Predicting user requests in practical VoD systems

In the last 2 decades, VoD (Video-on-Demand) becomes more and more popular. Its widely use threw out a big challenge to the limited bandwidth of the Internet. In this paper, we propose a new method of predicting user requests of videos based on data recorded during the last week. Then we improvement its accuracy by distinguishing between time. To “hot” hours (when there are lots of requests), we calculate the weights with differences between them and the benchmark time (e.g. 9 am) instead of the original requests, while “cold” hours just repeat the same trend of the previous day. With prediction, we are able to divide videos into different classes and adopt dedicated delivery strategy for each class, to minimize the total bandwidth cost. We conduct simulation on real data captured from real VoD systems and our method presents a good performance.