Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Heterogeneous Multi-treatment Uplift Modeling for Trade-off Optimization in Short-Video Recommendation

Chenhao Zhai, Chang Meng, Xueliang Wang, Shuchang Liu, Xiaolong Hu, Shisong Tang, Xiaoqiang Feng, Xiu Li

TL;DR

This work tackles the challenge of optimizing short-video recommendations under heterogeneous multi-treatment scenarios by introducing HMUM, a dual offline-online framework. The Offline Hybrid Uplift Modeling (HUM) module captures both individual causal effects and synergistic interactions across multiple treatments, while the Online Dynamic Decision-Making (DDM) module dynamically estimates user-level value weights for multiple response types to enable personalized trade-offs. Empirical results on public, industrial, and online A/B tests demonstrate HMUM's superiority in offline uplift ranking and real-world gains in consumption metrics with modest commercial improvements, ultimately leading to deployment on a platform serving hundreds of millions of users. The approach advances uplift modeling for video recommendation by jointly modeling heterogeneous treatments and multi-response decision making with KL-divergence constrained representations and MMOE-inspired value weighting, offering practical impact for large-scale systems.

Abstract

The rapid proliferation of short videos on social media platforms presents unique challenges and opportunities for recommendation systems. Users exhibit diverse preferences, and the responses resulting from different strategies often conflict with one another, potentially exhibiting inverse correlations between metrics such as watch time and video view counts. Existing uplift models face limitations in handling the heterogeneous multi-treatment scenarios of short-video recommendations, often failing to effectively capture both the synergistic and individual causal effects of different strategies. Furthermore, traditional fixed-weight approaches for balancing these responses lack personalization and can result in biased decision-making. To address these issues, we propose a novel Heterogeneous Multi-treatment Uplift Modeling (HMUM) framework for trade-off optimization in short-video recommendations. HMUM comprises an Offline Hybrid Uplift Modeling (HUM) module, which captures the synergistic and individual effects of multiple strategies, and an Online Dynamic Decision-Making (DDM) module, which estimates the value weights of different user responses in real-time for personalized decision-making. Evaluated on two public datasets, an industrial dataset, and through online A/B experiments on the Kuaishou platform, our model demonstrated superior offline performance and significant improvements in key metrics. It is now fully deployed on the platform, benefiting hundreds of millions of users.

Heterogeneous Multi-treatment Uplift Modeling for Trade-off Optimization in Short-Video Recommendation

TL;DR

This work tackles the challenge of optimizing short-video recommendations under heterogeneous multi-treatment scenarios by introducing HMUM, a dual offline-online framework. The Offline Hybrid Uplift Modeling (HUM) module captures both individual causal effects and synergistic interactions across multiple treatments, while the Online Dynamic Decision-Making (DDM) module dynamically estimates user-level value weights for multiple response types to enable personalized trade-offs. Empirical results on public, industrial, and online A/B tests demonstrate HMUM's superiority in offline uplift ranking and real-world gains in consumption metrics with modest commercial improvements, ultimately leading to deployment on a platform serving hundreds of millions of users. The approach advances uplift modeling for video recommendation by jointly modeling heterogeneous treatments and multi-response decision making with KL-divergence constrained representations and MMOE-inspired value weighting, offering practical impact for large-scale systems.

Abstract

The rapid proliferation of short videos on social media platforms presents unique challenges and opportunities for recommendation systems. Users exhibit diverse preferences, and the responses resulting from different strategies often conflict with one another, potentially exhibiting inverse correlations between metrics such as watch time and video view counts. Existing uplift models face limitations in handling the heterogeneous multi-treatment scenarios of short-video recommendations, often failing to effectively capture both the synergistic and individual causal effects of different strategies. Furthermore, traditional fixed-weight approaches for balancing these responses lack personalization and can result in biased decision-making. To address these issues, we propose a novel Heterogeneous Multi-treatment Uplift Modeling (HMUM) framework for trade-off optimization in short-video recommendations. HMUM comprises an Offline Hybrid Uplift Modeling (HUM) module, which captures the synergistic and individual effects of multiple strategies, and an Online Dynamic Decision-Making (DDM) module, which estimates the value weights of different user responses in real-time for personalized decision-making. Evaluated on two public datasets, an industrial dataset, and through online A/B experiments on the Kuaishou platform, our model demonstrated superior offline performance and significant improvements in key metrics. It is now fully deployed on the platform, benefiting hundreds of millions of users.

Paper Structure

This paper contains 34 sections, 11 equations, 3 figures, 4 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. PRELIMINARIES
  4. Definition of Uplift Modeling Problem
  5. Definition of Decision-making Problem
  6. METHOD
  7. Offline Hybrid Uplift Modeling
  8. Feature Selection Module.
  9. Individual Causal Effect Branches.
  10. Multi-treatment Joint Optimization and Prediction.
  11. Online Dynamic Decision-Making
  12. Pre-processing of Estimates.
  13. Value Weight Estimating.
  14. Decision-making for Multiple Responses.
  15. OFFLINE EVALUATIONS
  16. ...and 19 more sections

Figures (3)

  • Figure 1: Illustration of the proposed HMUM framework. For the convenience of demonstration, the offline part of the model takes two responses and two treatments as an example.
  • Figure 2: Illustration of the online pipeline.
  • Figure 3: Comparison of the non-treatment output distributions across different branches under APP usage time. The left figure shows independent modeling of each treatment, while the right figure shows joint modeling of each treatment (i.e., HMUM).