TL;DR
TriGen is a novel generative recommendation framework designed for "trigger-induced" scenarios, where a specific user click (the "trigger") should steer subsequent suggestions. By combining calibrated history injection, adaptive dual-path fusion, and position-specific decoder guidance, TriGen effectively balances long-term user preferences with immediate shopping intent, delivering significant revenue and GMV lifts in large-scale production environments.
What Problem Does This Paper Address?
Most modern Generative Recommendation (GR) models operate in a "passive" mode known as User-Induced Recommendation (UIR). They treat a user's history as a continuous sequence and try to predict the next item as a natural extension of that history.
However, real-world e-commerce platforms frequently use Trigger-Induced Recommendation (TIR). As shown in Figure 1, when a user clicks a specific product in a general feed, the platform often opens a follow-up feed organized around that "trigger" item.
This creates a structural tension:
- Continuity: The system must respect the user's long-term tastes.
- Redirection: The system must immediately pivot to the intent signaled by the trigger click.
Existing GR models lack mechanisms to reconcile this tension, often either ignoring the trigger or over-specializing on it. Meanwhile, traditional TIR methods are "discriminative"—they score a pre-defined pool of candidates—and cannot be easily adapted to the token-by-token generation process used in state-of-the-art generative models.
Methods
TriGen introduces an end-to-end encoder-decoder architecture that manages trigger influence across the entire generation pipeline. The process is broken down into four key stages:
- Input Representation: Items are represented by Semantic IDs (SIDs)—hierarchical sequences of discrete tokens (e.g., [Category, Sub-category, Brand, Item-specific-code]).
- Calibrated Tail Injection: To ensure the trigger isn't "diluted" by a long history, TriGen appends $N$ copies of the trigger item to the end of the historical sequence. It uses Rotary Position Embeddings (RoPE) to help the model understand the relative distance between the trigger and recent behaviors.
- Adaptive Dual-Path Fusion: The model decouples "habitual" preferences from "instant" intent. It processes the history and the trigger through two parallel paths. An input-dependent Gating Router then dynamically decides how much to weigh each path for every specific request.
- Position-Specific Trigger-PFFN: Because SIDs are hierarchical, the first token generated needs broad categorical guidance, while later tokens need fine-grained attribute guidance. TriGen uses independent feed-forward networks (MLPs) at each decoding step to provide "level-aware" trigger steering.
Datasets
The authors evaluated TriGen on two primary datasets:
- Industrial Dataset: A massive dataset from a major e-commerce platform containing over 49 million users and 1.4 billion samples, featuring real exposure, click, and follow-up feedback signals.
- Amazon Reviews 2018 (Industrial_and_Scientific): A public benchmark. Since this dataset lacks native "trigger" annotations, the authors simulated them by treating the most recent interaction as the trigger and the preceding items as history.
Key Insights and Ideas
- The "Bell Curve" of Repetition: Simply adding the trigger once isn't enough to steer a long history, but repeating it too many times overwhelms the user's persistent tastes. The authors found a "bell-shaped" performance curve (Figure 3), confirming that trigger influence must be carefully calibrated.
- Decoupling is Crucial: You cannot simply mix history and trigger intent in the encoder and expect the model to sort it out. By explicitly separating the paths and using a learned router, TriGen can handle both "exploratory" clicks (where the trigger should lead) and "accidental" clicks (where history should prevail).
- Hierarchical Guidance: In generative recommendation, the "meaning" of a token changes as you decode. TriGen’s use of position-specific MLPs in the decoder ensures that the trigger steers the generation at the right level of granularity—from broad categories down to specific attributes.
Key Results
TriGen demonstrated superior performance in both offline experiments and live production:
- Offline Performance: On the industrial dataset, TriGen achieved a Recall@10 of 15.50 and NDCG@10 of 7.33, significantly outperforming both generative baselines (like TIGER and LC-Rec) and discriminative TIR models (like DIHN and DIAN).
- Online A/B Tests: Deployed across six international markets, TriGen delivered a traffic-weighted overall lift of +3.27% in Revenue and +0.90% in GMV. In Singapore, GMV lift reached as high as +20.54%.
- Efficiency: Despite the added complexity, the online inference latency is only ~10.5 ms, making it virtually indistinguishable from simpler control models and well within strict production budgets.
Limitations and Future Work
- Simulated Triggers: The authors acknowledge that simulating triggers on public datasets provides a "lower bound" on performance, as a real-world click carries a much stronger intent signal than just the "last item in a list."
- Single Trigger Constraint: Currently, TriGen assumes exactly one trigger per request. In some scenarios, a user might have multiple recent signals (e.g., clicking several items in a row). Extending the framework to handle a set of triggers is a suggested direction for future research.
- Semantic ID Dependency: The model relies on pre-constructed Semantic IDs. While effective, the quality of these IDs heavily influences the generation quality. Integrating end-to-end ID learning with trigger modeling could be a promising next step.
Summary updated