Utilizing Variability of Time and Term Content, within and across Users in Session Detection

In this paper, we describe a SVM classification framework of session detection task on both Chinese and English query logs. With eight features on the aspects of temporal and content information extracted from pairs of successive queries, the classification models achieve significantly superior performance than the statof-the-art method. Additionally, we find through ROC analysis that there exists great discrimination power variability among different features and within the same feature across different users. To fully utilize this variability, we build local models for individual users and combine their predictions with those from the global model. Experiments show that the local models do make significant improvements to the global model, although the amount is small.

• Type: Journal
• Year: 2010
• Where: COLING
• Authors: Shu-Qi Sun, Sheng Li, Muyun Yang, Haoliang Qi, Tiejun Zhao
• Keywords: COLING 2010, Computational Linguistics, Global Model, Local Models, SVM Classification Framework