An End-to-End Personalized Preference Drift Aware Sequential Recommender System With Optimal Item Utilization

The user preference is dynamic and requires drift detection to capture changes for delivering relevant recommendations. A sequential recommender system with drift detection was proposed, where drift points are indicated by comparing characteristics of consecutive items. The model leverages drift points to retrieve only interactions with preferences relevant to the current user preference. Nonetheless, the number of utilized items is pre-defined and may not be optimal. It is also not a unified architecture that requires optimizing each part individually. Recently, a Content-Based Transformer has been proposed to consider only items with similar characteristics by leveraging a similarity function. Content-Based Transformer is trained in an end-to-end approach and can be applied for the sequential recommendation task, where the drift of the user preference is indicated as the point where the item’s group changes. However, Content-Based Transformer provides the item’s group as the hard label, ignoring the item characteristics in a real-world scenario where items can exist in many groups. For instance, most movies have multiple genres. This work proposes a unified sequential recommender system that detects the personalized drift pattern of user preference. It groups similar items with soft labels and utilizes the optimal amount of items. The model is trained in an end-to-end approach to jointly optimize for group items with similar characteristics and deliver relevant recommendations. We conducted the experiments to verify the effectiveness of the proposed method by comparing it with Content-Based Transformers and related methods. The evaluation results show that the proposed method consistently outperforms the baselines.