Grouped ECOC Conditional Random Fields for Prediction of Web User Behavior

Web page prefetching has shown to provide reduction in Web access latency, but is highly dependent on the accuracy of the Web page prediction method. Conditional Random Fields (CRFs) with Error Correcting Output Coding (ECOC) have shown to provide highly accurate and efficient Web page prediction on large-size websites. However, the limited class information provided to the binary-label sub-CRFs in ECOC-CRFs will also lead to inferior accuracy when compared to the single multi-label CRFs. Although increasing the minimum Hamming distance of the ECOC matrix can help to improve the accuracy of ECOC-CRFs, it is still not an ideal method. In this paper, we introduce the grouped ECOC-CRFs that allow us to obtain a prediction accuracy closer to that of single multi-label CRFs by grouping the binary ECOC vectors. We show in our experiments that by using the grouping method, we can maintain the efficiency of the ECOC-CRFs while providing significant increase in Web page prediction accuracy over ECOC-CRFs.