102. OPTIMIZING GRADED RELEVANCE RANKINGS IN MULITMEDIA DATA

Department: Electrical & Computer Engineering
Research Institute Affiliation: California Institute for Telecommunications and Information Technology (Calit2)
Faculty Advisor(s): Gert Lanckriet

Primary Student
Name: Janani Kalyanam
Email: jkalyana@ucsd.edu
Phone: 732-983-2437
Grad Year: 2014

Student Collaborators
Emanuele Coviello, emanuetre@gmail.com | Brian McFee, bmcfee@cs.ucsd.edu

Abstract
Machine learning plays an integral role in modern information retrieval systems, and in particular, the structured output prediction framework provides a natural learning model for ranking algorithms. Within the structured prediction paradigm, existing ranking algorithms are trained from binary relevance labels: a result is either relevant or irrelevant for the given query. This formalism is often too restrictive for practical applications involving rich multimedia data, where the line between relevance and irrelevance may not be clearly defi ned. In this paper, we present an extension to structured learning to rank which allows for multiple graded levels of relevance. To demonstrate the proposed method, we develop an embedding algorithm which jointly learns transformations across heterogeneous modalities (e.g., images and text) to construct a uni ed embedding space, within which query-by-example retrieval can be performed across modalities. Experimental results demonstrate that training with graded relevance improves accuracy over binary-relevance training, and other existing methods.

« Back to Posters or Search Results