Predictive Caching in Mobile Streaming Applications using Machine Learning Models

Abstract

Mobile video streaming constitutes the majority of internet traffic on cellular networks, yet the unpredictable nature of wireless links continues to challenge seamless playback. Traditional caching strategies (e.g., Least Recently Used and Least Frequently Used) respond only to past access patterns and often fail under real-time variability such as handovers, jitter, and throughput swings. This paper presents a Machine Learning (ML)-based predictive caching framework that proactively prefetches video segments by forecasting viewer intent and network conditions. The approach combines behavioral and contextual features—watch sequences, genre affinity, session depth, network bandwidth variance—and applies Random Forest and XGBoost classifiers to rank likely next requests. Prefetching is executed when a confidence threshold is exceeded, constrained by device storage and instantaneous throughput. Simulated experiments across 3G/4G/5G models demonstrate a 24% increase in cache hit ratio, 31% reduction in startup latency, and 47% drop in rebuffering relative to heuristic caching. The contributions include (i) a modular architecture deployable at the edge or client; (ii) an analysis of model trade-offs (accuracy vs. inference latency); and (iii) a discussion of privacy and deployment considerations for real-world OTT pipelines.