-resourced non-English languages. This makes them faster and more practical in such contexts (59–61).

The complexity of Portuguese creates challenges for sentiment analysis, necessitating specialized models for precise interpretation. This emphasizes a linguistic uniqueness that might be less evident in languages such as English (62, 63). In languages with limited NLP tools, such as Portuguese, lexicon-based methods can achieve competitive results. For instance, a Portuguese-specific lexicon demonstrated strong performance in sentiment analysis tasks (64).

For predictive modeling, the response variable (sentiment) was adjusted to fit a supervised machine-learning classification framework. Specifically, sentiment polarity was converted into a binary variable, with 0 indicating negative sentiment and 1 indicating positive sentiment.

2.2 Study area

Lisbon, the capital of Portugal, is located on the northern bank of the Tagus River estuary. The municipality covers 86.83 km², of which 70.24 km² (≈81%) is classified as urbanized territory (40). Lisbon has 545,796 inhabitants (41), with a population density of 6,286 inhabitants per km². Administratively, the municipality is subdivided into 24 parishes (Figure 2). Lisbon provides a relevant case study due to its diverse urban structure, combining historic areas, residential zones, and large green spaces such as Monsanto Forest Park. The city also presents spatial variability in environmental conditions, including differences in vegetation cover, accessibility to urban infrastructure, and exposure to air pollution and noise.

Figure 2

2.3 Selection of explanatory variables

The selection of the 30 independent variables was based on the conceptual framework of the relationship between urban form and mental well-being (65) and on the conceptual model of urban health-related well-being (UrbWellth) (66). Of all the variables, only those derived from the WorldView-2 satellite were originally found in raster format (2 m spatial resolution in multispectral bands). All other variables were subsequently converted to raster using these as references via distance, interpolation, and density operations (67).

The next stage involved a preliminary statistical analysis in which multicollinearity was assessed by calculating the variance inflation factor (VIF) (68, 69). VIF was computed in R software using a linear regression procedure to diagnose collinearity among the explanatory variables (70). Variables with a VIF greater than 5 were removed to reduce redundancy among predictors (71, 72). After calculating VIF and removing highly correlated variables, the model retained 10 explanatory predictors.

2.4 Machine learning models

The ML models were created using algorithms such as Random Forest (RF), Neural Networks (NNET), k-Nearest Neighbors (KNN), and XGBoost (XGB). The dataset was split into training (90%) and testing (10%) subsets. The training data was used to fit the models, while the test data were used to evaluate performance. Evaluation involved analyzing the confusion matrix to determine accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.

The performance of four machine-learning models in sentiment prediction was assessed. Random Forest (RF) achieved the highest accuracy (0.82), while XGBoost (XGB) showed the lowest (0.71). Despite its comparatively lower predictive accuracy, XGB was still used in the follow-up analysis due to its compatibility with SHAP-based local explanations and its ability to support the visualization of aggregated Shapley values.

2.5 Shannon entropy

Shannon entropy was introduced in information theory to quantify uncertainty in a probability distribution (73). Higher entropy values signify increased uncertainty in the information. In our sentiment modeling framework, the RF and XGB classifiers generate predicted class probabilities for each pixel or location, ranging from 0 to 1. These probabilities can be employed to calculate Shannon entropy, providing a measure of predictive uncertainty in the generated sentiment maps (Equation 1):

where denotes the entropy of the random variable , is the number of possible outcomes, and is the probability of outcome (73). The logarithm is base 2, so entropy is measured in bits (74).

Because sentiment is modeled as a binary outcome (positive vs. negative), with and , entropy is given by (Equation 2):

Entropy is minimal when is close to 0 or 1 (high-confidence predictions) and maximal when , when both outcomes are equally likely and predictive uncertainty is greatest (Figure 3).

Figure 3

We quantified uncertainty in the sentiment maps by using the binary Shannon entropy formula (Equation 2) on the predicted class probabilities at each pixel (see Figure 4). As previously mentioned, sentiment analysis is an NLP technique used to detect the emotional tone of a text and categorize it as positive or negative (75–77). In this context, each piece of text can be seen as a Bernoulli trial, where a random experiment produces just two possible outcomes, commonly called “success” and “failure,” such as positive or negative sentiment (78, 79).

Figure 4

In information theory, entropy indicates the uncertainty or unpredictability of a random variable. In Bernoulli trials, it reflects how hard it is to predict the outcome, making it a helpful measure of uncertainty in sentiment predictions (80). This leads to a specific form of entropy known as Binary entropy, which originates from information theory and quantifies the uncertainty or randomness in a binary variable. It is frequently used as a loss function in binary classification tasks like sentiment analysis. This function enhances the model’s predictions by penalizing mistakes and rewarding correct predictions, thus minimizing overall uncertainty (81, 82). The probabilistic perspective on binary cross-entropy helps estimate the chance of a specific sentiment, which is crucial for understanding the model’s variability.

Using binary entropy to evaluate sentiment prediction models involves measuring prediction uncertainty, optimizing with binary cross-entropy, and applying entropy-based techniques to enhance performance. This metric enables comparison among different sentiment analysis models, emphasizing which better handles uncertainty and produces more reliable predictions. Combining metrics like cross-entropy loss with accuracy, precision, recall, and F1-score provides a comprehensive view of the models’ predictive confidence and uncertainty (83).

This thorough assessment can uncover patterns of uncertainty unique to each model, assisting in choosing the most reliable sentiment analysis model. Moreover, it yields an entropy surface that highlights areas of higher predictive uncertainty, where sentiment classification should be interpreted with greater caution, and areas of lower uncertainty, where model outputs are more reliable (84, 85).

2.6 Explainable AI

Being able to accurately interpret a predictive model’s output is essential because it enhances user trust, offers insights for model improvement, and aids understanding of the modeled process. While simpler models like linear models are often chosen for their interpretability, the rise of big data has led to a preference for complex models, which increases the need to balance accuracy with interpretability. Since complex models are not easily understandable on their own, interpretation usually depends on an explanatory or simplified version of the original model (86).

Explainable AI (xAI), also known as model-agnostic modeling, is designed to interpret AI (and ML) decisions regardless of their complexity or internal structure, without requiring knowledge of how they function. Since an explainer should be capable of explaining any model, it must be model-agnostic (26). Model-agnostic methods can offer both overarching and specific explanations: overarching explanations outline what the model has learned from the input features, while specific explanations detail the reasoning behind individual predictions (12). Importantly, local fidelity does not necessarily mean global fidelity; variables that are important on a global scale might not be important locally, and the other way around (26).

