CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval

TL;DR

This work tackles trigger selection for multi-channel U2I2I retrieval by reframing it as a channel-aware routing problem that optimizes downstream utility. It introduces two components: the Value Attribution Module (VAM), which provides look-ahead supervision by crediting each trigger with the downstream engagement of items retrieved through it, and the Channel-Adaptive Trigger Routing (CATR) module, which learns per-channel value predictions and diversity-aware routing to coordinate trigger-to-channel allocation. The approach yields significant offline gains in multi-channel Recall@K and measurable online improvements in engagement metrics on a large-scale production platform, including a +0.351% lift in average time spent per device. These results demonstrate the practical impact of aligning trigger value with downstream utility and of coordinating multi-channel routing to reduce redundancy and improve coverage. Overall, CAPTS offers a scalable, production-friendly solution for trigger selection that enhances both the breadth and quality of multi-channel candidate generation.

Abstract

Large-scale industrial recommender systems commonly adopt multi-channel retrieval for candidate generation, combining direct user-to-item (U2I) retrieval with two-hop user-to-item-to-item (U2I2I) pipelines. In U2I2I, the system selects a small set of historical interactions as triggers to seed downstream item-to-item (I2I) retrieval across multiple channels. In production, triggers are often selected using rule-based policies or learned scorers and tuned in a channel-by-channel manner. However, these practices face two persistent challenges: biased value attribution that values triggers by on-trigger feedback rather than their downstream utility as retrieval seeds, and uncoordinated multi-channel routing where channels select triggers independently under a shared quota, increasing cross-channel overlap. To address these challenges, we propose Channel-Aware, Preference-Aligned Trigger Selection (CAPTS), a unified and flexible framework that treats multi-channel trigger selection as a learnable routing problem. CAPTS introduces a Value Attribution Module (VAM) that provides look-ahead supervision by crediting each trigger with the subsequent engagement generated by items retrieved from it on each I2I channel, and a Channel-Adaptive Trigger Routing (CATR) module that coordinates trigger-to-channel assignment to maximize the overall value of multi-channel retrieval. Extensive offline experiments and large-scale online A/B tests on Kwai, Kuaishou's international short-video platform, show that CAPTS consistently improves multi-channel recall offline and delivers a +0.351% lift in average time spent per device online.

Figures (6)

• Figure 1: Multi-channel retrieval with two candidate-generation paradigms: direct U2I and two-hop U2I2I. We study trigger selection in U2I2I, where a small set of history items is selected to seed downstream I2I retrieval channels.
• Figure 2: Challenges in multi-channel U2I2I trigger selection and CAPTS. (a) Conventional trigger selection often relies on direct feedback on the trigger item itself and feeds the selected triggers to multiple downstream I2I channels, which can lead to biased value attribution and uncoordinated multi-channel routing. (b) CAPTS addresses these issues with two modules: VAM for look-ahead value attribution and CATR for channel-adaptive trigger routing across I2I channels.
• Figure 3: Overview of the CAPTS framework. (a) VAM: at request time $\tau_0$, for each candidate trigger $t$ and retrieval channel $c$, VAM replays I2I retrieval to obtain $\mathcal{R}_c(t,\tau_0)$. It matches retrieved items to a fixed future window $\mathcal{W}_u^{\mathrm{fut}}(\tau_0)$ and aggregates subsequent engagement to form channel-specific supervision for each trigger. (b) CATR: a shared encoder with target attention produces a trigger-aware representation. For each channel, a value predictor with a bounded calibrator estimates the channel-specific trigger value, while a uniqueness head captures cross-channel complementarity. Routing scores combine value and uniqueness for trigger selection.
• Figure 4: Production deployment of CAPTS. Orange lines denote offline value attribution and CATR training, with periodic model-weight synchronization to online serving. Blue lines denote per-request online trigger scoring and routing for multi-channel I2I retrieval. Grey lines denote daily nearline jobs that refresh the nearline trigger cache for long-term triggers.
• Figure 5: Sensitivity of Recall@K to the future window size $w_s$ on Swing, Marm, and MMU I2I. We vary $w_s \in \{50,100,150,200\}$ in the value aggregation model while keeping other training settings fixed, and plot Recall@K for $K \in \{100,500,1000,2000\}$.