Abstract :
[en] Recent rapid growth of data traffic in mobile networks has stretched the capability of current network architectures. Proactively caching popular contents close to end-users has been proposed as a promising approach to mitigate the issue.
A full-fledged proactive cache management mechanism encompasses two interrelated algorithms which need to be carefully designed: content popularity prediction and caching policy. Abundant research has focused on the performance of various caching policies assuming that the content popularity is perfectly known. Nonetheless, the content popularity is unknown in practice and has to be predicted from users' requests. Due to non-deterministic and time-varying nature of the requests, the prediction is nontrivial. In this thesis, the main focus is to introduce efficient prediction algorithms from Bayesian viewpoint. The Bayesian approach provides a powerful framework to construct statistical models which capture uncertainty and are robust to "over-fitting" issue. Firstly, we consider the prediction problem under stationary scenario. To enhance the accuracy of prediction, content features are leveraged and a Bayesian Poisson regressor based on a Gaussian process is proposed. The model can automatically discover hidden patterns in the feature space among the already-existing or seen contents. It also allows to predict the popularities of newly-added or unseen contents whose statistical data is not available in advance. We show that these capabilities of the model can have significant impact on caching performance. Secondly, we formulate a cooperative content caching in order to optimize the aggregated network cost for delivering contents to users. An efficient caching policy requires an accurate prediction of time varying content popularity. The requests can potentially have interactions over time, among contents, and across locations. To exploit these patterns, a probabilistic dynamical model based on a canonical tensor decomposition is developed. Additionally, an online learning method that works with streaming data where content request arrives sequentially is designed. Numerical results confirm that modeling time-content-location interactions by the proposed model can improve the cooperative caching strategy performance. Last but not least, we take one step further and develop a dynamical model, which besides time-content-location interactions, it can also uncover a non-linear temporal trend structure in content requests through which a more accurate prediction can be attained. Subsequently, a cooperative caching policy is designed which adaptively performs network resource allocation and optimizes content delivery according to the dynamic of content requests. Therefore, the policy provides a more efficient utilization of network resources. Using simulations, we show that the developed caching mechanism outperforms reference methods which ignore the temporal trend information.
