Using clustering techniques to detect usage patterns in a Web-based information system

Different users of a Web-based information system will have different goals and different ways of performing their work. This article explores the possibility that we can automatically detect usage patterns without demographic information about the individuals. First, a set of 47 variables was defined that can be used to characterize a user session. The values of these variables were computed for approximately 257,000 sessions. Second, principal component analysis was employed to reduce the dimensions of the original data set. Third, a two-stage, hybrid clustering method was proposed to categorize sessions into groups. Finally, an external criteria-based test of cluster validity was performed to verify the validity of the resulting usage groups (clusters). The proposed methodology was demonstrated and tested for validity using two independent samples of user sessions drawn from the transaction logs of the University of California's MELVYL(R) on-line library catalog system (www.melvyl.ucop.edu). The results indicate that there were six distinct categories of use in the MELVYL system: knowledgeable and sophisticated use, unsophisticated use, highly interactive use with good search performance, known-item searching, help-intensive searching, and relatively unsuccessful use. Their characteristics were interpreted and compared qualitatively. The analysis shows that each group had distinct patterns of use of the system.