A Two-Stage Prediction Model for Web Page Transition

Utilizing data from a log file, a two-stage model for step-ahead web page prediction that permits adaptive page customization in real-time is proposed. The first stage predicts the next page of a viewer based on a variant of a Markov transition matrix computed from page sequences of other visitors who read the same pages as that viewer did thus far. The second stage re-analyzes the incorrect exit/continuation predictions of the first stage through data mining, incorporating the visitor's viewing behavior observed from the log file. The two-stage process takes advantage of a robust, theory-driven nature of statistical modeling for extracting the overall feature of the data, and a flexible, data-driven nature of data mining to capture any idiosyncrasies and complications unresolved in the first stage. The empirical result with a test site implies that the first stage alone is sufficiently accurate (50.3%) in predicting page transitions. Prediction of site exit was even better with 100% of the exit and 90.8% of the continuation predictions being correct. The result was compared against other models for predictive accuracy.