Predictive Caching via Learning Temporal Distribution of Content Requests

Abstract

In this letter, dynamic content placement of a local cache server that can store a subset of content objects in its cache memory is studied. Contrary to the conventional model in which content placement is optimized based on the time-invariant popularity distribution of content objects, we consider a general time-varying popularity distribution and such a probabilistic distribution is unknown for content placement. A novel learning method for predicting the temporal distribution of future content requests is presented, which utilizes the request histories of content objects whose lifespans are expired. Then we introduce the so-called predictive caching strategy in which content placement is periodically updated based on the estimated future content requests for each update period. Numerical evaluation is performed using real-world datasets reflecting the inherent nature of temporal dynamics, demonstrating that the proposed predictive caching outperforms the conventional online caching strategies.