Every online shopper leaves behind a trail of half-finished intentions. A pair of sneakers browsed at midnight, a coffee maker favorited on a lunch break, a winter coat dropped into a cart and then abandoned for weeks. To a recommendation algorithm, all of these signals look like interest. But as any human observer could guess, a click from three weeks ago and an add-to-cart from this morning do not mean the same thing. A new study published in Discover Artificial Intelligence argues that the field has been treating these decaying signals too crudely, and it proposes a fix that explicitly models how long each behavior stays alive before it expires.

The framework, called BLMRec for Behavior Lifetime Modeling-based Recommendation, was developed by Yu Chen of Jiangxi University of Software Professional Technology in Nanchang, China. Its central premise is deceptively simple: the purchase relevance of browsing, favoriting, and add-to-cart actions changes over time, and can become stale before the moment a prediction is made. Rather than treating every historical action as a fixed indicator of preference, BLMRec assigns each one a kind of internal clock, estimating how much purchase signal it still carries right now, and how much of that signal has already leaked away.

The technical machinery behind this idea rests on three cooperating modules. The first, multi-scale behavior lifetime modeling, encodes every historical event using the user, item, behavior type, category, and timestamp, then predicts two quantities from that representation: an initial effective strength describing how strongly the event signaled purchase intent when it occurred, and a lifetime scale describing how quickly that strength decays. The current validity of an event is computed as its initial strength multiplied by an exponential decay over the elapsed interval between the event and the prediction moment. Crucially, the decay rate is not fixed. It is learned per event, so a fleeting page view can expire quickly while a deliberate add-to-cart lingers longer.

On top of validity, the model estimates a stagnation risk. An add-to-cart that sits unconverted for a long time, or a favorite that is never followed by any further action, is not simply neutral; it may actively mislead the ranker. The risk head takes the event representation, an encoding of the elapsed interval, and the current validity, and outputs a score that down-weights such stagnant signals before they ever reach the matching stage. To train these lifetime states, the study constructs supervision labels across multiple observation windows, ranging from one to thirty days depending on the dataset. An event is labeled as valid feedback if the same item progresses through the interaction chain, or if a related item is purchased within the window; otherwise it counts as stagnation feedback.

The second module tackles a different question: even if a behavior is still temporally valid, does it actually support the specific candidate item being ranked? BLMRec answers this through candidate-conditioned relation matching, which builds explicit relation features between each historical item and each candidate. Four channels are used: category consistency, behavioral co-occurrence, co-purchase patterns, and auxiliary-to-purchase transition statistics, the last counting how often users who performed the same auxiliary action on the historical item went on to buy the candidate. These features are combined with the lifetime-calibrated event representation to compute a per-candidate attention weight, so the same history produces different aggregated signals for different candidate items.

The third module addresses perhaps the thorniest problem in multi-behavior recommendation: interest is not intent. A user may browse an item repeatedly out of curiosity or comparison shopping, never intending to buy. The purchase-intent calibration module computes an auxiliary-interest score and a target-purchase preference score separately, then learns a gate that controls how much of the auxiliary signal is allowed into the final purchase ranking. During training, the model is explicitly penalized when it ranks high-auxiliary non-conversion items, those that received strong auxiliary actions but were never purchased, above the items the user actually bought.

The empirical results are striking. Using two public datasets, Taobao UserBehavior with 33.5 million interactions and RetailRocket with 1.35 million interactions after preprocessing, BLMRec improved Recall@20 and NDCG@20 over the strongest baseline, RMBRec, by 6.0 and 7.3 percent on Taobao and by 6.9 and 9.2 percent on RetailRocket. Under a deliberately adversarial evaluation protocol, where negative candidates were drawn from repeatedly browsed, favorited, or carted items that were never purchased, the advantage grew to 8.8 and 10.1 percent. That the gains widen precisely when the candidate set is full of confusable, high-auxiliary traps suggests the framework is doing what it claims: suppressing noise rather than merely fitting the data more tightly.

The ablation experiments reveal which pieces matter most. Removing the lifetime modeling module cut Recall@20 from 0.1158 to 0.1061, the largest single drop, confirming that temporal calibration is the engine of the improvement. Replacing it with a fixed exponential time decay fared even worse, showing that a one-size-fits-all decay function cannot handle the heterogeneous temporal semantics of different behavior types. The learned lifetime statistics themselves are illuminating: favoriting showed a longer lifetime scale than browsing but lower current validity than add-to-cart and persistently high stagnation risk, consistent with favoriting reflecting delayed decision-making rather than imminent purchase. The calibration gate behaved sensibly across scenarios too, reaching its highest value of 0.676 for purchases within three days of an add-to-cart, and dropping to 0.443 when a cart sat unpaid for more than a week.

Efficiency numbers matter for any system pitched at real e-commerce scale, and here the design shows restraint. By retaining only the top twenty historical signals ranked by current validity, the matching cost scales with that fixed budget rather than the full history length, and relation statistics are precomputed and cached offline. Compared with RMBRec, BLMRec increased training time by 6.5 percent, GPU memory by 4.1 percent, and inference latency by 15.1 percent, a modest price for the accuracy gains. The study is candid about limitations: relation matching weakens when item metadata or interaction statistics are sparse, and lifetime supervision cannot capture substitute purchases outside the available relations or conversions beyond the predefined windows. Still, the broader message lands with force. Recommendation systems have long treated a user’s history as a pile of evidence, weighting it by type and recency. This work suggests something more nuanced: every behavioral signal has a lifetime, and knowing when it dies may matter as much as knowing that it happened.

Subject of Research: Time-sensitive modeling of multi-behavior user signals for purchase prediction in e-commerce recommendation systems

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

Article References: Chen, Y. (2026). Behavior lifetime modeling uses time sensitive behavioral signals to rank candidate items for purchase in E commerce recommendation. Discover Artificial Intelligence, 6(1), Article 1471. https://doi.org/10.1007/s44163-026-02419-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02419-1

Keywords: multi-behavior recommendation, e-commerce, purchase prediction, behavior lifetime modeling, temporal validity, recommendation systems, machine learning, attention mechanism, Taobao UserBehavior, RetailRocket, time decay, stagnation risk