In this study, we analyzed the XGBoost model by using SHapley Additive exPlanations (SHAP) with a model-agnostic Kernel SHAP method to evaluate how each explanatory variable influences the predicted outcome for individual observations (86). To support spatial analysis across Lisbon, SHAP values were visualized within a GIS platform and interpolated using Inverse Distance Weighting (IDW).

IDW is a well-known spatial interpolation technique, especially when the prediction depends on distance-based variables. It assumes that a known data point’s influence diminishes with distance, making it ideal for situations where proximity matters. The weights are inversely related to distance, so nearer points have greater influence on the results (87).

Furthermore, IDW can be improved by adding adaptive features, like distance-decay parameters that vary locally, which increase accuracy by considering spatial patterns and anisotropy in the data (88, 89). Using regression models or neural networks alongside, for example, improves predictions by combining distance-based weights with statistical trends, resulting in lower error rates than individual methods like Kriging (90).

Additionally, many predictor variables used in this study are distance-based, such as distance to green parks and sports facilities, and government official data, such as climate and pollution, are interpolated using IDW before being provided for public use.

The all process generated raster surfaces that highlight locations where certain factors have a greater impact on positive sentiment predictions. IDW proved its suitability for exploratory spatial analysis, while avoiding the imposition of statistical assumptions that may not be appropriate for SHAP-derived values. Such a workflow enables detailed, place-specific interpretation of the model outcomes and can guide targeted urban strategies.

2.6.1 Shapley additive exPlanations

SHAP is a post-hoc explainability framework that assigns a model’s prediction to individual input features by leveraging Shapley values derived from cooperative game theory (19). More specifically, it applies the most commonly used solution concept for non-cooperative games (91), i.e., Nash equilibrium (NE) is a situation in a game where no player can improve their outcome by unilaterally changing their strategy, assuming all other players’ strategies remain the same. This concept can be applied to ensure that sentiment analysis models are stable and robust (92). By treating the interactions between various parts of the sentiment analysis system, such as different classifiers or feature extractors, as a game, it is possible to identify an equilibrium state where the system operates at its best without any component having a reason to change deviate.

In this context, SHAP helps distribute the total gains or costs among players accurately according to their contributions. Using Shapley values, it explains machine learning model predictions, enhancing transparency. SHAP is commonly employed to interpret model predictions by evaluating feature importance, which is vital for understanding and refining sentiment analysis models (93–95). In this case, for a given observation , SHAP estimates how each feature contributes to the predicted outcome by distributing the “payout” (the model prediction) across the set of features in a theoretically grounded and additive manner (19).

Following the SHAP formulation, the explanation model is expressed as an additive feature-attribution model (Equation 3) (19):

where is the explanation model, is a coalition vector indicating whether a feature is present, and , represents the Shapley value (feature attribution) for feature . For the full coalition (all features present), the explanation simplifies to (Equation 4):

In this study, SHAP values were estimated using a model-agnostic Kernel SHAP approach to explain the predicted probability of positive sentiment for each observation.

2.7 Software and libraries

R (version 4.2.3) and RStudio (2023.12.1) were used for sentiment analysis, machine-learning modeling, and explainability. Sentiment and emotion labels were derived using the syuzhet package, applying the NRC method based on the NRC Word–Emotion Association Lexicon.

Machine-learning models were developed using the caret package, which provides a unified workflow for training and evaluating predictive models (96). The XGBoost model was implemented using the xgboost package. Model interpretation was conducted using Kernel SHAP implemented with the shapr package. In addition, the pdp package was used to generate partial dependence plots for the Random Forest model.

3 Results

3.1 Sentiment prediction maps (RF vs. XGB)

Predicted positive sentiment probabilities from the Random Forest (RF) and XGBoost (XGB) models were visualized across Lisbon (Figures 5A,B). Higher values denote a greater likelihood of positive sentiment. In Figure 5A (RF), areas with high predicted probabilities generally align with locations labeled as positive by EmoLex (green points), indicating broad spatial consistency between the probability surface and lexicon-based labels. Lower probabilities appear around Lisbon Airport and residential zones. Figure 5B (XGB) shows a similar spatial pattern, with higher probabilities mainly in the southwest of the city. However, localized differences are evident, especially in Marvila and Parque das Nações, where the distribution of higher probabilities diverges from the RF surface.

Figure 5

3.2 Uncertainty analysis (entropy)

Using Shannon entropy on predicted class probabilities allows for quantifying predictive uncertainty and visualizing spatial areas of varying confidence. After applying Equation 2, the XGB model generally exhibited higher entropy than the RF model, aligning with its lower predictive performance shown in Table 1.

The entropy maps identify regions with higher uncertainty (greater entropy), where sentiment classification is less reliable. In the RF model (Figure 6A), the areas with the highest entropy are mainly in Misericórdia, Avenidas Novas, Areeiro, Marvila, Olivais, and Santa Clara. These zones should be approached with more caution when interpreting RF predictions. For the XGB model (Figure 6B), the highest-entropy zones are again in Misericórdia, Avenidas Novas, and Olivais, showing clusters where the model shows similar probabilities for both classes, indicating increased uncertainty. Conversely, in other parishes, the entropy levels are lower, pointing to more confident predictions with probabilities near 0 or 1.

Figure 6

3.3 Global explanations

Using the XGBoost model, we generated a SHAP summary plot for the sentiment dataset (Figure 7). The dependent variable is binary, derived from the EmoLex-based classification of posts, while the explanatory variables are continuous. Figure 7 can be interpreted as follows (97):

• The y-axis lists the explanatory variables, ordered by importance (ranked by the mean absolute SHAP value);

• The x-axis shows the SHAP value, i.e., the contribution of a feature to the predicted probability of the positive class;

• The color gradient represents the original feature value (low to high). When high values (purple) cluster on the right, higher feature values tend to increase positive sentiment; when they cluster on the left, higher values tend to decrease it.

Figure 7

According to the XGBoost–SHAP framework (Figure 7), the most influential factors explaining positive sentiment in Lisbon in 2019 were the distance to cycling networks, the distance to fitness equipment, and the Normalized Difference Vegetation Index (NDVI). These results align with those from the Random Forest model using PFI (Figure 8). This consistency enhances confidence in the stability of the identified predictors, although the two interpretability methods offer different but complementary insights into how each variable influences model predictions.

