Matrix Factorization with Dynamic Multi-view Clustering for Recommender System

TL;DR

The paper addresses the scalability and interpretability gap in traditional matrix factorization-based recommender systems by introducing Matrix Factorization with Dynamic Multi-view Clustering (MFDMC). MFDMC unifies MF with dynamic multi-view clustering to learn view-specific, weight-adjusted cluster-centered representations and employs pruning to manage cluster complexity, enabling end-to-end training on web-scale data. The method yields improved RMSE on six real-world datasets and demonstrates interpretability through visualizations and semantic cluster analysis, with demonstrated applicability to computer vision tasks. This approach offers scalable, explainable representations that adapt to evolving user-item dynamics and can extend to downstream tasks beyond recommendation.

Abstract

Matrix factorization (MF), a cornerstone of recommender systems, decomposes user-item interaction matrices into latent representations. Traditional MF approaches, however, employ a two-stage, non-end-to-end paradigm, sequentially performing recommendation and clustering, resulting in prohibitive computational costs for large-scale applications like e-commerce and IoT, where billions of users interact with trillions of items. To address this, we propose Matrix Factorization with Dynamic Multi-view Clustering (MFDMC), a unified framework that balances efficient end-to-end training with comprehensive utilization of web-scale data and enhances interpretability. MFDMC leverages dynamic multi-view clustering to learn user and item representations, adaptively pruning poorly formed clusters. Each entity's representation is modeled as a weighted projection of robust clusters, capturing its diverse roles across views. This design maximizes representation space utilization, improves interpretability, and ensures resilience for downstream tasks. Extensive experiments demonstrate MFDMC's superior performance in recommender systems and other representation learning domains, such as computer vision, highlighting its scalability and versatility.

Figures (6)

• Figure 1: Schematic Illustration of the Multi-View Feature-Weighted Dynamic Cluster (MFDMC) Framework. User/item embeddings are projected into $t$ weighted cluster centers across $v$ distinct views, with each view’s representation formulated as a weighted summation of cluster centers. Distinct color groups denote different views, where darker shading intensity corresponds to higher feature importance.
• Figure 2: Mapping function of Eq. \ref{['equ:map']} for the weight. Using the function in the middle of the Figure, we can map the weights on the left to the right, which facilitates a relatively balanced optimization of weight in different views.
• Figure 3: Illustration of the MFDMC for computer vision tasks. The features are embedded into the weights of cluster centers in each view. The final representation of each image for downstream tasks is obtained by summing up the weighted centers of each view and concatenating them.
• Figure 4: Multi-view Clustering results. The four rows show the clustering centers of users/items, and the clustering results of users/items respectively. Each column represents a view. In the clustering result, each point corresponds to the value of the user/item latent vector in a specific view.
• Figure 5: The t-sne results of experiments on synthetic toy dataset. (a), (b), and (c) show the results of shape, color view, and the entire embedding respectively.