Revisiting the loss functions in sequential recommendation

There has been a growing interest in benchmarking sequential recommendation models and reproducing/improving existing models. However, the exploration of loss functions in this context has been relatively limited. To address this gap, we investigate the application of diverse loss functions in sequential recommendation, focusing on Cross-Entropy (CE), Binary Cross-Entropy (BCE), and Bayesian Personalized Ranking (BPR) losses. Our objective is to enhance model performance through refining these loss functions. Existing loss functions in sequential recommendation are discussed and analyzed, summarizing their pros and cons. Following this, we identify the two crucial characteristics that an efficient loss function in sequential recommendation should embody, namely Time-Awareness and Efficiency. Subsequently, we introduce a straightforward yet impactful approach that seamlessly integrates time-awareness and efficiency into the computation of the loss function. Further, by utilizing meta-loss learning, we train time-aware parameters and integrate them into the loss function, enabling the model to adaptively capture significant user interactions with items. Our methodology is extensively validated through comprehensive experiments involving various classical and advanced sequential recommendation models on multiple publicly available datasets. Impressively, our approach not only leads to substantial performance improvements but also enables baseline models to surpass state-of-the-art models.

Authors

• Fangyu Li JD Logistics Huazhong University of Science and Technology
• Hong Peng
• Shenbao Yu Xiamen University, China
• Feng Zeng
• Fei Chen
• Fan Yang Nanyang Technological University

Keywords

• Sequential recommendation
• Loss function
• Meta loss learning