Figure 8

3.4 Local explanations

Figure 9 shows the spatial mapping of Shapley-based explanations for Lisbon, emphasizing the variable distance to the cycling network. This map highlights regions where closer proximity to cycle lanes correlates with higher positive sentiment, while areas farther away tend to show negative contributions. Notably, negative contributions are observed around Humberto Delgado Airport and in Santa Maria Maior.

Figure 9

Figure 9 also illustrates the top ten most influential variables for observation 99, which shows a predicted probability of 0.59 for the positive class despite EmoLex labeling it as negative. In this case, the distances to cycling networks, fitness facilities, and green corridors have the largest absolute Shapley contributions. Notably, being 603 meters away from the cycling network lowers the likelihood of positive sentiment, while being 541 meters from green corridors increases it, as explained by the Kernel SHAP analysis. However, it is important to note that observation 99 is located in an area with high entropy, indicating higher uncertainty in the sentiment prediction. Therefore, caution is needed when interpreting the results for this location, as the model’s prediction is less reliable in regions with high uncertainty.

Figure 10 illustrates the spatial distribution of Shapley-based explanations for the variable “distance to fitness equipment.” The findings reveal that this feature contributes both positively and negatively across different areas, without a clear overall pattern. Additionally, the figure presents the Kernel SHAP explanation for observation 45, which has a predicted probability of 0.94 for positive sentiment. In this case, the variables with the most significant Shapley contributions are distance to fitness equipment, weighted average of building age, and place popularity (density of likes on Flickr photos). Specifically, a distance of 1,372 meters to fitness facilities positively influences the likelihood of a positive prediction, while a solar energy value of 5,099 kW·m⁻² has a negative impact. These SHAP values help clarify the model’s reasoning behind this prediction. Furthermore, observation 45 is located in an area with low entropy, which indicates higher certainty in the sentiment prediction. Therefore, the results for this observation can be considered more reliable.

Figure 10

As shown in Figure 11, analyzing NDVI-based Shapley explanations reveals the regions where vegetation significantly boosts the model’s predicted positive sentiment. Generally, higher NDVI values, which reflect denser vegetation such as in Monsanto Forest Park, positively influence sentiment. Conversely, areas with lower NDVI values, common in residential zones with minimal vegetation, tend to reduce the predicted positive sentiment. Figure 11 confirms this trend, indicating that an NDVI value of 0.0319 representing sparse or non-vegetated surfaces is linked to a decline in positive sentiment in the model’s output for this case.

Figure 11

4 Discussion

Generally positive sentiments prevail over negative ones. This comes in line to other works where sentiment analysis of online reviews in Portugal, such as on TripAdvisor and Twitter, indicated that most comments were positive, with prevalent emotions like joy and analytical responses. Negative feelings like anger or sadness were uncommon, possibly reflecting a more constructive or serious attitude toward online feedback (98).

The results indicate that proximity to cycling infrastructure and fitness facilities plays a significant role in shaping positive sentiments across Lisbon. These findings are consistent with previous studies highlighting associations between active mobility, physical activity and subjective well-being (99–101).

Cycling infrastructure is widely discussed as a lever for shifting urban travel toward active modes and reducing car dependence, particularly when implemented as connected, protected networks (102, 103). Evidence also suggests that protected cycling facilities are associated with lower injury risk compared with cycling in mixed traffic, supporting their relevance for perceived and objective safety (104). The SHAP results suggest that shorter distances to cycling infrastructure are generally linked to higher predicted positive sentiment, reinforcing the role of accessible active transport systems in enhancing urban well-being.

From an urban-planning perspective, incorporating sports and exercise infrastructure may support recreational activity, community engagement, and social interaction (105, 106). Access to fitness and sports facilities has been associated with higher levels of physical activity, supporting active lifestyles (107, 108). Regular exercise has also been associated with improved cardiovascular health, reduced risk of chronic diseases, and positive mental-health benefits, including reduced anxiety and improvements in mood and self-esteem (109). However, the spatial variability observed in SHAP contributions suggests that the relationship between fitness infrastructure and sentiment may be context-dependent, potentially reflecting differences in accessibility, quality, or surrounding urban conditions.

Vegetation, as captured by NDVI, emerges as a key factor influencing sentiment patterns. Higher NDVI values are associated with increased predicted positive sentiment, particularly in areas with dense vegetation such as Monsanto Forest Park. These findings are consistent with studies linking mental health and well-being with vegetation and positive sentiment (2, 110). One possible explanation is that NDVI may influence well-being through multiple pathways, including promoting physical activity and social interaction. However, it is important to note that NDVI captures overall vegetation coverage rather than accessibility or quality, and therefore may not fully reflect how green spaces are experienced or used by residents (110). Portuguese demonstrate unique emotional patterns in their comments. They tend to express positive feelings toward natural and ecological environments but are less detailed in their preferences than other linguistic groups (111). Furthermore, Portuguese social media users exhibit emotional intensity in their reactions to events, as in other cultures, although gender differences in emotional expression are less evident than in English-speaking contexts (112).

According to Rodrigues et al., in Lisbon, the distribution of environmental resources associated with well-being appears to be spatially uneven, while exposure to environmental risks is often concentrated in specific urban areas. Vegetation is often concentrated in large parks located outside the most central and densely built areas, while central districts tend to combine high accessibility with greater exposure to environmental stressors such as air pollution and noise, particularly in areas surrounding the airport (113).

More broadly, the findings of this study highlight the importance of integrating active mobility infrastructure and green environments in urban well-being. By combining machine learning predictions with uncertainty quantification and explainability techniques, this work provides a more transparent and spatially explicit framework for understanding sentiment–environment relationships. Such an approach can support more informed and evidence-based urban planning decisions, particularly in identifying areas where interventions may have the greatest impact.

5 Conclusion

This study demonstrates the potential of combining sentiment analysis, machine learning, uncertainty quantification, and explainable AI to better understand the relationship between urban environments and well-being. By integrating spatial prediction with entropy-based uncertainty measures and SHAP explanations, the proposed framework provides a more transparent and interpretable approach to analyzing sentiment patterns in urban contexts.

