Relevance feedback as distribution generation

Multimedia database interfaces should be designed to be very user-adaptive, since there is no generally applicable model of user's search behavior or of his search intention. First, the challenging task for the interface is to present the most representative objects in an appealing and concise manner. Second, the interface has to identify the user's search intention from very few positive feedbacks. In particular for the latter there exist a lot of Relevance Feedback implementations. While most of them are considered as more or less heuristically proved parameter adjustment procedures, we treat Relevance Feedback as direct probability density estimation. Our density is defined as the "hidden user's search intention" (HUSI), and - due to a lack of better criteria - it is derived directly from the feature space. Intuitively and positively associated objects influence a certain probability density by their characteristic, each for every dimension of its features. This dens ity may be derived from a Bayesian approach and is regarded as the marginal distribution of the joint distribution presenting the HUSI. Then we generate samples, satisfying the HUSI, and we present them for a further iterative refinement. In addition, since the user can express his global preference by influencing the linear PCA-compression, we adjust the metrics as well as the HUSI. Our technique provides search results very close to psychological expectations