Mining ordinal data under human response uncertainty

Analysis and interpretation of collective feedback on ordinal scales is an important issue for several disciplines, including social sciences, recommender systems research, marketing, political science, and many others. A "reasonable" model is expected to provide an "explanation" of collective user behaviour. Many existing data mining approaches employ for this purpose probabilistic models, based on distributions and mixtures from a certain parametric family. In real life, users meet their decisions with considerable uncertainty. Its assessment and use in probabilistic models for better interpretation of collective feedback is the key concern of this paper. In doing so, we introduce approaches for gathering individual uncertainty, and discuss their viability and limitations. Consequently, we enrich state of the art response mining models (especially focused on discovery of latent user groups) with uncertainty knowledge, and demonstrate resulting advantages in systematic experiments with real users.