Across Lisbon, the results consistently highlighted proximity to cycling infrastructure and higher vegetation cover (NDVI) as prominent predictors associated with positive sentiment. These findings reinforce the relevance of active mobility and green environments in shaping urban sentiment, which is closely linked to subjective well-being. By promoting active lifestyles and enhancing access to nature through urban infrastructure, cities can create conditions that support a more favorable emotional environment for residents.

Moreover, the choice of sentiment-analysis approach is critical, because the labels assigned to posts directly shape the subsequent machine-learning pipeline. In this study, several locations were associated with elevated levels of negative sentiment in the mapped outputs (Figures 3–4). This pattern is plausibly influenced by limitations of lexicon-based labeling, which does not account for contextual meaning. Such mislabeling can propagate into model training and bias spatial interpretations. Future work should therefore explore alternative sentiment-classification approaches, including supervised and contextual language models, and compare how sensitive downstream spatial patterns are to the labeling method.

Finally, machine-learning models can support urban planners, policymakers, and public-health professionals by identifying patterns in large and heterogeneous urban datasets and translating them into actionable, spatially explicit insights. However, when model outputs are used to inform interventions, it is essential to report not only point predictions but also the uncertainty associated with them. Prediction uncertainty provides an explicit measure of confidence in each estimate, helping distinguish areas where results are comparatively robust from areas where additional data, validation, or cautious interpretation are needed. Incorporating uncertainty analysis into the workflow strengthens evidence-informed decision-making by clarifying the reliability of model-based recommendations in complex and dynamic urban settings. The use of spatial prediction, uncertainty mapping, and model-agnostic explanations increases the transparency of sentiment analysis, enabling more confident identification of areas where sentiment patterns and their drivers can be inferred.

References

• 1.

  FathiSSajadzadehHSheshkalFMAramFPinterGFeldeIet al. The role of urban morphology design on enhancing physical activity and public health. Int J Environ Res Public Health. (2020) 17:1–29. doi: 10.3390/ijerph17072359,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2.

  HeHSunRLiJLiW. Urban landscape and climate affect residents’ sentiments based on big data. Appl Geogr. (2023) 152:102902. doi: 10.1016/j.apgeog.2023.102902

  • CrossRef
  • Google Scholar

• 3.

  MeltonC. The Drivers of Health: 4 Factors that Influence Individual and Population Health. Tennessee: (2018). Available online at: https://sycamoretn.org/wp-content/uploads/2018/07/2018.07.18-final-drivers-of-health.pdf

  • Google Scholar

• 4.

  GuHYanWElahiECaoY. Air pollution risks human mental health: an implication of two-stages least squares estimation of interaction effects. Environ Sci Pollut Res Int. (2020) 27:2036–43. doi: 10.1007/s11356-019-06612-x,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5.

  HaoYLiuSRuiXHuXWangLWangSet al. Exploring the psychological impacts of air contamination: a cross-national study. J Environ Manag. (2025) 390:126207. doi: 10.1016/j.jenvman.2025.126207,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6.

  JangEChoiHBKimM. The restorative effects of urban parks on stress control ability and community attachment. Sustainability. (2024) 16:2113. doi: 10.3390/su16052113

  • CrossRef
  • Google Scholar

• 7.

  ZhaoZHeQZhangYLiuSYangY. Assessing cross cultural urban park emotional wellbeing impact in Shanghai and London. Sci Rep. (2025) 15:18892. doi: 10.1038/s41598-025-03599-z,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8.

  FanCGaiZLiSCaoYGuYJinCet al. Does the built environment of settlements affect our sentiments? A multi-level and non-linear analysis of Xiamen, China, using social media data. Front Public Health. (2023) 10:1094036. doi: 10.3389/fpubh.2022.1094036,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9.

  ChatterjeeMModakSSarkarD. "Mental health predictions through online social media analytics". In: BhowmickPDasSMazumdarK, editors. Cognitive Cardiac Rehabilitation. Hershey, PA: Using IoT and AI Tools (2023). p. 44–66.

  • Google Scholar

• 10.

  IlievaRTMcPhearsonT. Social-media data for urban sustainability. Nat Sustain. (2018) 1:553–65. doi: 10.1038/s41893-018-0153-6

  • CrossRef
  • Google Scholar

• 11.

  YangLMarmolejo DuarteCMartí CiriquiánP. Quantifying the relationship between public sentiment and urban environment in Barcelona. Cities. (2022) 130:103977. doi: 10.1016/j.cities.2022.103977

  • CrossRef
  • Google Scholar

• 12.

  VianaCMSantosMFreireDAbrantesPRochaJ. Evaluation of the factors explaining the use of agricultural land: a machine learning and model-agnostic approach. Ecol Indic. (2021) 131:108200. doi: 10.1016/j.ecolind.2021.108200

  • CrossRef
  • Google Scholar

• 13.

  CartoneAPostiglioneP. Principal component analysis for geographical data: the role of spatial effects in the definition of composite indicators. Spat Econ Anal. (2021) 16:126–47. doi: 10.1080/17421772.2020.1775876

  • CrossRef
  • Google Scholar

• 14.

  DemšarUHarrisPBrunsdonCFotheringhamASMcLooneS. Principal component analysis on spatial data: an overview. Ann Assoc Am Geogr. (2013) 103:106–28. doi: 10.1080/00045608.2012.689236

  • CrossRef
  • Google Scholar

• 15.

  JombartTDevillardSDufourA-BPontierD. Revealing cryptic spatial patterns in genetic variability by a new multivariate method. Heredity. (2008) 101:92–103. doi: 10.1038/hdy.2008.34,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 16.

  LimaARCannonAJHsiehWW. Nonlinear regression in environmental sciences using extreme learning machines: a comparative evaluation. Environ Model Softw. (2015) 73:175–88. doi: 10.1016/j.envsoft.2015.08.002

  • CrossRef
  • Google Scholar

• 17.

  RenXMiZGeorgopoulosPG. Comparison of machine learning and land use regression for fine scale spatiotemporal estimation of ambient air pollution: modeling ozone concentrations across the contiguous United States. Environ Int. (2020) 142:105827. doi: 10.1016/j.envint.2020.105827,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18.

  ShortridgeJEGuikemaSDZaitchikBF. Machine learning methods for empirical streamflow simulation: a comparison of model accuracy, interpretability, and uncertainty in seasonal watersheds. Hydrol Earth Syst Sci. (2016) 20:2611–28. doi: 10.5194/hess-20-2611-2016

  • CrossRef
  • Google Scholar

• 19.

