Leveraging User Interaction Signals for Web Image Search

摘要

User interfaces for web image search engine results diu000ber signirnfcantly from interfaces for traditional (text) web searchrnresults, supporting a richer interaction. In particular, usersrncan see an enlarged image preview by hovering over a resultrnimage, and an `image preview' page allows users to browsernfurther enlarged versions of the results, and to click-throughrnto the referral page where the image is embedded. No existingrnwork investigates the utility of these interactions asrnimplicit relevance feedback for improving search ranking,rnbeyond using clicks on images displayed in the search resultsrnpage. In this paper we propose a number of implicitrnrelevance feedback features based on these additional interactions:rnhover-through rate, `converted-hover' rate, referralrnpage click through, and a number of dwell time features.rnAlso, since images are never self-contained, but always embeddedrnin a referral page, we posit that clicks on other imagesrnthat are embedded on the same referral webpage as arngiven image can carry useful relevance information aboutrnthat image. We also posit that query-independent versionsrnof implicit feedback features, while not expected to capturerntopical relevance, will carry feedback about the quality orrnattractiveness of images, an important dimension of relevancernfor web image search. In an extensive set of rankingrnexperiments in a learning to rank framework, using a largernannotated corpus, the proposed features give statisticallyrnsignifcant gains of over 2% compared to a state of the artrnbaseline that uses standard click features.