A Novel Adaptive Caching Framework for Dynamic Web Portals
Explored adaptive caching strategies using machine learning to improve dynamic web portal performance. Demonstrated significant latency reduction by predicting temporal data access patterns.
Research Motivation
Dynamic web portals suffer from high latency during traffic spikes due to rigid, heuristic-based caching policies (like LRU or LFU) that fail to adapt to complex user access patterns.
Problem Statement
Can machine learning models accurately predict near-future data access patterns to proactively cache resources and evict stale data, thereby maximizing cache hit rates?
Methodology & Architecture
Implemented a predictive model using sequence-to-sequence LSTM networks to forecast URL access probability. The output dynamically adjusts cache TTLs and pre-fetches high-probability resources.
Results & Findings
Achieved a 25% improvement in cache hit ratio and a 15% reduction in tail latency under synthetic load testing.
Discussion
Lessons Learned
Model inference overhead can negate caching benefits if not heavily optimized; lightweight models (e.g., Random Forests) often outperform deep networks in strict real-time caching.