부산대학교 도서관

Evaluations of website effectiveness have evolved from metrics based on page hits and visits to assessments of the impact of site changes on the behavior of particular user groups. Typically, this involves the use of clustering algorithms to generate behavior-based groupings, followed by profile-based analyses. One weakness of this approach is that resulting groups are often based on aspects of user behavior not relevant for the revision under study. A second weakness is that profiles may be difficult to interpret. To address these weaknesses we employ 1) a semi-supervised clustering algorithm seeded with evaluator-selected, representative sessions; 2) Time Series Analysis - Selected Episode Graphs (TSA-SEGs), annotated graphical representations of “interesting” aspects of behavior, based on episodes (sequences of requests) and time windows selected by the evaluator. Our approach compared favorably with a leading approach, H-UNC and analysis of profiles created from the resultant clusters, on a case study of an important revision to a biological database site (www.cryptodb.org).