[Paper Review] Hybrid Deep-Semantic Matrix Factorization for Tag-Aware Personalized Recommendation

Zhenghua Xu, Cheng Chen+2 more|arXiv (Cornell University)|Aug 12, 2017

Recommender Systems and Techniques15 references3 citations

TL;DR

This paper proposes Hybrid Deep-Semantic Matrix Factorization (HDMF), a novel model that integrates deep-semantic representation learning with matrix factorization using a hybrid learning signal to jointly optimize reconstruction and matrix factorization errors. HDMF significantly outperforms state-of-the-art baselines, achieving 1.52× higher mean reciprocal rank and 1.66× higher mean average precision than the best baseline, DSPR, in tag-aware personalized recommendation.

ABSTRACT

Matrix factorization has now become a dominant solution for personalized recommendation on the Social Web. To alleviate the cold start problem, previous approaches have incorporated various additional sources of information into traditional matrix factorization models. These upgraded models, however, achieve only "marginal" enhancements on the performance of personalized recommendation. Therefore, inspired by the recent development of deep-semantic modeling, we propose a hybrid deep-semantic matrix factorization (HDMF) model to further improve the performance of tag-aware personalized recommendation by integrating the techniques of deep-semantic modeling, hybrid learning, and matrix factorization. Experimental results show that HDMF significantly outperforms the state-of-the-art baselines in tag-aware personalized recommendation, in terms of all evaluation metrics, e.g., its mean reciprocal rank (resp., mean average precision) is 1.52 (resp., 1.66) times as high as that of the best baseline.

Motivation & Objective

• To address the limitations of existing tag-aware recommendation models, including poor optimization in deep-semantic models and marginal performance gains in additional-information-based matrix factorization (AMF) models.
• To overcome the cold start problem in matrix factorization by effectively leveraging social tags as a primary source of user and item representation.
• To improve model interpretability by making the decomposed user and item matrices in matrix factorization semantically meaningful through deep-semantic learning.
• To enhance model optimization and generalization by introducing a hybrid learning signal that combines reconstruction and matrix factorization errors.
• To achieve superior recommendation performance by maximizing the contribution of tag information to the optimization process, unlike AMF models that use it only as a weak regularization term.

Proposed method

• HDMF employs two deep autoencoders to process tag-based user and item matrices, learning deep-semantic representations at the code layer and reconstructing the original matrices at the output layer.
• The model is trained using a hybrid learning signal that minimizes both reconstruction error (difference between input and output matrices) and matrix factorization error (difference between predicted and actual user-item ratings derived from deep-semantic matrices).
• User and item latent factors are derived from the code-layer representations of the autoencoders, which are then used to predict user-item ratings via dot product, forming the core of matrix factorization.
• The hybrid loss function enables better gradient flow, helping the model escape local minima and improving optimization stability.
• The model is interpretable because it identifies the most influential input tags for each dimension of the deep-semantic matrices, enabling meaningful interpretation of latent factors.
• HDMF directly uses tag information to shape the decomposed user and item matrices in matrix factorization, maximizing its impact on optimization, unlike AMF models that treat it as a weak regularizer.

Experimental results

Research questions

• RQ1 Can deep-semantic modeling be effectively combined with matrix factorization to improve tag-aware personalized recommendation?
• RQ2 Does a hybrid learning signal that combines reconstruction and matrix factorization errors lead to better optimization and performance than single-objective training?
• RQ3 Can the deep-semantic matrices in HDMF be made interpretable by identifying key input tags per dimension, thus overcoming the non-interpretability of traditional matrix factorization?
• RQ4 How does HDMF compare to state-of-the-art models like DSPR and AMF variants in terms of recommendation accuracy and robustness across different evaluation metrics?
• RQ5 To what extent does the integration of collaborative filtering signals through matrix factorization enhance the performance of deep-semantic models in cold-start scenarios?

Key findings

• HDMF significantly outperforms all seven baselines across all evaluation metrics on the Delicious dataset, including P@k, R@k, F@k, MAP, and MRR.
• The mean reciprocal rank (MRR) of HDMF is 1.52 times higher than that of the best baseline, DSPR, and its mean average precision (MAP) is 1.66 times higher.
• The relative performance gains of HDMF over DSPR increase with the cut-off rank k, reaching more than double improvement in P@k at k=50, indicating stronger performance on longer recommendation lists.
• HDMF improves MAP and MRR by approximately 70% and 130%, respectively, over the standard matrix factorization model (MF), demonstrating a substantial performance leap from effective tag integration.
• The marginal improvements seen in AMF models like MF_sf and MF_tc (only 2.4% and 6.8% better than MF in MAP and MRR) contrast sharply with HDMF’s gains, highlighting the superiority of HDMF’s direct modeling of tag information in the decomposed matrices.
• The hybrid learning signal enables better optimization and gradient flow, allowing HDMF to escape local minima and achieve superior convergence compared to models relying on single-objective training.