Conditional diffusion model for recommender systems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Information Retrieval

Background:

• Recommender systems combat information overload by filtering personalized content.
• Diffusion models, particularly conditional diffusion models, are effective for reconstructing user interaction vectors and predicting preferences.
• Existing methods often use generic labels or features as guidance, limiting personalization.

Purpose of the Study:

• To develop an improved recommendation method using conditional diffusion models.
• To enhance recommendation performance by integrating user-specific preference features into the reverse diffusion process.
• To propose a novel strategy using user interaction vectors as conditional guidance.

Main Methods:

• Utilizing user interaction vectors as conditional guidance information within a conditional diffusion model framework.
• Employing neural networks as encoders to process user data.
• Conducting experiments on five diverse datasets (movies, music, beauty, sports) with varying sizes and sparsity levels.

Main Results:

• The proposed method achieved a 7.41% and 6.00% performance improvement through distinct approaches.
• The model enhanced Top-10 and Top-20 recommendation metrics by 5.59% and 4.38%, respectively, compared to baselines.
• Hyper-parameter analysis indicated optimal performance with small diffusion steps and moderate noise levels.

Conclusions:

• The conditional diffusion model effectively leverages user interaction data for improved personalized recommendations.
• The proposed approach offers significant performance gains over existing recommender system baselines.
• Further research is needed to address limitations in relationship network applications and model scalability.