Contrastive Learning and Deep Fusion Recommendation Model Based on ID Features

Abstract

In recent years, the application of deep learning in recommendation systems has achieved breakthrough progress. Neural networks have captured the complex nonlinear relationships between user behaviors and item characteristics. As indispensable elements, ID (Identity) features are widely used to construct user-item association models. The embedding of ID features plays a crucial role in neural network-based recommendation models, enhancing the model’s learning ability by converting them into one-hot vectors as input to the deep learning network. ID features are discrete and high-dimensional sparse, containing critical information for reasoning in recommendation systems. However, most models use the same embedding method to encode all features, leading to inaccurate representation of ID features and affecting the model’s generalization ability. This paper proposes a Contrastive Learning and Deep Fusion Recommendation Model based on ID features (CDM-ID) to address this issue. This model independently encodes ID features first, then accurately re-encodes ID and attribute features through a contrastive learning module, generating high-quality inputs for subsequent feature interaction modules. Finally, it achieves feature fusion through an attention mechanism, significantly improving the performance of the recommendation model. Experiments conducted on three public datasets demonstrate that CDM-ID achieves an average improvement of 8.95% and 5.17% in AUC and F1 scores, respectively, compared to the various baseline models. Additionally, it reduces MAE and RMSE by an average of 9.62% and 12.75%, respectively, further validating the effectiveness and superiority of the proposed model. The contributions of this paper lie in optimizing ID feature embeddings, enhancing embedding accuracy, generating multi-order interaction features, and applying an attention mechanism for feature fusion, thereby significantly improving the performance and prediction accuracy of recommendation systems.

Keywords

• Recommendation model
• ID features
• contrastive learning
• feature interaction
• attention mechanism