Web-log cleaning for constructing sequential classifiers

With millions of Web users visiting Web servers each day, the Web log contains valuable information about users' browsing behavior. In this work, we construct sequential classifiers for predicting the users' next visits based on the current actions using association rule mining. The domain feature of Web-log mining entails that we adopt a special kind of association rules we call latest-substring rules, which take into account the temporal information as well as the correlation information. Furthermore, when constructing the classification model, we adopt a pessimistic selection method for choosing among alternative predictions. To make such prediction models useful, especially for small devices with limited memory and bandwidth, we also introduce a model compression method, which removes redundant association rules from the model. We empirically show that the resulting prediction model performs very well.