Towards a Better Understanding of Linear Models for Recommendation

Zhi Liu, Ruoming Jin, Dong Li, Jing Gao, Li Chen, Yang Zhou

Abstract: Linear regression models (EASE, SLIM) often match or outperform deep models, yet their relation to (weighted) matrix factorization is unclear. We derive and analyze closed-form solutions for both approaches, revealing they both “scale-down” singular values but differ in regularization implications. We introduce an algorithm for hyper-parameter search over the closed-form solution, discovering nearby models. Experiments on benchmark datasets confirm that these basic models are competitive with state-of-the-art methods, validating the theoretical insights.