Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hierarchical Contextual Uplift Bandits for Catalog Personalization

Anupam Agrawal, Rajesh Mohanty, Shamik Bhattacharjee, Abhimanyu Mittal

TL;DR

The paper addresses rapid non-stationarity in fantasy-sports catalog personalization by introducing the Hierarchical Contextual Uplift Bandit (HCUB), which uses a coarse-to-fine hierarchical context tree and an uplift-based reward objective to enable efficient policy transfer and mitigate cold-start. HCUB constructs an 81-arm action space by jointly adjusting four contest attributes, computes uplifts via hierarchical bootstrapping, and propagates rewards top-down while selecting arms bottom-up with Bayesian UCB. In online A/B tests on the Dream11 platform, HCUB yields revenue uplifts of approximately $0.42\%$ in a 6M-user stage and $0.51\%$ after production deployment, with modest, statistically insignificant changes in engagement metrics; offline simulations show 4–5\% improvements in regret when employing reward inheritance. The results demonstrate practical impact for large-scale catalog personalization in dynamic environments and motivate future work on adaptive bandits with non-stationarity handling to broaden applicability beyond fantasy sports.

Abstract

Contextual Bandit (CB) algorithms are widely adopted for personalized recommendations but often struggle in dynamic environments typical of fantasy sports, where rapid changes in user behavior and dramatic shifts in reward distributions due to external influences necessitate frequent retraining. To address these challenges, we propose a Hierarchical Contextual Uplift Bandit framework. Our framework dynamically adjusts contextual granularity from broad, system-wide insights to detailed, user-specific contexts, using contextual similarity to facilitate effective policy transfer and mitigate cold-start issues. Additionally, we integrate uplift modeling principles into our approach. Results from large-scale A/B testing on the Dream11 fantasy sports platform show that our method significantly enhances recommendation quality, achieving a 0.4% revenue improvement while also improving user satisfaction metrics compared to the current production system. We subsequently deployed this system to production as the default catalog personalization system in May 2025 and observed a further 0.5% revenue improvement.

Hierarchical Contextual Uplift Bandits for Catalog Personalization

TL;DR

The paper addresses rapid non-stationarity in fantasy-sports catalog personalization by introducing the Hierarchical Contextual Uplift Bandit (HCUB), which uses a coarse-to-fine hierarchical context tree and an uplift-based reward objective to enable efficient policy transfer and mitigate cold-start. HCUB constructs an 81-arm action space by jointly adjusting four contest attributes, computes uplifts via hierarchical bootstrapping, and propagates rewards top-down while selecting arms bottom-up with Bayesian UCB. In online A/B tests on the Dream11 platform, HCUB yields revenue uplifts of approximately in a 6M-user stage and after production deployment, with modest, statistically insignificant changes in engagement metrics; offline simulations show 4–5\% improvements in regret when employing reward inheritance. The results demonstrate practical impact for large-scale catalog personalization in dynamic environments and motivate future work on adaptive bandits with non-stationarity handling to broaden applicability beyond fantasy sports.

Abstract

Contextual Bandit (CB) algorithms are widely adopted for personalized recommendations but often struggle in dynamic environments typical of fantasy sports, where rapid changes in user behavior and dramatic shifts in reward distributions due to external influences necessitate frequent retraining. To address these challenges, we propose a Hierarchical Contextual Uplift Bandit framework. Our framework dynamically adjusts contextual granularity from broad, system-wide insights to detailed, user-specific contexts, using contextual similarity to facilitate effective policy transfer and mitigate cold-start issues. Additionally, we integrate uplift modeling principles into our approach. Results from large-scale A/B testing on the Dream11 fantasy sports platform show that our method significantly enhances recommendation quality, achieving a 0.4% revenue improvement while also improving user satisfaction metrics compared to the current production system. We subsequently deployed this system to production as the default catalog personalization system in May 2025 and observed a further 0.5% revenue improvement.
Paper Structure (13 sections, 2 equations, 2 figures, 1 table, 1 algorithm)

This paper contains 13 sections, 2 equations, 2 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Literature
  3. Hierarchical Bandits
  4. Causal Bandits
  5. Methodology
  6. Catalog Representation and Action Space
  7. Reward Function
  8. Hierarchical Context and Reward Inheritance
  9. Evaluation
  10. Online Experiments
  11. Offline Simulations
  12. Results
  13. Conclusions and Future Work

Figures (2)

  • Figure 1: The Contest Selection page on the Dream11 mobile app. It shows the current state of available contests for a particular match. More contests are available by scrolling up or down.
  • Figure 2: Hierarchical Contextual Tree representation illustrating the relationship between system-level and user-level features in contextual bandits. The root node represents system-level features, with branches corresponding to distinct cohorts. Feature granularity increases along the depth of the tree, culminating in user-level features at the leaf nodes. The shaded regions highlight an example context path from the root to leaf nodes. The uplift for each context-action pair are computed relative to a base action. Uplift values calculated at leaf nodes are aggregated upwards (represented by the solid arrows), while rewards propagate downward from parent to child nodes (represented by the dotted arrows).