Skip to main content
Back to timeline
arXivSource publication:

Triggered feeds switch to generative recommendation: TriGen lifts revenue 3.27% in six-market A/B test

Synopsis

When a shopper clicks an item and a follow-up feed opens around it, that feed has mostly been ranked by discriminative click models, while generative recommenders assume interests continue smoothly and have no mechanism for the instant trigger. TriGen repeats the trigger at the tail of the history sequence, uses a learned gate to weight long-term preference against immediate intent per request, and injects a position-specific trigger condition at every decoding step. Offline it beats sequential, generative and trigger-specific discriminative baselines on an industrial e-commerce dataset and an Amazon benchmark; an online A/B test across six Southeast Asian markets reports traffic-weighted lifts of 3.27% in revenue and 0.90% in GMV, followed by full deployment.

Source-provided article image: TriGen: A Generative Paradigm for Trigger-Induced Recommendation
arXiv

Interpretation

Generative recommendation now covers the case where a click opens a follow-up feed: the model no longer just continues the long-term history, because the clicked item actively steers the whole semantic-ID generation. Earlier trigger-aware recommenders were built for discriminative click-through prediction and emit only ranking scores or static representations, so they cannot steer token-by-token generation; existing generative recommenders assume behavioral continuity and lack a mechanism for sudden intent shifts. On the industrial e-commerce dataset TriGen reaches Recall@5 of 5.58 and NDCG@10 of 7.33, above three simple trigger-conditioned generative baselines sharing its backbone, codebook and parameter budget, where GR + trigger append scores 4.35 and 5.62 and GR + trigger conditioning scores 4.46 and 5.51.

The method works in three stages: the trigger is repeated at the tail of the history so encoder self-attention aligns it with recent behavior; two parallel paths extract immediate intent and long-term preference, fused per request by an input-dependent gate; and each decoder step receives a trigger condition specific to that position. The repetition count controls how much query-key capacity the trigger occupies in self-attention, the gate replaces a fixed weighting, and the position-specific feed-forward layers match the coarse-to-fine levels of a semantic ID. On the industrial dataset, ablations drop Recall@5 to 2.88 without the trigger, 4.16 without calibrated tail injection, 5.11 without dual-path fusion and 4.28 without the position-specific trigger feed-forward network, against 5.58 for the full model; replacing the gate with uniform averaging drops it to 4.15.

Perspective

The demonstrated scope is the single-trigger setting: offline evaluation on an industrial e-commerce dataset, the Amazon Industrial_and_Scientific benchmark with the most recent interaction simulating the trigger, and an online A/B test in Indonesia, Malaysia, Singapore, the Philippines, Thailand and Vietnam; the public-data trigger is not a real click, and the authors read that dataset's margin as a conservative lower bound.

The public benchmark has no native trigger annotation, so the trigger is simulated from the most recent history interaction, which the authors say makes the signal weaker and noisier than in production; the method models exactly one trigger and does not yet cover surfaces where several recent signals arrive at once.