Recommender systems act as skilled agents to assist users to conquer information overload while making selection decisions over items by providing customized recommendations. Users and items are general phrases denoting, respectively, entities actively browsing and making choices and entities being selected, such as goods and services.

Formally, recommender systems leverage one or more of the following three information sources to discover user preferences and generate recommendations: user-item interactions, side information, and contexts. User-item interactions, or ratings, are collected explicitly by prompting users to provide numerical feedback on items and are acquired implicitly by tracking user behaviors such as clicks, browsing time, or purchase history. The data are commonly represented as a matrix that encodes the preferences of users and is naturally sparse, since users normally interact with a limited fraction of items. Side information is rich information attached to an individual user or item that depicts user characteristics such as education and job or item properties such as description and product categories. Side information can take diverse structures with rich meaning, ranging over numerical status, texts, and images to videos, locations, and networks. On the other hand, context refers to all the information collected when a user interacts with an item, such as timestamps, locations, or textual reviews. This contextual information usually serves as an additional information source appended to the user-item interaction matrix.

The goal of recommender systems is to disclose unknown user preferences over items that users never interact with and recommend the most preferred items to them. In practice, recommender systems learn to generate recommendations based on three types of approaches: pointwise, pairwise, and listwise. The pointwise approach is the most common approach and demands recommendation systems to provide accurate numerical predictions on observed ratings. Items that a user never interacts with are then sorted by their rating predictions, and a number of items with the highest ratings are recommended to the user. On the other hand, a pairwise approach seeks to preserve the ordering of any pair of items based on ratings, while in the listwise approach, recommender systems aim to preserve the relative order of all rated items as a list for each user. The pairwise approach and listwise approach are together considered item rankings that only require recommender systems to output the ordering of items but not ratings for individual items.

In most of the works covered in this paper, the task of attribute-aware recommendation is defined as predicting unknown ratings, that is: given N_u users, N_i items, a user-item rating matrix R∈ℝNu×Ni with only a small portion N_r ratings observed (i.e., there are a total of N_u × N_i − N_r missing ratings), side information of users X∈ℝKX×Nu (assuming each user has K_X attributes), side information of items Y∈ℝKY×Ni, and contexts Z∈ℝKZ×Nr, the goal is to build a model that is capable of predicting each of the unknown ratings in R.

Then, given any specific user, a recommender system can make recommendations based on the predicted ratings. Normally, items with higher ratings are recommended first. Note that the dimension K_X, K_Y of side information attribute matrix X, Y might be zero, denoting that there is no side information about users or items. Likewise, if there is no contextual information about user-item interactions, K_Z will be zero.

The core techniques or algorithms for realizing recommender systems are generally classified into three categories: content-based filtering, collaborative filtering, and hybrid filtering (Bobadilla et al., 2013; Shi et al., 2014b; Isinkaye et al., 2015). Content-based filtering generates recommendations based on properties of items and user-item interactions. Content-based techniques exploit domain knowledge and seek to transform item properties in raw attribute structures such as texts, images, or locations into numerical item profiles. Each item is represented as a vector, and the matrix of side information of items Y is constructed. A representation of each user is then created by aggregating profiles of items that this user interacted with, and a similarity measure is leveraged to retrieve a number of the most similar items as recommendations. Note that content-based filtering does not require information from any other user to make recommendations. Collaborative filtering strives to identify a group of users with similar preferences based on past user-item interactions and recommends items preferred by these users. Since discovering users with common preferences is generally based on user-item ratings R, collaborative filtering becomes the first choice when item properties are inadequate in describing their content, such as movies or songs. Hybrid filtering is the extension or combination of content-based and collaborative filtering. Examples of this are building an ensemble of the two techniques, using the item rating history of collaborative filtering as part of the item profiles for content-based filtering or extending collaborative filtering to incorporate user characteristics X or item properties Y. This review focuses on attribute-aware recommender systems that shed light not only on user-item interactions R but also on the side information of users or items X, Y and contexts Z, which is a subset of hybrid filtering.

Collaborative filtering (CF) has become the prevailing technique for realizing recommender systems in recent years (Adomavicius and Tuzhilin, 2005, 2011; Shi et al., 2014b; Isinkaye et al., 2015). It assumes that the preferences users exhibit for items they have interacted with can be generalized and used to infer their preferences regarding items they have never interacted with by leveraging the records of other users with similar preferences. This section briefly introduces conventional CF techniques, which assume the availability of only user-item interactions or the rating matrix R. In practice, they are commonly categorized into memory-based CF and model-based CF (Adomavicius and Tuzhilin, 2005; Shi et al., 2014b; Isinkaye et al., 2015).

