Beyond Time as Context: Joint Item-Time Prediction for Sequential Recommendation

Time-aware sequential recommendation aims to predict which item a user will purchase, while temporal information helps characterize evolving user preferences and item dependencies based on historical interactions. Existing time-aware sequential recommendation models often use temporal information mainly as auxiliary features such as time intervals or cyclical patterns extracted from purchase records. In this work, we propose a recommendation-oriented perspective in which temporal signals are explicitly modeled as an additional learning objective to support item recommendation, rather than being used only as contextual features. Within a unified framework, our approach jointly learns next-item prediction together with the expected time of the next interaction, where item recommendation remains the primary objective and temporal prediction provides explicit temporal supervision for representation learning and reranking. Specifically, we propose a unified Item-Time Dual Recommendation framework named ITDRec. Built on a standard time-aware sequential backbone, ITDRec jointly learns next-item prediction and candidate-conditioned future time estimation within a unified architecture. To support this formulation, ITDRec employs a time decoder with a candidate-aware Chronological Preference Mapping (CPM) module based on cross-attention, together with three complementary time predictors for modeling temporal signals from multiple perspectives. Extensive experiments on multiple real-world datasets demonstrate that ITDRec consistently and significantly outperforms strong baselines on item recommendation. Additional results on time prediction show that the temporal branch learns meaningful temporal patterns that provide useful signals for the recommendation task.