  MolnarC. Interpretable Machine Learning: A Guide for Making Black Box Models Explainable. Munich, Germany: Independently published (2019).

  • Google Scholar

• 20.

  CracknellMJReadingAM. Geological mapping using remote sensing data: a comparison of five machine learning algorithms, their response to variations in the spatial distribution of training data and the use of explicit spatial information. Comput Geosci. (2014) 63:22–33. doi: 10.1016/j.cageo.2013.10.008

  • CrossRef
  • Google Scholar

• 21.

  GiraudC. Introduction to High-Dimensional Statistics. 2nd ed. New York: Chapman & Hall/CRC (2021).

  • Google Scholar

• 22.

  CoulstonJWBlinnCEThomasVAWynneRH. Approximating prediction uncertainty for random forest regression models. Photogramm Eng Remote Sens. (2016) 82:189–97. doi: 10.14358/PERS.82.3.189

  • CrossRef
  • Google Scholar

• 23.

  CarvalhoDVPereiraEMCardosoJS. Machine learning interpretability: a survey on methods and metrics. Electronics. (2019) 8:832. doi: 10.3390/electronics8080832

  • CrossRef
  • Google Scholar

• 24.

  MurdochWJSinghCKumbierKAbbasi-AslRYuB. Definitions, methods, and applications in interpretable machine learning. Proc Natl Acad Sci. (2019) 116:22071–80. doi: 10.1073/pnas.1900654116,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25.

  RibeiroM. T.SinghS.GuestrinC. (2016). Model-Agnostic Interpretability of Machine Learning. ICML Workshop on Human Interpretability in Machine Learning

  • Google Scholar

• 26.

  RibeiroM. T.SinghS.GuestrinC. (2016). “Why should i trust you? explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1135–1144. doi: 10.48550/arXiv.1602.04938

  • CrossRef
  • Google Scholar

• 27.

  FouedjioFKlumpJ. Exploring prediction uncertainty of spatial data in geostatistical and machine learning approaches. Environ Earth Sci. (2019) 78:1–24. doi: 10.1007/s12665-018-8032-z

  • CrossRef
  • Google Scholar

• 28.

  KirkwoodCCaveMBeamishDGrebbySFerreiraA. A machine learning approach to geochemical mapping. J Geochem Explor. (2016) 167:49–61. doi: 10.1016/j.gexplo.2016.05.003

  • CrossRef
  • Google Scholar

• 29.

  VaysseKLagacherieP. Using quantile regression forest to estimate uncertainty of digital soil mapping products. Geoderma. (2017) 291:55–64. doi: 10.1016/j.geoderma.2016.12.017

  • CrossRef
  • Google Scholar

• 30.

  HomerCGHuangCYangLWylieBKCoanM. Development of a 2001 National Land Cover Database for the United States. Photogramm Eng Remote Sens. (2004) 70:829–40. doi: 10.14358/PERS.70.7.829

  • CrossRef
  • Google Scholar

• 31.

  CoulstonJWMoisenGGWilsonBTFincoMVCohenWBBrewerCK. Modeling percent tree canopy cover: a pilot study. Photogrammetric Engineering and Remote Sensing. (2012) 78:715–27. doi: 10.14358/PERS.78.7.715

  • CrossRef
  • Google Scholar

• 32.

  MoisenGGFrescinoTS. Comparing five modelling techniques for predicting forest characteristics. Ecol Model. (2002) 157:209–25. doi: 10.1016/S0304-3800(02)00197-7

  • CrossRef
  • Google Scholar

• 33.

  WielandMPittoreM. Performance evaluation of machine learning algorithms for urban pattern recognition from multi-spectral satellite images. Remote Sens. (2014) 6:2912–39. doi: 10.3390/rs6042912

  • CrossRef
  • Google Scholar

• 34.

  HePYuBMaJLuoKChenSShenZ. Exploring the non-linear relationship and synergistic effect between urban built environment and public sentiment integrating macro- and micro-level perspective: a case study in San Francisco. Front Psychol. (2024) 15:1276923. doi: 10.3389/fpsyg.2024.1276923,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35.

  LuZLuYChenYChenS. Effects of streetscapes on residents’ sentiments during heatwaves in Shanghai: evidence from multi-source data and interpretable machine learning for urban sustainability. Sustainability. (2025) 17:10281. doi: 10.3390/su172210281

  • CrossRef
  • Google Scholar

• 36.

  KongLLiuZPanXWangYGuoXWuJ. How do different types and landscape attributes of urban parks affect visitors’ positive emotions?Landsc Urban Plan. (2022) 226:104482. doi: 10.1016/j.landurbplan.2022.104482

  • CrossRef
  • Google Scholar

• 37.

  WangSLiangCGaoYYeYQiuJTaoCet al. Social media insights into spatio-temporal emotional responses to COVID-19 crisis. Health Place. (2024) 85:103174. doi: 10.1016/j.healthplace.2024.103174,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38.

  GaoYChenYMuLGongSZhangPLiuY. Measuring urban sentiments from social media data: a dual-polarity metric approach. J Geogr Syst. (2022) 24:199–221. doi: 10.1007/s10109-021-00369-z

  • CrossRef
  • Google Scholar

• 39.

  MelloCCheemaGSThakkarG. Combining sentiment analysis classifiers to explore multilingual news articles covering London 2012 and Rio 2016 Olympics. Int J Digit Humanit. (2023) 5:131–57. doi: 10.1007/s42803-022-00052-9,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 40.

  DGT (Direção-Geral do Território) (2018) Carta de Uso e Ocupação do Solo (COS) versão 2 - Série 1 (antiga). Available online at: https://www.dgterritorio.gov.pt/Carta-de-Uso-e-Ocupacao-do-Solo-para-2018

  • Google Scholar

• 41.

  INE (2021) População residente (N.^o) por Local de residência à data dos Censos [2021] (NUTS - 2024 - geografia 2025), Sexo, Grupo etário, Condição perante o trabalho e Escalão de dimensão populacional (do lugar); Decenal - INE, Recenseamento da população e habitação. Available online at: https://www.ine.pt/xportal/xmain?xpid=INE&xpgid=ine_base_dados&contexto=bd&bdtemas=1115&bdsubtemas=111514&bdfreetext=habit&xlang=pt

  • Google Scholar

• 42.