Memory-based CF directly exploits rows or columns in the rating matrix R as representations of users or items and identifies a group of similar users or items using a pre-defined similarity measure. Commonly used similarity metrics include the Pearson correlation, the Jaccard similarity coefficient, the cosine similarity, or their variants. Memory-based CF techniques can be divided into user-based and item-based approaches, which identify either a group of similar users or similar items, respectively. For user-based approaches, K nearest neighbors—or the K most similar users—are extracted, and their preferences or ratings regarding a target item are aggregated into a rating prediction using similarities between users as weights. The rating prediction for user u on item i, r^ui, can be formulated as:

where function sim(·) is a similarity measure, Z is the normalization constant, and U_u is the set of similar users to user u (Shi et al., 2014b). Rating predictions of item-based approaches can be formulated in a similar way. The calculated pairwise similarities between users or items act as the memory of the recommender system since they can be saved for generating later recommendations.

Model-based CF, on the other hand, takes the rating matrix R to train a predictive model with a set of parameters θ to make recommendations (Adomavicius and Tuzhilin, 2005; Shi et al., 2014b). Predictive models can be formulated as a function that outputs ratings for rating predictions or numerical preference scores for item ranking given a user-item pair (u, i):

Model-based CF then ranks and selects the K items with the highest ratings or scores r_ui as recommendations. Common core algorithms for model-based CF involve Bayesian classifiers, clustering techniques, graph-based approaches, genetic algorithms, and dimension-reduction methods such as Singular Value Decomposition (SVD) (Adomavicius and Tuzhilin, 2005, 2011; Bobadilla et al., 2013; Shi et al., 2014b; Isinkaye et al., 2015). Over the last decade, a class of latent factor models, called matrix factorization, has been popularized and is commonly adopted as the basis of advanced techniques because of its success in the development of algorithms for the Netflix competition (Koren et al., 2009; Koren and Bell, 2011). In general, latent factor models aim to learn a low-dimensional representation, or latent factor, for each entity and combine the latent factors of different entities using specific methods such as inner product, bilinear map, or neural networks to make predictions. As one of these latent factor models, matrix factorization for recommender systems characterizes each user and item by a low-dimensional vector and predicts ratings based on inner product.

Matrix factorization (MF) (Paterek, 2007; Koren et al., 2009; Koren and Bell, 2011; Shi et al., 2014b), in the basic form, represents each user u as a parameter vector wu∈ℝK and each item i as hi∈ℝK, where K is the dimension of latent factors. The prediction of user u's rating or preference regarding item i, denoted as r^ui, can be computed using the inner product:

which captures the interaction between them. MF seeks to generate rating predictions that are as close as possible to those recorded ratings. In matrix form, it can be written as finding W, H such that R ≈ W^⊤H where R∈ℝNu×Ni. MF is essentially learning a low-rank approximation of the rating matrix, since the dimension of representations K is usually much smaller than the number of users N_u and items N_i. To learn the latent factors of users and items, the system tries to find W, H that minimizes the regularized square error on the set of entries of known ratings in R [denoted as δ(R)]:

where λ_W and λ_H are regularization parameters. MF tends to cluster users or items with similar rating configurations into groups in the latent factor space, which implies that similar users or items will be close to each other. Furthermore, MF assumes that the rank of rating matrix R or the dimension of the vector space generated by the rating configuration of users is far smaller than the number of users N_u. This implies that each user's rating configuration can be obtained by a linear combination of ratings from a group of other users since they are all generated by K principal vectors. Thus, MF is in the spirit of collaborative filtering, which is to infer a user's unknown ratings by the ratings of several other users.

Biased matrix factorization (Paterek, 2007; Koren et al., 2009; Koren and Bell, 2011), an improvement of MF, models the characteristics of each user and each item and the global tendency that are independent of user-item interactions. The obvious drawback of MF is that only user-item interactions wu⊤hi are considered in rating predictions. However, ratings usually contain universal shifts or exhibit systematic tendencies with respect to users and items. For instance, there might be a group of users inclined to give significantly higher ratings than others or a group of items widely considered to be of high-quality and that receive higher ratings. Besides, it is common that all ratings are non-negative, which implies that the overall average might not be close to zero and causes a difficulty for training small-value-initialized representations. Due to the issues mentioned above, biased MF augments MF rating predictions with linear biases that account for user-related, item-related, and global effects. The rating prediction is extended as follows:

where μ, c_i, d_j are global bias, bias of user i, and bias of item j, respectively. Biased MF then finds the optimal W, H, c, d, μ that minimizes the regularized square error as follows:

where ||W||F2=∑u=1Nu||wu||22 denotes the squared Frobenius norm. The regularization parameter λ is tuned by cross-validation.

