Heterogeneous Recommendation via Deep Low-Rank Sparse Collective Factorization.

Area of Science:

• Recommender Systems
• Machine Learning
• Data Mining

Background:

• Real-world recommender systems utilize diverse user feedback, including numerical and binary ratings.
• Existing collective factorization methods often overlook shared rating patterns and hierarchical data structures in heterogeneous recommendations.
• Data sparsity poses a significant challenge, particularly when transferring knowledge from sparse binary ratings to denser numerical ratings.

Purpose of the Study:

• To propose a novel Deep Low-rank Sparse Collective Factorization (DLSCF) framework for heterogeneous recommendation.
• To address challenges of data sparsity and hierarchical structures in recommendation systems.
• To effectively transfer knowledge from binary user ratings to numerical user ratings.

Main Methods:

• Developed a Deep Low-rank Sparse Collective Factorization (DLSCF) framework.
• Employed low-rank sparse decomposition to capture common rating patterns across domains while isolating domain-specific patterns.
• Utilized multi-layer factorization to model hierarchical relationships between latent categories and sub-categories.
• Proposed both batch and Stochastic Gradient Descent (SGD) based optimization algorithms for DLSCF.

Main Results:

• The proposed DLSCF framework demonstrated effectiveness across multiple benchmark datasets (MoviePilot, Netfilx, Flixter, MovieLens10M, MovieLens20M).
• DLSCF outperformed several state-of-the-art batch and SGD-based recommendation approaches.
• The framework successfully addressed data sparsity and leveraged hierarchical structures for improved recommendation accuracy.

Conclusions:

• The DLSCF framework offers a robust solution for heterogeneous recommendation scenarios with sparse and hierarchical data.
• The proposed methods provide effective knowledge transfer from binary to numerical ratings, enhancing recommender system performance.
• DLSCF represents a significant advancement in collective factorization techniques for complex recommendation tasks.