Behavior lifetime modeling uses time sensitive behavioral signals to rank candidate items for purchase in E commerce recommendation

Multi-behavior recommendation relies on heterogeneous user actions to infer purchase intent, but the purchase relevance of browsing, favoriting, and add-to-cart behaviors can change over time and become stale before prediction. To address this problem, we propose behavior lifetime modeling-based recommendation (BLMRec), a candidate-item purchase ranking framework that explicitly models the temporal validity of historical behaviors. BLMRec encodes each behavior according to its type and temporal distance from the prediction moment, and employs multi-window lifetime learning to estimate both short-term activation and long-term persistence. The resulting temporal validity and stagnation signals are used to suppress outdated auxiliary behaviors. BLMRec further performs candidate-conditioned matching between historical and candidate items through category, co-occurrence, co-purchase, and behavior-transition relations, followed by purchase-intent calibration for final ranking. Experiments on Taobao UserBehavior and RetailRocket, with 33.5 M and 1.35 M interactions after preprocessing, show that BLMRec improves Recall@20/NDCG@20 over the strongest baseline by 6.0%/7.3% and 6.9%/9.2%, respectively. Further analyses indicate that the gains are associated with improved suppression of stale, weakly related, and high-auxiliary non-conversion signals.