面向序列推荐的个性化时间感知注意力模型OA

Sequential recommendation is one of the research hotspots in the recommendation field in recent years,where the key challenge is to predict the next item of interest from users'interaction histories.However,most approaches simplify users'interaction histories as ordered sequences while ignoring crucial temporal information,which leads to sub-optimal performance by failing to capture the complex dependencies between user preferences and temporal contexts.Some studies have considered the timestamps or time intervals,yet still fail to fully exploit the multidimensional temporal information(e.g.,year,month,day).Moreover,they ignore moment when the recommendations are delivered,which prevents them from capturing behavioral patterns linked to that target time and ultimately degrades model performance.To address these issues,this paper proposes a sequential recommendation method with a personalized time-aware attention mechanism that fuses multidimensional temporal information via multi-head attention to more accurately capture users'evolving preferences.Meanwhile,this paper designs a personalized target time-aware module that models users'real-time preferences by assigning higher weights to past actions occurring in time range similar to the target time at which the recommendation is delivered.Experiments on five real-world public datasets demonstrate that the proposed method outperforms all baseline models,achieving average improvements of 5.51%and 5.87%in NDCG@5 and MAP@5,5.08%and 5.66%in NDCG@10 and MAP@10 over the SOTA model,validating the effectiveness and superiority of the proposed method.