Utilizing group prediction by users' interests to improve the performance of web proxy servers

Companies and institutions often use Web proxy servers to service the multiple requests of the same Web pages (or Web objects) from users therein to save the network bandwidth and reduce the Internet latency. Web proxy servers usually are geographically close to their clients (users). If Web proxy serves have cached valid copies of requested Web objects, they can be directly delivered to users. Otherwise, users need to spend a long time on getting Web objects from their hosting Web servers. Both cases will require users to wait for their requests. This period of latency can be largely reduced if Web proxy servers could predict what Web objects users may need in the near future, and send those predicted Web objects to the client sites before their actual usage. Various predicting algorithms, such as those based on temporal locality and data mining, had been proposed to enable proxy servers to make prediction of what Web objects users may access. However, they often need to constantly update and maintain their models in realtime to make their schemes effective. As a result, for proxy servers servicing a large number of users, schemes demanding for complicated and realtime calculation will not be as useful as expected in practice. Based on Web sites and Web pages commonly visited by users, we proposed a new model with an offline learning algorithm to help Web proxy servers make prediction about upcoming requests of Web objects from users. Compared with the hit ratio achieved by the original environment without prediction, our model can improve the caching performance of users' Web browsers by up to 51.37%.