Deep Learning-Based View Count Prediction for Content Caching Services

초록

As media data consumption increases, content caching becomes important for low-latency services. To maximize the effect of content caching, we need to find the contents expected to be viewed frequently in the near future. However, the number of expected views for each content is hard to predict because it changes over time, and is intertwined with view trends and correlations with other contents. In this study, we propose a deep learning technique to predict the number of views of contents with high accuracy. In particular, we propose a new data mapping/demapping method before and after the learning model, allowing the learning model to focus on learning trends of view counts excluding the magnitude of view counts. In addition, we propose a deep learning model based on the 1DCNN-LSTM-Dense layers to learn the correlation among multi-attribute data and the correlation of data on the time axis. Based on the YouTube dataset, we evaluate the performance of the proposed learning technique and various previously proposed ones, including heuristic algorithms, normalization techniques, and other data mapping techniques. The results shows that the proposed learning technique achieves the highest caching performance without requiring any future information in advance. © 2023, Korean Institute of Communications and Information Sciences. All rights reserved.

키워드