On a Scale-Invariant Approach to Bundle Recommendations in Candy Crush Saga

TL;DR

This work addresses bundle recommendations in a large-scale mobile game by proposing a scale-invariant, attention-based three-step pipeline that combines supervised prediction, unsupervised clustering, and discrete bundle construction. It employs TabNet for per-user item-proportion prediction, uses $K$-means to define a discrete preference space, and rounds clusters into actual bundles, optimizing a total cosine-distance error across $d_p$, $d_c$, and $d_o$. The solution is productionized via a four-pipeline MLOps platform, with strong emphasis on reproducibility, configuration management, and monitoring of business metrics (take rate, click rate) and diversity to mitigate degenerate feedback loops. Offline and online experiments show substantial engagement uplifts compared to baselines (up to ~130% in novelty scenarios and consistent improvement in take and click rates), while also highlighting the persistence of diversity drift and the importance of continuous retraining. Overall, the paper delivers a scalable, maintainable pipeline for in-game bundle recommendations and contributes practical lessons on tech debt prevention and performance-variance trade-offs in production ML systems.

Abstract

A good understanding of player preferences is crucial for increasing content relevancy, especially in mobile games. This paper illustrates the use of attentive models for producing item recommendations in a mobile game scenario. The methodology comprises a combination of supervised and unsupervised approaches to create user-level recommendations while introducing a novel scale-invariant approach to the prediction. The methodology is subsequently applied to a bundle recommendation in Candy Crush Saga. The strategy of deployment, maintenance, and monitoring of ML models that are scaled up to serve millions of users is presented, along with the best practices and design patterns adopted to minimize technical debt typical of ML systems. The recommendation approach is evaluated both offline and online, with a focus on understanding the increase in engagement, click- and take rates, novelty effects, recommendation diversity, and the impact of degenerate feedback loops. We have demonstrated that the recommendation enhances user engagement by 30% concerning click rate and by more than 40% concerning take rate. In addition, we empirically quantify the diminishing effects of recommendation accuracy on user engagement.

Figures (5)

• Figure 1: In-game bundle recommendation.
• Figure 2: Pipeline Overview
• Figure 3: Novelty effect in experiment 1.
• Figure 4: Take rates uplifts in experiment 4.
• Figure 5: Recommendation diversity development over time.