  WangW-JHaaseTWYangC-H. Warning message elements and retweet counts: an analysis of tweets sent during hurricane Irma. Nat Hazards Rev. (2020) 21:1–11. doi: 10.1061/(asce)nh.1527-6996.0000351

  • CrossRef
  • Google Scholar

• 43.

  KlašnjaMBarberáPBeauchampNNaglerJTuckerJA. "Measuring public opinion with social media data". In: AtkesonLRAlvarezRM, editors. Oxford: The Oxford Handbook of Polling and Polling Methods (2015)

  • Google Scholar

• 44.

  MenaouerBFairouzSMeriemMBMohammedSNadaM. A sentiment analysis of the Ukraine-Russia war tweets using knowledge graph convolutional networks. Int J Inf Technol. (2025). doi: 10.1007/s41870-024-02357-0

  • CrossRef
  • Google Scholar

• 45.

  PawarRChinchmalatpurePChinchmalatpureSPujeriRVWategaonkarDNJadhavNet al. Social threat identification on X via sentiment analysis: an SVM and LSTM-based approach. Int J Appl Math. (2025) 38:565–81. doi: 10.12732/ijam.v38i3s.167

  • CrossRef
  • Google Scholar

• 46.

  SinghASinghAKumariAChauhanASrivastavaR. "Comparative analysis of machine learning models for sentiment analysis on X (twitter) dataset". In: SachanVMalikSGautamRKumarP, editors. Advances in AI for Biomedical Instrumentation, Electronics and Computing. Boca Raton, FL: CRC Press (2024). p. 600–6.

  • Google Scholar

• 47.

  TracyA.KlucikR.Javernick-WillA.PoleacovschiC. (2018). New disasters in the twittersphere: how communities utilize social media to seek and share information in the wake of induced seismicity. Construction Research Congress 2018: Safety and Disaster Management - Selected Papers from the Construction Research Congress 2018, 2018-April, 524–534.

  • Google Scholar

• 48.

  DoSTNguyenVTBanlasanD. Social media sensing framework for urban infrastructure management: a Philippine case study. Constr Innov. (2024) 24:1117–36. doi: 10.1108/CI-04-2022-0082

  • CrossRef
  • Google Scholar

• 49.

  Encalada-AbarcaLFerreiraCCRochaJ. Revisiting city tourism in the longer run: an exploratory analysis based on LBSN data. Curr Issues Tour. (2023) 27:584–99. doi: 10.1080/13683500.2023.2182669

  • CrossRef
  • Google Scholar

• 50.

  BurtonSHTannerKWGiraud-CarrierCGWestJHBarnesMD. “Right time, right place” health communication on twitter: value and accuracy of location information. J Med Internet Res. (2012) 14:e156. doi: 10.2196/jmir.2121,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51.

  Koed MadsenA. Digital methods as ‘experimental a priori’ – how to navigate vague empirical situations as an operationalist pragmatist. Convergence. (2024) 30:94–115. doi: 10.1177/13548565221144260

  • CrossRef
  • Google Scholar

• 52.

  MadsenAKGrundtvigAThorsenS. Soft city sensing: a turn to computational humanities in data-driven urbanism. Cities. (2022) 126:103671. doi: 10.1016/j.cities.2022.103671

  • CrossRef
  • Google Scholar

• 53.

  TantoushMSenguptaUHanL. Exploring city dynamics through tweets: a framework for capturing urban activities as complex spatiotemporal patterns. Cities. (2025) 162:105894. doi: 10.1016/j.cities.2025.105894

  • CrossRef
  • Google Scholar

• 54.

  MedhatWHassanAKorashyH. Sentiment analysis algorithms and applications: a survey. Ain Shams Eng J. (2014) 5:1093–113. doi: 10.1016/j.asej.2014.04.011

  • CrossRef
  • Google Scholar

• 55.

  ZunicACorcoranPSpasicI. Sentiment analysis in health and well-being: systematic review. JMIR Med Inform. (2020) 8:e16023. doi: 10.2196/16023,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 56.

  MohammadSMTurneyPD. Crowdsourcing a word–emotion association lexicon. Comput Intell. (2013) 29:436–65. doi: 10.1111/j.1467-8640.2012.00460.x

  • CrossRef
  • Google Scholar

• 57.

  BetcoIRibeiroAIValeDSEncalada-AbarcaLVianaCMRochaJ. Sentiment analysis using a lexicon-based approach in Lisbon, Portugal. Geospat Health. (2025) 20:1344. doi: 10.4081/gh.2025.1344,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58.

  Encalada-AbarcaLBoavida-PortugalICardoso FerreiraCRochaJ. Identifying tourist places of interest based on digital imprints: towards a sustainable smart city. Sustainability. (2017) 9:2317. doi: 10.3390/su9122317

  • CrossRef
  • Google Scholar

• 59.

  BasiriM EKabiriA. (2017). Translation is not enough: comparing lexicon-based methods for sentiment analysis in Persian. 2017 International Symposium on Computer Science and Software Engineering Conference (CSSE), 36–41.

  • Google Scholar

• 60.

  Moreno-OrtizAHernándezCP. Lexicon-based sentiment analysis of twitter messages in Spanish. Procesamiento Del Lenguaje Natural. (2013) 50:93–100. Available online at: http://journal.sepln.org/sepln/ojs/ojs/index.php/pln/article/view/4664

  • Google Scholar

• 61.

  BasiriMEKabiriA. Words are important: improving sentiment analysis in the Persian language by lexicon refining. ACM Trans Asian Low-Resour Lang Inf Process. (2018) 17:1–18. doi: 10.1145/3195633

  • CrossRef
  • Google Scholar

• 62.

  AlmeidaT. G.SouzaB. A.MenezesA. A. F.FigueiredoC. M. S.NakamuraE. F. (2016). Sentiment analysis of Portuguese comments from foursquare. WebMedia 2016 - Proceedings of the 22nd Brazilian Symposium on Multimedia and the Web, 355–358

  • Google Scholar

• 63.

  VillalobosC. M.ForeroM. L. A.De MelloH. D.CavalcantiM. P.ValenciaC. H.Orjuela-CañonA. D. (2022). Sentimental analysis on social media comments with recurring models and Pretrained word Embeddings in Portuguese. ACM International Conference Proceeding Series, 205–209.

  • Google Scholar

• 64.

