Recommender systems are intrinsically tied to a reliability/coverage dilemma: The more reliable we desire the forecasts, the more conservative the decision will be and thus, the fewer items will be ...recommended. This causes a detriment to the predictive capability of the system, as it is only able to estimate potential interest in items for which there is a consensus in their evaluation, rather than being able to estimate potential interest in any item. In this paper, we propose the inclusion of a new term in the learning process of matrix factorization-based recommender systems, called recklessness, that takes into account the variance of the output probability distribution of the predicted ratings. In this way, gauging this recklessness measure we can force more spiky output distribution, enabling the control of the risk level desired when making decisions about the reliability of a prediction. Experimental results demonstrate that recklessness not only allows for risk regulation but also improves the quantity and quality of predictions provided by the recommender system.
•Adjusting the variance in probability-based recommendation systems modifies the shape of the generated probabilities.•High variance produces many low-precision predictions, while low variance yields fewer high-precision predictions.•The variance can be adjusted by adding it to the cost function and tuning its weight with a hyper-parameter.
Recently, recommender systems have witnessed the fast evolution of Internet services. However, it suffers hugely from inherent bias and sparsity issues in interactions. The conventional uniform ...embedding learning policies fail to utilize the imbalanced interaction clue and produce suboptimal representations to users and items for recommendation. Towards the issue, this work is dedicated to bias-aware embedding learning in a decomposed manner and proposes a counterfactual graph convolutional learning (CGCL) model for personalized recommendation. Instead of debiasing with uniform interaction sampling, we follow the natural interaction bias to model users’ interests with a counterfactual hypothesis. CGCL introduces bias-aware counterfactual masking on interactions to distinguish the effects between majority and minority causes on the counterfactual gap. It forms multiple counterfactual worlds to extract users’ interests in minority causes compared to the factual world. Concretely, users and items are represented with a causal decomposed embedding of majority and minority interests for recommendation. Experiments show that the proposed CGCL is superior to the state-of-the-art baselines. The performance illustrates the rationality of the counterfactual hypothesis in bias-aware embedding learning for personalized recommendation.
In today’s digital landscape, recommender systems have gained ubiquity as a means of directing users towards personalized products, services, and content. However, despite their widespread adoption ...and a long track of research, these systems are not immune to shortcomings. A significant challenge faced by recommender systems is the presence of biases, which produces various undesirable effects, prominently the popularity bias. This bias hampers the diversity of recommended items, thus restricting users’ exposure to less popular or niche content. Furthermore, this issue is compounded when multiple stakeholders are considered, requiring the balance of multiple, potentially conflicting objectives.
In this paper, we present a new approach to address a wide range of undesired consequences in recommender systems that involve various stakeholders. Instead of adopting a consequentialist perspective that aims to mitigate the repercussions of a recommendation policy, we propose a deontological approach centered around a minimal set of ethical principles. More precisely, we introduce two distinct principles aimed at avoiding overconfidence in predictions and accurately modeling the genuine interests of users. The proposed approach circumvents the need for defining a multi-objective system, which has been identified as one of the main limitations when developing complex recommenders. Through extensive experimentation, we show the efficacy of our approach in mitigating the adverse impact of the recommender from both user and item perspectives, ultimately enhancing various beyond accuracy metrics. This study underscores the significance of responsible and equitable recommendations and proposes a strategy that can be easily deployed in real-world scenarios.
Collaborative Filtering (CF) is achieving a plateau of high popularity. Still, recommendation success is challenged by the diversity of user preferences, structural sparsity of user-item ratings, and ...inherent subjectivity of rating scales. The increasing user base and item dimensionality of e-commerce and e-entertainment platforms creates opportunities, while further raising generalization and scalability needs. Moved by the need to answer these challenges, user-based and item-based clustering approaches for CF became pervasive. However, classic clustering approaches assess user (item) rating similarity across all items (users), neglecting the rich diversity of item and user profiles. Instead, as preferences are generally simultaneously correlated on subsets of users and items, biclustering approaches provide a natural alternative, being successfully applied to CF for nearly two decades and synergistically integrated with emerging deep learning CF stances. Notwithstanding, biclustering-based CF principles are dispersed, causing state-of-the-art approaches to show accentuated behavioral differences. This work offers a structured view on how biclustering aspects impact recommendation success, coverage, and efficiency. To this end, we introduce a taxonomy to categorize contributions in this field and comprehensively survey state-of-the-art biclustering approaches to CF, highlighting their limitations and potentialities.
Learning how to represent users based on historical interactions is a crucial problem for recommender systems. Unavoidable noise in interactions and long-tail items composed of a large number of ...unpopular items bring more challenges for learning better user representations and still limit the performance of existing models. Aiming to design a simple model that can alleviate both the noise problem and the long-tail item problem, we propose a Hybrid Normalization strategy via feature statistics for Collaborative Filtering (HyNCF). After each user is represented by his/her interacted items, the feature statistics of a target user are mixed with that of another randomly sampled user. In addition, the uncertainty estimation of the target user’s feature statistics is calculated by a Gaussian sampling technique. Both kinds of improved feature statistics are separately used to normalize the target user’s embedding, and then normalized embeddings are aggregated to generate two representations of the user. Based on the fusion of the two representations, the cosine contrastive loss is used to train HyNCF. The effectiveness of the proposed model is evaluated on five benchmark datasets.
•Utilizing feature statistics in embedding space to learn user representations.•Designing mixing up and Gaussian sampling to improve feature statistics’ diversity.•Using improved feature statistics to normalize a user’s interacted items.
