Modelling long-term relevance feedback

We propose a general relevance model, called the User Relevance Model, that formalises the decisions taken by a user during a query with respect to relevance judgements. Starting from a keyword-based query, the user is allowed to refine the document search using relevance feedback iterations where some subset of the result set is marked as relevant, and another subset is marked as non-relevant. The model postulates that observed relevance judgements stem from the existence or lack thereof of underlying topics or concepts common to both documents and the query. Furthermore, it explains the underlying concepts through the estimation parameters using a latent-variable model, non-negative matrix factorisation. Experiments are carried out on artificial relevance feedback judgements generated using the model.