Popularity Tracking for Proactive Content Caching with Dynamic Factor Analysis

The performance of all proactive edge-caching policies is largely influenced by the content popularity prediction algorithm, which in turn relies on having a good understanding of the underlying request process. However, due to the non-deterministic and time-varying nature of popularities, prediction is not a trivial task. In this work, we suggest a probabilistic dynamic factor analysis model to describe content requests for real-world time-varying scenarios. The dynamic factor analyzer is a flexible model to capture common patterns among content requests through temporal stochastic processes, which lay in a low-dimensional latent space. Modeling these common dynamic patterns provides a better accuracy for tracking and predicting the evolution of content popularities. Model learning is performed from the Bayesian aspect which provides a systematic mechanism to incorporate uncertainty for robust prediction. Due to the model complexity, we derive a simple approximation method based on variational Bayes (VB) to infer the model parameters. Finally, we show simulation results where the proposed popularity prediction method outperforms the traditional ones available in the literature on a real-world dataset.