Big data prediction in location-aware wireless caching: A machine learning approach

Abstract

This article investigates a wireless caching framework based on tweets and their location data collected from Twitter. The tweet texts are associated with the location information of the corresponding base stations (BSs) for improving the caching efficiency at BSs. Extracted latent topics and predicted content probability are applied to reduce caching redundancy at BSs. A machine learning approach, namely latent Dirichlet allocation (LDA), is invoked to extract location-aware latent topics for better caching performances. In an effort to predict content probability for caching, a novel skip-gram based long short-term memory (LSTM) model is proposed to cluster words with similar semantics for content probability prediction. Moreover, practical data collected from Twitter is tackled to verify the performance of the proposed framework. Extensive practical tests demonstrate that: 1) Our proposed framework is capable of perceiving caching peaks while the conventional counting method fails; 2) The proposed machine learning approaches are capable of generating accurate topics extraction and content probability prediction results; 3) Our proposed framework maintains superiority over conventional caching approaches and possesses considerable application potential due to its ability of associating with indigenous public preferences.