STCRank: Spatio-temporal Collaborative Ranking for Interactive Recommender System at Kuaishou E-shop

Boyang Xia, Ruilin Bao, Hanjun Jiang, Jun Wang, Wenwu Ou

TL;DR

Kuaishou researchers have developed STCRank, a ranking framework designed for immersive, swipe-based shopping feeds. By balancing conflicting objectives (like "buying" vs. "keep swiping") and using a look-ahead mechanism to optimize the entire user session rather than just the next item, they achieved significant growth in both purchases and daily active users.

What Problem Does This Paper Address?

Traditional e-commerce recommendation happens on a homepage (the "Explore" stage). However, Kuaishou E-shop uses an Interactive Recommender System (IRS): when a user clicks an item, they enter a "Focus stage"—a full-screen, immersive vertical feed where they can swipe down to see related products.

This UI creates two major headaches for standard ranking algorithms:

  1. Objective Interference (Spatial Challenge): In a swipe-feed, behaviors overlap. A "view" often precedes a "purchase," but a "purchase" often leads to an "exit," which conflicts with the goal of "keep swiping" (engagement). Standard models struggle to balance these competing signals.
  2. The Temporal Greedy Trap: Because the UI is full-screen, the system only shows one item at a time. If the ranker greedily picks the item most likely to be bought right now, it might inadvertently end the session early, missing out on a sequence of items that could have led to higher total value or better user exploration.

Methods

The authors propose STCRank, which consists of two main modules:

1. Multi-Objective Collaboration (MOC)

To handle conflicting goals, MOC refines how labels are defined:

  • VTR (View-Through Rate) Refinement: Instead of a generic "view" label, they found a "sweet spot" at 5 seconds. Shorter views are too noisy; longer views overlap too much with purchase intent.
  • SDR (Swipe-Down Rate) Filtering: They implemented a "Conflict Sample Filter." Since users often exit immediately after buying, a raw swipe-down model would learn that "purchased items are bad for engagement." MOC redefines these samples to ensure engagement goals don't suppress conversion goals.

2. Multi-Slot Collaboration (MSC)

This module moves from "point-wise" ranking (is this item good?) to "session-wise" ranking (is this sequence good?):

  • Cross-Stage Look-Ahead: The homepage (E-stage) ranker is modified to favor items that are likely to trigger high-value "Focus" sessions.
  • Single-Stage Look-Ahead: Within the swipe feed, the system uses Beam Search to evaluate permutations of items. It calculates the "cumulative sequence value," accounting for the probability that a user will actually swipe down to reach the 3rd or 4th item in the list.

Datasets

The study was conducted using massive, real-world production data from Kuaishou E-shop.

  • Training: Uses real-time user feedback loops (clicks, views, purchases, swipes).
  • Evaluation: Results are based on large-scale online A/B tests against the live production environment.

Key Insights and Ideas

  • The 5-Second Rule: Correlation between viewing time and purchase isn't linear. By setting a 5-second threshold for "quality views," the model captures interest without becoming redundant with the purchase model.
  • Purchases are "Exit" Events: In immersive feeds, a conversion is often the end of a session. If you don't account for this, your engagement model will actively avoid showing "buyable" items because they "kill" the swipe streak.
  • Permutation Matters: The order of items significantly changes the "exposure probability" of subsequent items. MSC treats the feed as a path rather than a bucket of items.

Key Results

The framework was deployed at Kuaishou E-shop in June 2025, showing impressive gains:

  • Overall Engagement: +9.65% increase in Item-detail Page Views (IPV).
  • Business Impact: +1.55% increase in total purchases and a +0.03% lift in Daily Active Users (DAU).
  • High-Involvement Categories: The system performed exceptionally well for "high-decision-cost" goods. For example, it saw a +42% contribution to purchase gain in Women’s Shoes, where users typically need to compare multiple items before committing.

Limitations and Future Work

  • Manual Thresholds: While the 5-second VTR threshold worked well, it was determined empirically. Future work could involve dynamic, user-specific thresholds for "quality engagement."
  • Long-term LTV: The current look-ahead mechanism focuses on the immediate session. The authors suggest that extending this to optimize for long-term User Lifetime Value (LTV) across days or weeks is a promising next step.
  • Computational Cost: Using Beam Search for re-ranking adds latency. While the authors used pruning to keep it manageable, scaling this to even larger candidate sets remains a challenge.

Summary updated

Abstract

As a popular e-commerce platform, Kuaishou E-shop provides precise personalized product recommendations to tens of millions of users every day. To better respond real-time user feedback, we have deployed an interactive recommender system (IRS) alongside our core homepage recommender system. This IRS is triggered by user click on homepage, and generates a series of highly relevant recommendations based on the clicked item to meet focused browsing demands. Different from traditional e-commerce RecSys, the full-screen UI and immersive swiping down functionality present two distinct challenges for regular ranking system. First, there exists explicit interference (overlap or conflicts) between ranking objectives, i.e., conversion, view and swipe down. This is because there are intrinsic behavioral co-occurrences under the premise of immersive browsing and swiping down functionality. Second, the ranking system is prone to temporal greedy traps in sequential recommendation slot transitions, which is caused by full-screen UI design. To alleviate these challenges, we propose a novel Spatio-temporal collaborative ranking (STCRank) framework to achieve collaboration between multi-objectives within one slot (spatial) and between multiple sequential recommondation slots. In multi-objective collaboration (MOC) module, we push Pareto frontier by mitigating the objective overlaps and conflicts. In multi-slot collaboration (MSC) module, we achieve global optima on overall sequential slots by dual-stage look-ahead ranking mechanism. Extensive experiments demonstrate our proposed method brings about purchase and DAU co-growth. The proposed system has been already deployed at Kuaishou E-shop since 2025.6.
BibTeX
@inproceedings{xia2026stcrank,
  author = {Boyang Xia and Ruilin Bao and Hanjun Jiang and Jun Wang and Wenwu Ou},
  title = {{STCRank: Spatio-temporal Collaborative Ranking for Interactive Recommender System at Kuaishou E-shop}},
  booktitle = {The Web Conference},
  year = {2026},
  doi = {10.1145/3774905.3794662},
}
Figure