Representation learning for recommender systems
Vinh Tran Lucas · 2020
Recommender systems are widely used in many big companies such as Facebook, Google, Twitter, LinkedIn, Amazon and Netflix.They help to deal with the problem of information overload by filtering the important information fragments efficiently according to users' preferences and interests.However, it is challenging to develop highly effective models for handling various recommendation problems, from individual-level to group-level tasks.To be specific, the standard recommendation problem is personalized with respect to a single user, where it aims to identify the most interesting and relevant items to make personalized recommendation.In contrast, group level recommendation deals with groups of users instead of individuals, in which fine-grained, intricate group dynamics and preferences have to be considered.With the success of deep learning, many neural architectures have been proposed for learning-to-rank recommendation.Indeed, recent studies have shown their effectiveness and efficiency in providing better recommendations in terms of user and group-ofusers satisfaction with the involvement of deep learning.The key intuition behind many successful end-to-end neural architectures for recommendation is to design appropriate frameworks that are not only able to learn rich user and item representations, but also well capture the implicit and hidden relationships behind users and items.Therefore, the objective of designing such effective neural architectures remains a challenge, especally in the context of different recommendation tasks such as general recommendation, next-item recommendation, shopping-basket recommendation, etc.This dissertation focuses on designing neural architectures for personalized and group recommendation.More specifically, we explore representation learning techniques, both Euclidean and non-Euclidean representation, for learning-to-rank user/group-of-users and item pairs.The key contributions of this dissertation are listed below.Personalized Recommendation.Our contributions are summarized as follows:• Wasserstein based Metric Learning Representation for Recommendation.We introduce a novel Wasserstein distance-based Metric Learning Chain (W-MLC) model.Our W-MLC model employs a series of metric learning, together with a vi Wasserstein distance to constraint on the user/item projection transformations, allowing us to encode user-item interactions better through a deeper chain.In addition, we propose a hinge loss function with personalized adaptive margins for different users.Extensive experiments on eight datasets of three different recommendation tasks reveal the effectiveness of our proposed model over eight strong state-of-the-art baselines.• Going Beyond Euclidean: Hyperbolic Representation for Recommendation.We investigate the notion of learning user and item representations in non-Euclidean space.Specifically, we study the connection between metric learning in hyperbolic space and collaborative filtering by exploring Möbius gyrovector spaces where the formalism of the spaces could be utilized to generalize the most common Euclidean vector operations.Overall, this work aims to bridge the gap between Euclidean and hyperbolic geometry in recommender systems through metric learning approach.We propose HyperML (Hyperbolic Metric Learning), a conceptually simple but highly effective model for boosting the performance.Via a series of extensive experiments, we show that our proposed HyperML not only outperforms their Euclidean counterparts, but also achieves state-of-the-art performance on multiple benchmark datasets, demonstrating the effectiveness of personalized recommendation in hyperbolic geometry.7.1 Dataset Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .114 7.2 Performance comparison between MF-AVG and ATT-AVG (Ablation study) on four datasets.Results show that our proposed attention mechanism is significantly better than a standard attention-based aggregation. . . . .115 7.3 Performance comparison between MoSAN and PIT on Meetup and Plancast datasets in terms of ndcg@5 by removing top-K high weight users. .118 xviii