Can readers understand their profiles? A study of human involvement in reader profiling

The aim in information filtering is to provide users with a personalised selection of information, based on a description of their interest profile. In some domains, users will want access to such profiles even if they are system generated. We have performed a study of the effects of combining automatic profiling with explicit user involvement. Firstly, we wanted to explore if a machine-learned profile would benefit from being based on an initial explicit user profile. Secondly, we tested if profiles that provided better filtering also were better liked by users. Finally, we tested if users could make improvements to machine-learned profiles. We found that the initial setup of a personal profile was effective, and yielded performance improvements even after feedback training. However, the study showed no correlation between users´ ratings of profiles and their filtering performance, and neither did user modifications to learned profiles improve filtering performance.