  MachadoMTPardoTASRuizEES. "Creating a Portuguese context sensitive lexicon for sentiment analysis". In: VillavicencioAMoreiraVAbadACaseliHGamalloPRamischCet al, editors. BT - Computational Processing of the Portuguese Language. PROPOR 2018. Lecture Notes in Computer Science. Cham: Springer International Publishing (2018). p. 335–44.

  • Google Scholar

• 65.

  HajrasoulihAdel RioVFrancisJEdmondsonJ. Urban form and mental wellbeing. J Urban Design Mental Health. (2018) 5:10. Available online at: https://www.urbandesignmentalhealth.com/journal-5---urban-form-and-mental-wellbeing.html

  • Google Scholar

• 66.

  Von SzombathelyMAlbrechtMAntanaskovicDAugustinJAugustinMBechtelBet al. A conceptual modeling approach to health-related urban well-being. Urban Sci. (2017) 1:17. doi: 10.3390/urbansci1020017

  • CrossRef
  • Google Scholar

• 67.

  BetcoIRochaJ. A relação entre o ambiente urbano e o bem-estar: Análise em Lisboa, Portugal, utilizando redes sociais. Rev Cartogr. (2024) 108:7–28. doi: 10.35424/rcarto.i108.4496

  • CrossRef
  • Google Scholar

• 68.

  DohooIRDucrotCFourichonCDonaldAHurnikD. An overview of techniques for dealing with large numbers of independent variables in epidemiologic studies. Prev Vet Med. (1997) 29:221–39. doi: 10.1016/S0167-5877(96)01074-4,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 69.

  LinF-J. Solving multicollinearity in the process of fitting regression model using the nested estimate procedure. Qual Quant. (2008) 42:417–26. doi: 10.1007/s11135-006-9055-1

  • CrossRef
  • Google Scholar

• 70.

  NaimiBHammNASGroenTASkidmoreAKToxopeusAG. Where is positional uncertainty a problem for species distribution modelling?Ecography. (2014) 37:191–203. doi: 10.1111/j.1600-0587.2013.00205.x

  • CrossRef
  • Google Scholar

• 71.

  JamesGWittenDHastieTTibshiraniR. Springer Texts in Statistics An Introduction to Statistical Learning. 7th ed. New York: Springer (2013).

  • Google Scholar

• 72.

  JohnstonRJonesKManleyD. Confounding and collinearity in regression analysis: a cautionary tale and an alternative procedure, illustrated by studies of British voting behaviour. Qual Quant. (2018) 52:1957–76. doi: 10.1007/s11135-017-0584-6,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 73.

  ShannonCE. A mathematical theory of communication. Bell Syst Tech J. (1948) 27:379–423. doi: 10.1002/j.1538-7305.1948.tb01338.x

  • CrossRef
  • Google Scholar

• 74.

  FloresASJ. Bits, bans y nats: unidades de medida de cantidad de información. Gandia: Universitat Politècnica de València (2017).

  • Google Scholar

• 75.

  DewiCChenR-CChristantoHJCauteruccioF. Multinomial naïve Bayes classifier for sentiment analysis of internet movie database. Vietnam J Comput Sci. (2023) 10:485–98. doi: 10.1142/S2196888823500100

  • CrossRef
  • Google Scholar

• 76.

  GandhiB. S.KumarS.VictorA.SelvanambiR. (2022). Transfer learning using BERT & Comparative Analysis of ML algorithms for opinion mining. 2022 IEEE 3rd Global Conference for Advancement in Technology, GCAT 2022.

  • Google Scholar

• 77.

  ShettyS. M.PushpaD. (2024). An overview of the literature on sentiment analysis methods for online and social media platforms. 2024 12th International Conference on Intelligent Systems and Embedded Design, ISED 2024.

  • Google Scholar

• 78.

  MullerT. An unforeseen equivalence between uncertainty and entropy. IFIP Adv Inform Commun Technol. (2019) 563:57–72. doi: 10.1007/978-3-030-33716-2_5

  • CrossRef
  • Google Scholar

• 79.

  SasonI. (2012). On the entropy of sums of Bernoulli random variables via the Chen-stein method. 2012 IEEE Information Theory Workshop, ITW 2012, 542–546

  • Google Scholar

• 80.

  Valverde-AlbaceteFJCarrillo-de-AlbornozJPeláez-MorenoC. A proposal for new evaluation metrics and result visualization technique for sentiment analysis tasks. Lecture Notes Computer Sci. (2013) 8138:41–52. doi: 10.1007/978-3-642-40802-1_5

  • CrossRef
  • Google Scholar

• 81.

  AslamA.QamarU.SaqibP.AyeshaR.QadeerA. (2020). A novel framework for sentiment analysis using deep learning. International Conference on Advanced Communication Technology, ICACT, 2020.

  • Google Scholar

• 82.

  ShuklaA.AroraD. (2023). Deep learning model for identification and classification of web based toxic comments. 2023 International Conference in Advances in Power, Signal, and Information Technology, APSIT 2023, 274–279

  • Google Scholar

• 83.

  DuruISunarAS. Transformer and pre-transformer model-based sentiment prediction with various Embeddings: a case study on Amazon reviews. Entropy. (2025) 27:1202. doi: 10.3390/e27121202,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 84.

  AroraDTayalDKYadavSK. Entropy-driven sentiment rating generation for online textual reviews with type-2 fuzzy sets. Int J Inf Technol. (2025) 17:4631–41. doi: 10.1007/s41870-025-02656-0

  • CrossRef
  • Google Scholar

• 85.

  BishtRK. "Entropy-modulated sentiment distribution law: a novel approach to modeling sentiment dynamics in a review". In: KumarSByeR. TPrasadM, editors. Singapore: Lecture Notes in Networks and Systems (2026)

  • Google Scholar

• 86.

  LundbergSMLeeS. A Unified Approach to Interpreting. Long Beach, CA: Model Predictions. (2017). p. 1–10.

  • Google Scholar

• 87.

  FattoriniLFranceschiSMarcheselliMPisaniCPratelliL. Design-based spatial interpolation with data driven selection of the smoothing parameter. Environ Ecol Stat. (2023) 30:103–29. doi: 10.1007/s10651-023-00555-w

  • CrossRef
  • Google Scholar

• 88.

  LiZZhangXZhuRZhangZWengZ. Integrating data-to-data correlation into inverse distance weighting. Comput Geosci. (2020) 24:203–16. doi: 10.1007/s10596-019-09913-9

  • CrossRef
  • Google Scholar

• 89.

