Application and Evaluation of Multi-Dimensional Diversity

Traditional information retrieval (IR) systems mostly focus on finding documents relevant to queries without considering other documents in the search results. This approach works quite well in general cases; however, this also means that the set of returned documents in a result list can be very similar to each other. This can be an undesired system property from a user’s perspective. The creation of IR systems that support the search result diversification present many challenges, indeed current evaluation measures and methodologies are still unclear with regards to specific search domains and dimensions of diversity. In this paper, we highlight various issues in relation to image search diversification for the ImageClef 2009 collection and tasks. Furthermore, we discuss the problem of defining clusters/subtopics by mixing diversity dimensions regardless of which dimension is important in relation to information need or circumstances. We also introduce possible applications and evaluation metrics for diversity based retrieval.