Recently, the user-side fairness issue in Collaborative Filtering (CF) algorithms has gained considerable attention, arguing that results should not discriminate an individual or a sub-user group ...based on users’ sensitive attributes (e.g., gender). Researchers have proposed fairness-aware CF models by decreasing statistical associations between predictions and sensitive attributes. A more natural idea is to achieve model fairness from a causal perspective. The remaining challenge is that we have no access to interventions, i.e., the counterfactual world that produces recommendations when each user has changed the sensitive attribute value. To this end, we first borrow the Rubin-Neyman potential outcome framework to define average causal effects of sensitive attributes. Next, we show that removing causal effects of sensitive attributes is equal to average counterfactual fairness in CF. Then, we use the propensity re-weighting paradigm to estimate the average causal effects of sensitive attributes and formulate the estimated causal effects as an additional regularization term. To the best of our knowledge, we are one of the first few attempts to achieve counterfactual fairness from the causal effect estimation perspective in CF, which frees us from building sophisticated causal graphs. Finally, experiments on three real-world datasets show the superiority of our proposed model.
Online interactive recommender systems strive to promptly suggest users appropriate items (e.g., movies and news articles) according to the current context including both user and item content ...information. Such contextual information is often unavailable in practice, where only the users' interaction data on items can be utilized by recommender systems. The lack of interaction records, especially for new users and items, inflames the performance of recommendation further. To address these issues, both collaborative filtering, one of the most popular recommendation techniques relying on the interaction data only, and bandit mechanisms, capable of achieving the balance between exploitation and exploration, are adopted into an online interactive recommendation setting assuming independent items (i.e., arms). This assumption rarely holds in reality, since the real-world items tend to be correlated with each other. In this paper, we study online interactive collaborative filtering problems by considering the dependencies among items. We explicitly formulate item dependencies as the clusters of arms in the bandit setting, where the arms within a single cluster share the similar latent topics. In light of topic modeling techniques, we come up with a novel generative model to generate the items from their underlying topics. Furthermore, an efficient particle-learning based online algorithm is developed for inferring both latent parameters and states of our model by taking advantage of the fully adaptive inference strategy of particle learning techniques. Additionally, our inferred model can be naturally integrated with existing multi-armed selection strategies in an interactive collaborative filtering setting. Empirical studies on two real-world applications, online recommendations on movies and news, demonstrate both the effectiveness and efficiency of our proposed approach.
Collaborative filtering (CF) approaches are widely applied in recommender systems. Traditional CF approaches have high costs to train the models and cannot capture changes in user interests and item ...popularity. Most CF approaches assume that user interests remain unchanged throughout the whole process. However, user preferences are always evolving and the popularity of items is always changing. Additionally, in a sparse matrix, the amount of known rating data is very small. In this paper, we propose a method of online collaborative filtering with dynamic regularization (OCF-DR), that considers dynamic information and uses the neighborhood factor to track the dynamic change in online collaborative filtering (OCF). The results from experiments on the MovieLens100K, MovieLens1M, and HetRec2011 datasets show that the proposed methods are significant improvements over several baseline approaches.
Recommender systems filter information to meet users’ personalized interests actively. Existing graph-based models typically extract users’ interests from a heterogeneous interaction graph. They do ...not distinguish learning between users and items, ignoring the heterogeneous property. In addition, the interaction sparsity and long-tail bias issues still limit the recommendation performance significantly. Fortunately, hidden homogeneous correlations that have a considerable volume can entangle abundant CF signals. In this paper, we propose a light dual hypergraph convolution (LDHC) for collaborative filtering, which designs a hypergraph to involve heterogeneous and homogeneous correlations with more CF signals confronting the challenges. Over the integrated hypergraph, a two-level interest propagation is performed within the heterogeneous interaction graph and between the homogeneous user/item graphs to model users’ interests, where learning on users and items is distinguished and collaborated by the homogeneous propagation. Specifically, hypergraph convolution is lightened by removing unnecessary parameters to propagate users’ interests. Extensive experiments on publicly available datasets demonstrate that the proposed LDHC outperforms the state-of-the-art baselines.
•Homogeneous user/item correlations bring entangled CF signals beyond the single CF signal in heterogeneous user-item interactions, which enables hybrid learning with much more reliable CF signals to alleviate interaction sparsity and confront long-tail bias.•This work proposes a light dual hypergraph convolution (LDHC) model by hybrid learning with a hypergraph convolution to predict interests, which performs a two-level interest propagation within the heterogeneous correlations and between the homogeneous ones.•A light convolution is introduced in hypergraph-based interest propagation to lighten the LDHC model for reducing the burden and difficulty in training.
In the era of big data, recommender system (RS) has become an effective information filtering tool that alleviates information overload for Web users. Collaborative filtering (CF), as one of the most ...successful recommendation techniques, has been widely studied by various research institutions and industries and has been applied in practice. CF makes recommendations for the current active user using lots of users' historical rating information without analyzing the content of the information resource. However, in recent years, data sparsity and high dimensionality brought by big data have negatively affected the efficiency of the traditional CF-based recommendation approaches. In CF, the context information, such as time information and trust relationships among the friends, is introduced into RS to construct a training model to further improve the recommendation accuracy and user's satisfaction, and therefore, a variety of hybrid CF-based recommendation algorithms have emerged. In this paper, we mainly review and summarize the traditional CF-based approaches and techniques used in RS and study some recent hybrid CF-based recommendation approaches and techniques, including the latest hybrid memory-based and model-based CF recommendation algorithms. Finally, we discuss the potential impact that may improve the RS and future direction. In this paper, we aim at introducing the recent hybrid CF-based recommendation techniques fusing social networks to solve data sparsity and high dimensionality and provide a novel point of view to improve the performance of RS, thereby presenting a useful resource in the state-of-the-art research result for future researchers.