Intelligent Emoji Prediction with Dynamic Fusion for Mobile Keyboards

Abstract

In the realm of mobile keyboard technology, the integration of emoji prediction has transformed user interactions. However, deploying predictive models on edge servers poses challenges, especially without prior knowledge of critical metrics like prediction accuracy and inference latency. This paper introduces "dynamic model fusion," a novel concept that addresses this challenge. By dynamically fusing multiple predictive models based on real-time feedback and historical performance data, our approach surpasses conventional model selection techniques. Through the integration of history-aware online learning, online control, and dynamic model fusion methodologies, our scheme optimizes energy efficiency while achieving unprecedented levels of accuracy and latency reduction in emoji prediction tasks. Theoretical analysis establishes the robustness of our approach, showcasing a sub-linear round-averaged regret bound, while extensive simulation results underscore its superior performance in real-world scenarios. This paradigm shift towards dynamic model fusion not only redefines mobile keyboard intelligence but also heralds a new era of adaptive and resource-efficient computing paradigms.