Personalized Recommendation Systems Based on Multi-Source Heterogeneous Data

With the rapid development of Internet technology and the widespread generation of multi-source heterogeneous data, various types of data such as user behavior, social relationships, and text information jointly depict the complex and changeable interests and preferences of users. However, traditional recommendation system faces challenges in effectively integrating heterogeneous information sources and providing diverse recommendation results, often leading to limited recommendation diversity and suboptimal user satisfaction when processing these data. In order to solve the challenge resulting from the diversity of user interests and the fusion of multi-source heterogeneous data. Traditional multimodal fusion techniques predominantly depend on predetermined methodologies for integrating diverse data sources. However, these conventional approaches often prove inadequate in comprehensively capturing the intricate interdependencies among the data sources, consequently leading to suboptimal performance in recommendation systems. Therefore, we proposed a recommendation framework named GraphSelfRec, the fundamental new of GraphSelfRec lies in its development of a multi-level adaptive attention mechanism, which dynamically adjusts the weights of different data sources and effectively captures the heterogeneous nature of user interests through its fusion strategy. In addition, the framework introduces efficient neighbor sampling and graph compression techniques to meet the needs of large-scale data scenarios. Experimental results show that the proposed model significantly outperform existing methods in terms of recommendation accuracy and diversity, while also achieving high efficiency and transparency.