Neural matrix factorization++ based recommendation system

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Traditional Collaborative Filtering (CF) methods like Matrix Factorization (MF) have limited non-linear learning capabilities.
• Neural Collaborative Filtering (NCF) methods incorporate Deep Neural Networks (DNNs) but still face challenges with data sparsity and the cold-start problem.
• Existing hybrid models often struggle with effectively learning latent user and item representations.

Purpose of the Study:

• To propose an improved hybrid recommender system, Neural Matrix Factorization++ (NeuMF++), designed to enhance recommendation accuracy.
• To alleviate the persistent issues of cold start and data sparsity in recommender systems.
• To effectively learn intricate user and item features for superior recommendation performance.

Main Methods:

• Incorporation of Stacked Denoising Autoencoders (SDAE) to generate effective latent representations within the Neural Matrix Factorization (NMF) framework.
• Fusion of Generalized Matrix Factorization (GMF++) and Multilayer Perceptrons (MLP++) components, allowing for separate feature extraction to enhance flexibility.
• Development of NeuMF++ as an extension of the NCF framework, combining linearity and non-linearity for improved feature learning.

Main Results:

• NeuMF++ achieved a test root-mean-square error (RMSE) of 0.8681 on a real-world dataset, demonstrating outstanding performance.
• The integration of SDAE-derived latent representations significantly enhanced the learning capability for user and item features.
• Allowing separate feature extraction for GMF++ and MLP++ components led to substantial performance improvements.

Conclusions:

• NeuMF++ represents a significant advancement in recommender systems, effectively addressing limitations of traditional and existing NCF methods.
• The proposed model demonstrates superior performance in recommendation accuracy and robustness against data sparsity and cold-start issues.
• Future research can extend NeuMF++ by incorporating auxiliary data and exploring diverse neural network architectures for further enhancements.