|
چکیده
|
Recommender systems often face a significant challenge of data sparsity, as user-item interaction matrices in practical implementations typically exhibit over 95% sparsity. This issue negatively impacts the accuracy and utility of recommendations, especially for new users and niche products. While current methodologies aim to mitigate sparsity by incorporating additional information or modifying system designs, they often overlook the core problem of insufficient interaction data. This study introduces a novel framework that employs a diffusion model to generate high-quality synthetic ratings for recommender systems. Specifically, we use a diffusion model designed for tabular data to learn the complex joint distribution of user-item-rating triples and produce synthetic ratings that are statistically coherent. To ensure the quality of the generated ratings, we propose an innovative method for detecting noise through behavioral pattern analysis. This method identifies and removes noisy ratings that do not align with user preferences and item characteristics. We employ traditional collaborative filtering and neural collaborative filtering models to assess the effectiveness of the proposed framework. Experiments conducted on two real-world datasets, featuring varying levels of sparsity (ranging from 5% to 100% data retention), demonstrate significant improvements. Specifically, traditional collaborative filtering can reduce the root mean square error by as much as 51% and enhance rating coverage by 36% points under sparse conditions. Additionally, neural collaborative filtering provides more nuanced responses and is most effective with a low augmentation ratio, indicating that different models require distinct forms of augmentation.
|