How current web generation affects prediction algorithms performance

Web prefetching techniques have been widely used to reduce the latency perceived by users. An important part of these techniques is the prediction algorithm whose main aim is to predict the future users’ accesses based on the previous visiting patterns. For this purpose, the algorithms proposed in the open literature usually consider a training period before making predictions. The length of this period may impact on performance, either improving or degrading it. In addition, this length may involve a high amount of information and therefore important computer resources are consumed. In this work we analyze how the training affects the prediction performance using current and old web traces, still used in the recent literature. Our results show that while in old traces the training, in general, improves performance, when using recent traces this training may degrade the performance. This fact happens because those users’ visiting patterns have changed, what means that old traces are not appropriate for prediction studies in current web environments.