Enhancing Personalized Recommendation with Collaborative Filtering: A Multi-Dimensional Perspective

Abstract

Within the rapidly evolving digital landscape, recommender systems have become pivotal in assisting individuals to locate and filter relevant information, thereby alleviating information overload. To identify potential targets of interest from vast information sources, recommender systems need to enhance their ability to capture user preferences. These preferences are often reflected in users' historical behaviors, particularly their interactions with other targets within the recommendation context. Consequently, the extraction of user preferences from interactions between users and recommended targets, or items, has emerged as a core task within recommender systems. This task, known as collaborative filtering, has been extensively studied in recent years by academia and industry.

Collaborative filtering can be conceptualized as a form of information retrieval, and its efficacy can thus be assessed via the accuracy of retrieval. However, as applications deployed and serving real human users, our expectations from today's recommender systems go beyond mere accuracy. Factors such as the capacity to provide recommendations within complex contextual scenarios, scalability in dealing massive users and items, interpretability, novelty, and fairness of recommendations have also been integrated into the assessment of recommender system performance.

In this thesis, we study recommender systems with collaborative filtering in conjunction with various evaluation dimensions, and propose methods that incorporate multidimensional advantages. First, most traditional collaborative filtering methods lack explainability of recommendation results. To address this issue, we propose a new graph convolutional network (GCN) model to simultaneously learn item recommendation and explanation ranking tasks. We introduce new information propagation strategies and loss functions to mitigate performance degradation caused by conflicts between different task objectives. Experimental results validate the effectiveness of these strategies in mitigating task conflicts, providing recommendation results that surpass existing models in accuracy, while also offering explanations for the recommendations.

Secondly, we address the fact that most collaborative filtering methods focus on interactions between users and items, neglecting the context of the scenario. We concentrate on scenarios where both user preferences and contextual information are pivotal in recommendation, such as when users review specific products or services. We model this scenario as a collaborative filtering problem involving user-item-context relationships. Correspondingly, we propose a sparse factorization model and an accelerated optimization algorithm. These methods have been demonstrated to yield more accurate recommendation results in these scenarios, and exhibit faster convergence rates compared to existing methods.

While the aforementioned methods show strong recommendation performances in small-scale academic scenarios, they may incur prohibitive computational costs in real scenarios with a vast number of items and users. To address this issue, we develop a method to boost efficiency and scalability in large-scale item scenarios. Specifically, through theoretical and experimental analysis of existing GCN methods, we uncover their implicit utilization of community structures within the item set in to reduce the cost of constructing graphs. Based on this, we propose new methods that reduce the learning scale of linear autoencoders through graph partitioning, and introduce new strategies to minimize information loss due to partitioning. Experiments on large-scale real-world datasets demonstrate that our model outperforms existing models, offering both high training efficiency and robust scalability.

Finally, traditional recommendation methods have been found to exhibit a significant popularity bias, with more popular items appearing more frequently in recommendation results. Therefore, we explore the relationship between recommendation system accuracy and novelty, where novelty refers to the proportion of less popular items. Specifically, we view user-item interactions as graph signals and model the collaborative filtering problem based on metric learning. Further theoretical analysis uncovers the relationship between the normalization strength of graph signals and the novelty of recommendations. Based on these findings, we propose a new model to learn a generalized distance metric by modeling the residuals of user-item distances to capture user preferences. The role of the proposed method in enhancing recommendation accuracy and novelty is validated in experiments.

Overall, we explore the performance of recommender systems using collaborative filtering across different scenarios and evaluation perspectives. Our proposed models have been validated to ensure accuracy, while possessing characteristics expected of practical, human-centered modern recommender systems.

Date of Award	27 Jun 2024
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Wai Shing Tommy CHOW (Supervisor)