Neural Embedding Singular Value Decomposition for Collaborative Filtering.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Singular Value Decomposition (SVD) is a key algorithm in Recommender Systems (RSs).
• Initialization significantly impacts SVD convergence and performance in RSs.
• Current SVD methods often use random initialization, failing to leverage available data effectively.

Purpose of the Study:

• To develop an efficient initialization method for SVD algorithms in RSs.
• To propose a general neural embedding initialization framework for user and item features.
• To support both explicit and implicit feedback datasets within the framework.

Main Methods:

• A low-complexity probabilistic autoencoder neural network is employed for feature initialization.
• The framework is designed to initialize user and item embeddings.
• The approach is validated on explicit and implicit feedback datasets.

Main Results:

• Recommender systems utilizing the proposed initialization framework outperform state-of-the-art methods in rating prediction.
• The framework demonstrates significant improvements in item ranking, ranging from 2.20% to 5.74% over existing SVD and matrix factorization techniques.

Conclusions:

• The proposed neural embedding initialization framework offers a superior alternative to random initialization for SVD in RSs.
• This data-driven initialization enhances both rating prediction and item ranking accuracy.
• The framework effectively utilizes data information, leading to improved recommender system performance.