Recommendation systems have seen significant advancements with the application of machine learning techniques, yet challenges remain in maintaining optimal performance throughout training. Contrastive learning, while effective in enhancing user and item representations, often suffers from performance degradation over time. In this paper, we present a novel approach that incorporates time-variant objectives (TVO) to address this issue. By integrating a scheduler with various time-variant functions into the contrastive learning framework, we dynamically balance the recommendation loss and contrastive loss during training. This method stabilizes model performance and mitigates the typical decline observed in traditional approaches. Our experimental results show that the TVO-enhanced model achieves more reliable and precise recommendations compared to existing methods. This approach offers a promising solution for improving the consistency and accuracy of contrastive learning-based recommendation systems. Keywords: Recommendation Systems, Contrastive Learning, Time Variant Objectives, TVO, GCN
