Discovering Hidden Contextual Factors for Implicit Feedback

This paper presents a statistical framework based on Principal Component Analysis (PCA) for discovering the contextual factors which most strongly influence user behavior during information-seeking activities. We focus particular attention on explaining how PCA can be used to assist in the discovery of contextual factors. As a demonstration of the utility of PCA, we employ it in an Implicit Relevance Feedback (IRF) algorithm that observes features of user interaction, computes the feature co-variances from a few seen documents, and calculates the eigenvectors of the co-variance matrix to be used as the basis for ranking the unseen documents. This ranking is then compared with the ideal ranking that could be computed if the ratings explicitly given by the user were known. The most effective eigenvector, in terms of impact on retrieval performance, was chosen as representative of each user’s intent. Our experiments showed that each aspect of user behavior is influenced by different contextual factors, yet there exist some important features common to most factors. Our findings demonstrate both the effectiveness of the IRF algorithm and the potential value of incorporating multiple sources of interaction evidence in their development. In particular, it was shown that IRF was more effective when the eigenvectors are personal-