A generic pseudo relevance feedback framework with heterogeneous social information

Social information has been widely applied in the research fields of information retrieval and recommendation systems. However, most research efforts have been conducted to investigate social information with a specific type (e.g. tags), which ignores the relationships among these various types of heterogeneous social information. In this paper, taking Pseudo Relevance Feedback (PRF) as an example of classical IR models, we propose a novel and generic PRF framework for searching and recommendation with heterogeneous social information. Firstly, through the representations, social information is categorized into structured and unstructured information, then a correlation graph for terms occurred in queries and social information is built to express rich semantic contexts for their relationships. Inspired by the correlation graph, we respectively design diverse term selection strategies for structured information (e.g. tags, ratings) and unstructured information (e.g. reviews, annotations). Particularly, we modify the conventional distribution-based term selection models for structured information due to the “semantic indivisibility”, meanwhile we also introduce several transformation term selection models for unstructured information with an additional consideration of ratings when constructing weighting functions. Finally, we construct an effective searching system, where learning-to-rank is utilized adaptively to combine results from various PRF strategies with different social information. Extensive evaluations and comparisons on CLEF/INEX Social Book Search Track (SBS) and three other collections verify the effectiveness and robustness of the proposed framework and term selection models. Furthermore, statistics indicate that the historic records of all 4-year SBS tracks are refreshed by our proposed method.