Probabilistic matrix factorization (PMF, Figure 2 and Biased PMF, Figure 3) (Salakhutdinov and Mnih, 2007, 2008) is a probabilistic linear model with observed Gaussian noise and can be viewed as a probabilistic extension of MF. PMF adopts the assumption that users and items are independent and represents each user or each item with a zero-mean spherical multivariate Gaussian distribution as follows:

where σW2 and σH2 are observed user-specific and item-specific noise. PMF then formulates the conditional probability over the observed ratings as

where δ(R) is the set of known ratings and N(x∣μ,σ2) denotes the Gaussian distribution with mean μ and variance σ². Learning in PMF is conducted by maximum a posteriori (MAP) estimation, which is equivalent to maximizing the log of the posterior distribution of W, H:

where C is a constant independent of all parameters and K is the dimension of user or item representations. With Gaussian noise σR2,σW2,σH2 observed, maximizing the log-posterior is identical to minimizing the objective function with the form:

where λW=σR2/σW2,λH=σR2/σH2. Note that (10) has exactly the same form as the regularized square error of MF, and gradient descent or its extensions can then be applied in training PMF.

Since collaborative filtering techniques only consider rating matrix R in making recommendations, they cannot discover the preferences of users or for items with scant user-item interactions. This problem is referred as the cold-start issue. In section 3, we will review recommendation systems that extend CF to incorporate contexts or rich side information regarding users and items to alleviate the cold-start problem.

Attribute-aware recommendation models are proposed to tackle the challenges of integrating additional information from user/item/rating. There are two strategies for designing attribute-aware collaborative filtering-based systems. One direction is to combine content-based recommendation models with CF models, which can directly accept attributes as content to perform recommendation. On the other hand, researchers also try to extend an existing collaborative filtering algorithm such that it leverages attribute information.

Rather, we will focus on four important factors in the design of an attribute-aware recommender system in current research, as shown in Figure 4. They are specifically discussed from sections 3.2 to 3.5. With respect to input data, attribute sources determine whether an attribute vector is relevant to users, items, or ratings. For example, age describes a user instead of an item; rating time must be appended to ratings, representing when the rating event occurred. Different models impose distinct strategies to integrate attributes of specific sources. Additionally, a model may constrain attribute types that can be used. For instance, graph-based collaborative filtering realizations define attributes as node types, which is not appropriate for numerical attributes. Rating type is the factor that is emphasized by most model designers. Besides usual numerical ratings, many recommendation models concentrate on binary rating data, where the ratings represent whether users interact with items. Finally, different recommender systems emphasize different recommendation goals. One is to predict the ratings from users to items through minimizing the error between the predicted and real ratings. Another is to produce the ranking among items given a user instead of caring about the real rating value of a single item. We then summarize the design categories of all the surveyed papers in a table in section 3.6.

Throughout this paper, we will use X=[x1x2…xN]∈ℝK×N to denote the attribute matrix, where each column x_i represents a K-dimensional attribute vector of entity i. Here, an entity can refer to a user, an item, or a rating, determined by attribute sources (discussed in section 3.2). If attributes are limited categorically, then X ∈ {0, 1}^K×N can be represented by one-hot encoding (discussed in section 3.3). Note that our survey does not include models designed specifically for a certain type of attribute but rather covers models that are general enough to accept different types of attributes. For example, Collaborative Topic Regression (CTR) (Wang and Blei, 2011) extends matrix factorization with Latent Dirichlet Allocation (LDA) to import text attributes. Social Regularization (Ma et al., 2011a) specifically utilizes user social networks to regularize the learning of matrix factorization. Neither model is included since they are not general enough to deal with general attributes.

Attributes usually come from a variety of sources. Typically, side information refers to the attributes appended to users or items. In contrast, keyword contexts indicate the attributes relevant to ratings. Ratings from the same user can be attached to different contexts, such as “locations where users rate items.” Recommendation models considering rating-relevant attributes are usually called context-aware recommender systems. Although contexts in some papers could include user-relevant or item-relevant ones, in this paper, we tend to be precise and use the term contexts only for rating-relevant attributes.

Sections 3.2.1 and 3.2.2, respectively introduce different attribute sources. It is worth mentioning our observation as follows. Even though some of the models we surveyed demand side information while others require context information, we discover that the two sets of attributes can be represented in a unified manner and thus that both types of models can be applied. We will discuss such unified representation in sections 3.2.3 and 3.2.4.

In most cases, attribute-aware recommender systems accept a real-value attribute matrix X. However, we notice that some attribute-aware recommender systems require attributes to be categorical, which is typically represented by binary encoding. Specifically, this approach demands a binary attribute matrix where attributes of value 1 are modeled as discrete latent information in some way. A summary of the two types of attributes is given in sections 3.3.1 and 3.3.2.

