Adaptive Embedding Using Contrastive Learning with Median Absolute Deviation and Variance for Recommendation

Recommendation systems have increasingly relied on embedding techniques to capture user preferences and item features. Adaptive temperature techniques have rapidly developed to provide more accurate recommendations. However, these methods often struggle with popularity bias and a lack of adaptability in dynamic environments, which will end in a decline of the robustness and the bias in temperature prediction. To address these challenges, we propose integrating noise-based data augmentation and InfoNCE loss from the XSimGCL framework into the adaptively modulating embedding magnitude model. Our method enhances traditional adaptive temperature adjustment by incorporating statistical measures like median absolute deviation and variance of embedding norms, aiming to create more robust and adaptive embeddings. This approach eliminates the necessity for adapting temperature by loss, instead focusing on incorporating statistical measures and adding noise to enhance robustness. Extensive evaluations on several real-world datasets demonstrate that our approach effectively mitigates popularity bias and significantly improves overall recommendation quality. This suggests that the nuanced integration of contrastive learning with adaptive embedding techniques can lead to more personalized and equitable recommendations.