  YanJWuBHeQ. An anisotropic IDW interpolation method with multiple parameters cooperative optimization. Cehui Xuebao/Acta Geodaetica et Cartographica Sinica. (2021) 50:675–84. doi: 10.11947/j.AGCS.2021.20200148

  • CrossRef
  • Google Scholar

• 90.

  ZhaoZLuoSZhaoXZhangJLiSLuoYet al. A novel interpolation method for soil parameters combining RBF neural network and IDW in the Pearl River Delta. Agronomy. (2024) 14:2469. doi: 10.3390/agronomy14112469

  • CrossRef
  • Google Scholar

• 91.

  KumarI. E.VenkatasubramanianS.ScheideggerC.FriedlerS. A. (2020). Problems with Shapley-value-based explanations as feature importance measures. 37th International Conference on Machine Learning, ICML 2020, PartF16814, 5447–5456.

  • Google Scholar

• 92.

  BirthriyaSKAhlawatPJainAK. Multi-objective feature selection for phishing detection using a hybrid Nash equilibrium and LightGBM model. IETE J Res. (2025) 71:2686–99. doi: 10.1080/03772063.2025.2498608

  • CrossRef
  • Google Scholar

• 93.

  BrandtF.BrillM.FischerF.HarrensteinP. (2009). Computational aspects of shapley’s saddles. Proceedings of the International Joint Conference on Autonomous Agents and Multiagent Systems, AAMAS, 1, 181–188.

  • Google Scholar

• 94.

  JavedD.JhanjhiN. Z.KhanN. A. (2023). Explainable twitter bot detection model for limited features. IET Conference Proceedings, 2023(11), 476–481

  • Google Scholar

• 95.

  ZhangKFengLYuX. Shap-PreBiNT: a sentiment analysis model based on optimized transformer. Lecture Notes Computer Sci. (2023):444–58. doi: 10.1007/978-3-031-25198-6_33

  • CrossRef
  • Google Scholar

• 96.

  KuhnM. (2019). The caret Package Available online at: https://topepo.github.io/caret/index.html (Accessed January 20, 2026).

  • Google Scholar

• 97.

  CasasP. (2019) A Gentle Introduction to SHAP Values in R. Available online at: https://www.r-bloggers.com/2019/03/a-gentle-introduction-to-shap-values-in-r/

  • Google Scholar

• 98.

  BredaZCostaRDinisGMartinsAA. "eWOW of guests regarding their hotel experience: sentiment analysis of TripAdvisor reviews". In: Information Resources Management Association, editor. Research Anthology on Implementing Sentiment Analysis across Multiple Disciplines, vol. 1–4. Hershey, PA: IGI Global Scientific Publishing (2022). p. 1831–44.

  • Google Scholar

• 99.

  KroesenMvan WeeB. Understanding how accessibility influences health via active travel: results from a structural equation model. J Transp Geogr. (2022) 102:103379. doi: 10.1016/j.jtrangeo.2022.103379

  • CrossRef
  • Google Scholar

• 100.

  LamuANJbailyAVerguetSRobberstadBNorheimOF. Is cycle network expansion cost-effective? A health economic evaluation of cycling in Oslo. BMC Public Health. (2020) 20:1869–11. doi: 10.1186/s12889-020-09764-5,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 101.

  TorresECMde Picado-SantosLG. Using sentiment analysis to study the potential for improving sustainable mobility in university campuses. Sustainability (Switzerland). (2025) 17:6645. doi: 10.3390/su17146645

  • CrossRef
  • Google Scholar

• 102.

  BarbosaMRegoE. The transition to bicycle mobility: energetic impacts analysis the case of São Paulo. IEEE Lat Am Trans. (2017) 15:1691–4. doi: 10.1109/TLA.2017.8015074

  • CrossRef
  • Google Scholar

• 103.

  SoltaniAAllanAJavadpoorMLellaJ. Space syntax in analysing bicycle commuting routes in inner metropolitan Adelaide. Sustainability. (2022) 14:3485. doi: 10.3390/su14063485

  • CrossRef
  • Google Scholar

• 104.

  AdamsTAldredR. Cycling injury risk in London: impacts of road characteristics and infrastructure. Transport Findings. (2020). doi: 10.32866/001c.18226

  • CrossRef
  • Google Scholar

• 105.

  DavisAJMacCarronPCohenE. Social reward and support effects on exercise experiences and performance: evidence from parkrun. PLoS One. (2021) 16:1–18. doi: 10.1371/journal.pone.0256546,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 106.

  FurberSPomroyHGregoSTavener-SmithK. People’s experiences of using outdoor gym equipment in parks. Health Promot J Austr. (2014) 25:211. doi: 10.1071/HE14038,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 107.

  CopelandJLCurrieCWalkerAMasonEWilloughbyTNAmsonA. Fitness equipment in public parks: frequency of use and community perceptions in a small urban Centre. J Phys Act Health. (2017) 14:344–52. doi: 10.1123/jpah.2016-0277,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 108.

  LeeSAJuYJLeeJEHyunISNamJYHanK-Tet al. The relationship between sports facility accessibility and physical activity among Korean adults. BMC Public Health. (2016) 16:893. doi: 10.1186/s12889-016-3574-z,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 109.

  Health, D. of. Choosing Activity: a Physical Activity action plan. London: Department of Health Publications (2005).

  • Google Scholar

• 110.

  WangRHelbichMYaoYZhangJLiuPYuanYet al. Urban greenery and mental wellbeing in adults: cross-sectional mediation analyses on multiple pathways across different greenery measures. Environ Res. (2019) 176:108535. doi: 10.1016/j.envres.2019.108535,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 111.

  WangMYanYLiMZhouL. Differences in emotional preferences toward urban green spaces among various cultural groups in Macau and their influencing factors. Land. (2024) 13:414. doi: 10.3390/land13040414

  • CrossRef
  • Google Scholar

• 112.

  MacedoMSaxenaA. Gender biases in online communication: a case study of soccer. Appl Intell. (2026) 56:33. doi: 10.1007/s10489-025-06988-z

  • CrossRef
  • Google Scholar

• 113.

  RodriguesALGiannottiMMirandaBMorgadoP. Intra-urban inequalities in opportunities for pedestrian mental well-being: a GIScience framework using 15-minute isochrones and spatial clustering. J Geogr Syst. (2026). doi: 10.1007/s10109-025-00487-y

  • CrossRef
  • Google Scholar