Personalized Interface Generation via Constrained Contextual Bandits at Billion-User Scale
DOI:
https://doi.org/10.63282/3117-5481/WFCMLS26-104Keywords:
Server-Driven UI, Adaptive Interfaces, ML-Optimized Rendering, Emerging Markets, Contextual Bandits, Engagement Optimization, Device- Tier AdaptationAbstract
The next billion mobile users are arriving on devices that flagship- market UI frameworks were never designed to serve. Rather than maintaining parallel codebases—as exemplified by Facebook Lite and Instagram Lite—this paper proposes a unified architecture where the UI is dynamically composed by the server and continuously optimized by ML models that learn the right interface for every user-device-context combination. We present AIDE (Adaptive Interface Delivery Engine), a frame- work combining Server-Driven UI (SDUI) for real-time layout com- position with contextual bandit models that personalize UI con- figurations across 40+ emerging-market signals including device capability, measured network quality, battery state, and data-cost sensitivity. The server determines not only what content to show, but how to render it—which component variants to assemble, at what fidelity, in what layout—without requiring client binary up- dates. Evaluation over six-week A/B experiments across 2.4B+ monthly active devices shows statistically significant improvements on Lite- tier devices versus dedicated lite applications: 34% lower crash rates (???? < 0.001), 41% faster Time-to-Interactive (???? < 0.001), 22% longer session duration (???? < 0.01), and 50% broader feature usage—all from a single codebase with under 3% code divergence. We also report an ablation study isolating the ML model’s contribution at +18% session duration over rule-based SDUI.
References
[1] Meta Engineering. Building Facebook Lite: Engineered for the Developing World. Meta Engineering Blog, 2016.
[2] Meta Engineering. Reintroducing Instagram Lite: A Lighter Way to Use Instagram. Meta Engineering Blog, 2021.
[3] M. Hasan et al. Characterizing the Android Ecosystem in Emerging Markets. In ACM MobiSys, 2024.
[4] Y. Liu et al. Adaptive Bitrate Streaming: A Survey. IEEE Communications Surveys & Tutorials, 26, 2024.
[5] Google. Android Vitals: Building for Billions. Android Developer Documentation, 2025.
[6] A. Pathak et al. Fine-Grained Energy Profiling for Power-Aware Application Design. In ACM EuroSys, 2023.
[7] S. Kumar et al. Network Quality Estimation on Mobile Devices. IEEE Transactions on Mobile Computing, 23, 2024.
[8] J. Chen et al. Adaptive Rendering for Heterogeneous Mobile GPUs. In ACM SIGGRAPH Mobile, 2024.
[9] R. Rosen et al. Thermal-Aware Task Scheduling in Mobile Systems. In USENIX ATC, 2023.
[10] P. Aggarwal et al. Understanding Mobile User Behavior in Bandwidth- Constrained Environments. In ACM CHI, 2024.
[11] D. Ferreira et al. Battery-Aware Application Design Patterns for Mobile Platforms. In IEEE PerCom, 2024.
[12] L. Li et al. A Contextual-Bandit Approach to Personalized News Article Recom- mendation. In WWW, 2010.
[13] A. Agarwal et al. Exploration–Exploitation in Mobile App UI Optimization. In KDD, 2023.
[14] Meta. Server-Driven Rendering for Mobile Applications at Scale. Meta Systems Research, 2024.
[15] W. Enck et al. Memory Management Strategies for Resource-Constrained Android Devices. In ACM MobiSys, 2023.
[16] GSMA. The Mobile Economy 2025: Emerging Markets. GSMA Intelligence Report, 2025.
[17] D. Agarwal et al. Overlapping Experiment Infrastructure at LinkedIn. In KDD, 2017.
[18] M. Dudik et al. Doubly Robust Policy Evaluation and Optimization. In ICML, 2014.
