[Paper Review] Top-N Recommender System via Matrix Completion

Zhao Kang, Chong Peng+1 more|arXiv (Cornell University)|Jan 19, 2016

Recommender Systems and TechniquesComputer Science27 references22 citations

TL;DR

This paper proposes a nonconvex matrix completion method for Top-N recommender systems that improves recommendation accuracy by approximating the low-rank user-item matrix using a log-determinant rank surrogate, outperforming state-of-the-art methods like SLIM and LorSLIM across six real-world datasets with significant gains in HR and ARHR metrics.

ABSTRACT

Top-N recommender systems have been investigated widely both in industry and academia. However, the recommendation quality is far from satisfactory. In this paper, we propose a simple yet promising algorithm. We fill the user-item matrix based on a low-rank assumption and simultaneously keep the original information. To do that, a nonconvex rank relaxation rather than the nuclear norm is adopted to provide a better rank approximation and an efficient optimization strategy is designed. A comprehensive set of experiments on real datasets demonstrates that our method pushes the accuracy of Top-N recommendation to a new level.

Motivation & Objective

• Address the low recommendation quality in Top-N recommender systems due to data sparsity and suboptimal rank approximation.
• Overcome limitations of existing methods like SLIM and LorSLIM that fail to fully preserve original user-item ratings or struggle with sparse data.
• Improve recommendation accuracy by reconstructing the user-item matrix through a nonconvex rank minimization that better approximates true low-rank structure.
• Design an efficient optimization algorithm using ADMM with dynamic penalty parameter to balance convergence speed and accuracy.
• Demonstrate superior performance and robustness across diverse real-world datasets with varying sparsity levels and sizes.

Proposed method

• Formulate matrix completion as a nonconvex rank minimization problem using the log-determinant function to approximate the matrix rank.
• Introduce a constraint to preserve observed entries in the user-item matrix, ensuring original data is retained during reconstruction.
• Use the nuclear norm as a convex surrogate in prior methods, but replace it with a nonconvex log-determinant function for better low-rank approximation.
• Apply the Alternating Direction Method of Multipliers (ADMM) with a dynamically increasing penalty parameter μ to accelerate convergence.
• Solve the optimization problem iteratively, leveraging efficient SVD computation for low-rank approximation to reduce time complexity.
• Reconstruct the full user-item matrix X such that X_ij = M_ij for observed entries, and predict unobserved ratings via the completed matrix.

Experimental results

Research questions

• RQ1 Can a nonconvex rank approximation using the log-determinant function improve Top-N recommendation accuracy compared to convex nuclear norm relaxation?
• RQ2 How does the proposed method perform in comparison to state-of-the-art methods like SLIM, LorSLIM, BPRMF, and WRMF across diverse real-world datasets?
• RQ3 To what extent does the proposed method preserve original user-item ratings during matrix reconstruction, especially in sparse settings?
• RQ4 How sensitive is the method’s performance to the choice of auxiliary parameters like initial μ and γ in the ADMM framework?
• RQ5 What is the computational efficiency of the proposed method relative to existing matrix completion-based recommenders like LorSLIM?

Key findings

• The proposed method achieves an average improvement of 48.55% in Hit Rate (HR) and 60.38% in Average Reciprocal Hit Rate (ARHR) over ItemKNN, outperforming all other baselines.
• On the ML100K dataset, the method recovers all zero entries with a mean of 0.554, significantly higher than LorSLIM’s 0.046, indicating superior matrix reconstruction fidelity.
• The method outperforms LorSLIM, the second-best baseline, by 14.15% on average in HR and 14.15% in ARHR, demonstrating consistent superiority across all six datasets.
• The method is robust to parameter choices: HR remains stable across different initial μ⁰ and γ values, indicating low sensitivity to auxiliary parameters.
• The algorithm is significantly faster than LorSLIM, requiring only 9 seconds on ML100K and 5080 seconds on lastfm, compared to 617 and 32974 seconds respectively, with γ enabling further speedup without major accuracy loss.