It is trivial to put one-hot categorical attributes into numerical attribute-aware recommender systems, since binary values {0, 1} ⊂ ℝ. Nonetheless, putting numerical attributes into categorical attribute-aware recommendation systems runs the risk of losing attribute information (e.g., quantization processing).

Although we always define the term ratings as the interactions between users and items in this paper, some previous works claim a difference between explicit opinions and implicit feedback. Taking the dataset MovieLens, for example, a user gives a rating value in {1, 2, 3, 4, 5} toward an item. The value denotes the explicit opinion, which quantifies the preference of the user for that item. How recommendation methods handle such type of ratings will be introduced in section 3.4.1.

Even though modeling explicit opinions is more beneficial for future recommendation, such data is more difficult to gather from users. Users may hesitate to show their preferences due to privacy considerations, or they may not be willing to spend time labeling explicit ratings. Instead, recommender systems are more likely to collect implicit feedback, such as user browsing logs. Such datasets record a series of binary values, each of which implies whether a user ever saw an item. User preferences behind implicit feedback assume that the items seen by a user must be more preferred by the user than those items that have never been seen. We discuss this type of rating in detail in section 3.4.2.

Some numerical rating data are controversial, like “the number of times a user clicks on the hyperlink to visit the page of an item.” Some of the related work may define such data as implicit feedback because the number of clicks is not equivalent to explicit user preferences. However, in this paper, we still identify them as explicit opinions. With respect to model designs, related recommendation approaches do not differentiate such data.

Any recommender system needs human intervention to set up a training goal. Since collaborative filtering-based recommender systems rely on ratings, the most straightforward goal is to infer what rating will be given by a user for an unseen item, called rating prediction. If the ratings of every item can be accurately predicted, then for any user, a recommender system can just sort and recommend items based on highest predicted rating. In machine learning, such a goal for model-based recommender systems can be described as pointwise learning. That is, given a user-item pair, a pointwise learning recommendation model directly minimizes the error of predicted ratings and true ones. The related mathematical details are presented in section 3.5.1.

However, in general, our ultimate goal is to recommend unseen items to users without being concerned about how these items are rated. All unseen items in pointwise learning are finally ranked in descending order of their ratings. In other words, what we truly care about is the order of ratings, but not the true rating values. Also, some research papers figure out that a low error of rating prediction is not always equivalent to a high quality of recommended item lists. Recent model-based collaborative filtering models have begun to set optimization goals of item ranking. That is, for the same user, such models maximize the differences between high-rated items and low-rated ones in training data. The implementation of item ranking includes pairwise learning and listwise learning in machine-learning domains. Both learning ideas try to compare the potentially related ranks between at least two items for the same user. section 3.5.2 will present how to define optimization criteria for item ranking.

Having introduced the above categories for attribute-aware recommender systems, we list which category each publication we have surveyed belongs to in Tables 2–4. We trace back 10 years to summarize the recent trends in attribute-aware recommender systems.

Intuitively, the goal of an attribute-aware recommender system is to import attributes to improve recommendation performance (either rating prediction or item ranking). In the framework of matrix factorization, an item is rated or ranked according to the latent factors of the item and its corresponding users. In order words, the learning of latent factors in classical matrix factorization depends only on ratings. Thus, the learning may fail due to a lack of training ratings. If we can regularize the latent factors using attributes or make attributes determine how to rate items then matrix factorization methods can be more robust to compensate for the lack of rating information in the training data, especially for those users or items that have very few ratings. In the following, we choose to describe attribute participation from probabilistic perspectives. Learning in Probabilistic Matrix Factorization (PMF) tries to maximize the posterior probability p(W, H ∣ R) of two latent factor matrices W (for users) and H (for items) given observed entries of training rating matrix R. Clearly, attribute-aware recommender systems claim that we are given an extra attribute matrix X. Then by Bayes' rule, the posterior probability can be shown as follows:

We eliminate the denominator p(R ∣ X) since it does not contain variables W, H for maximization. At the prior part, we follow the independence assumption W⊥H of PMF, though here the independence is given attribute matrix X. Now, compared with classical PMF, both likelihood p(R ∣ W, H, X) and prior p(W ∣ X)p(H ∣ X) could be affected by attributes X. Attributes in the likelihood can directly help predict or rank ratings, while attributes in the priors regularize the learning directions of latent factors. Moreover, some current works assume additional independence between attributes and the matrix factorization formulation. We give graphical interpretation of Discriminative Matrix Factorization in Figure 5. For ease of explanation, we suppose that all the random variables follow a normal distribution p(x)=N(x∣μ,σ2) with mean μ and variance σ², or a multivariate normal distribution p(x)=N(x∣μ,Σ) with mean vector μ and covariance matrix Σ. Theoretically, the following models accept other probability distributions.

We further introduce the sub-categories below.