Interactive search with weak supervisionTools Wood, Stuart Thomas (2017) Interactive search with weak supervision. [Dissertation (University of Nottingham only)]
AbstractThis work investigates the possible applications of functional clustering algorithms in relevance feedback algorithms for interactive search. A modified version of the Rocchio algorithm for relevance feedback is proposed based off the result of an LGA clustering of the data. The effectiveness of this algorithm is assessed based on simulations using a tiered probabilistic model for user feedback. It is found that the for a poor initial query (one that is far from the users target query in a vector space model), the simulations run with the modified algorithm reach a close vicinity of the target search faster than the standard Rocchio algorithm even when the increased distance between the current query and the options presented is accounted for. A strong inverse relationship is observed between the strength of the linear structures in the data (measured based on the gap statistic of the data set clustered under LGA) and the improvement in search refinement. Data with gap statistic of 0 offers similar performance to the standard Rocchio algorithm once corrected for the increased distance, suggesting that exploiting linear structures in the data can offer more efficient searching.
Actions (Archive Staff Only)
|
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)
