Relevance feedback strategies for artistic image collections tagging
Proceedings of the 1st ACM International Conference on Multimedia Retrieval
Feature space warping relevance feedback with transductive learning
ACIVS'11 Proceedings of the 13th international conference on Advanced concepts for intelligent vision systems
Toward consistent evaluation of relevance feedback approaches in multimedia retrieval
AMR'05 Proceedings of the Third international conference on Adaptive Multimedia Retrieval: user, context, and feedback
| Hi-index | 0.00 |
Many authors (1, 2, 3, 5, 6, 7) have suggested that overall performance of a document retrieval system is improved by relevance feedback. Relevance feedback denotes the last three steps in the following process: 1) the searcher enters a query, 2) the system prepares a ranked list of suggested documents, 3) the searcher judges some of the documents for relevancy, 4) the searcher informs the system of these documents judged and of the judgement, 5) the system constructs a new query based on the descriptors used in the original query and the descriptors used in the documents judged, 6) the system prepares a second ranked list of suggested documents. The presumption is that the second list is better than the first. By all performance measures (e.g. -&-ldquo;fluid ranking-&-rdquo; and -&-ldquo;frozen ranking-&-rdquo;), the second list is better than the first. However, if one reranks documents in the original list so as to reflect the searcher's efforts (step 3), the corresponding performance measures are comparable to those for the second list. The marginal difference between the performance measures for the -&-rdquo;reranked original-&-rdquo; list (searcher's efforts alone) and the second list (which includes computer efforts) makes it unclear if the cost of steps 4 through 6 above can be justified. It is hoped that advocates of relevance feedback will present -&-ldquo;reranked original-&-rdquo; performance measures as a basis for any performance improvement claims. This paper also presents three reasonable, easily understood retrieval procedures for which the frozen ranking, the fluid ranking, and the reranked original evaluations are -&-ldquo;obviously-&-rdquo; the pertinent way to evaluate. Relevance feedback techniques as implemented in Salton's SMART DRS appear to show that it is worthwhile for user's to read abstracts prior to evaluation of full texts. The last indication presented in this paper is that the relevance feedback performance improvements noted using SMART are due mostly to the user making assessments; subsequent computer efforts appear to be most likely to result in no further change. For a query for which there is a subsequent change, the change is as likely to be harmful as helpful.