Global and Local Tensor Factorization for Multi-criteria Recommender System.

Area of Science:

• Computer Science
• Artificial Intelligence
• Recommender Systems

Background:

• Matrix factorization models are standard for recommender systems, using latent factors from user-item ratings.
• Traditional 2D models struggle with multi-criteria recommendation data involving specific rating criteria.
• Existing methods may not fully capture diverse user preferences across different user groups.

Purpose of the Study:

• To introduce a tensor factorization method for handling three-dimensional user-item-criterion rating data.
• To propose a combined global and local tensor factorization (GLTF) approach for improved multi-criteria recommendations.
• To address limitations of single global tensor factorization in characterizing diverse user preferences.

Main Methods:

• Developed a tensor factorization technique to process user-item-criterion rating data.
• Proposed the Global and Local Tensor Factorization (GLTF) method, integrating global patterns with local user-subset behaviors.
• Inferred latent factor representations for users, items, and criteria using the GLTF model.

Main Results:

• The GLTF method effectively leverages both global user-item-criterion patterns and local user-specific behaviors.
• Experimental results on real-world datasets demonstrate the superiority of GLTF over established baseline methods.
• GLTF successfully infers richer latent factor representations for enhanced recommendation accuracy.

Conclusions:

• Tensor factorization provides a robust framework for multi-criteria recommendation systems.
• The proposed GLTF method offers a significant advancement by combining global and local factorization approaches.
• GLTF demonstrates superior performance, paving the way for more personalized and accurate multi-criteria recommendations.