8. Current Trends and Future Directions

Relevance feedback is a powerful tool for users to interactively improve their search results. There is extensive evidence in the fields of Information Retrieval and Multimedia Retrieval to support continued efforts in improving techniques to learn from relevance feedback. Several promising directions for using relevance feedback are being explored, including techniques to speed up relevance feedback iterations and finding better ways to identify relevant objects.

8.1 Kernel Approaches

A promising research direction is to explore features in multidimensional spaces of much higher dimensionality than that of the original feature space. For example, [6][9] use kernels distances to effectively map a feature representation space into a much higher dimensional space where they search for suitable separators that cannot be easily represented in the original feature space. [6] uses a support vector machine approach to find linear separators in a very high dimensional space to better characterize relevant results. Another support vector machine approach described in [30] does not return the closest results to the user's query at each iteration; instead it shows different images in hopes to maximize the learning at each iteration.

8.2 Feature Selection

Another approach to using a high dimensional search space is MEGA [1]. MEGA is a query refinement method that starts a general initial query for all users that is composed of a k-CNF Query Concept Space (QCS) and a k-DNF Candidate Concept Space (CCS). This method is based on Valiant's PAC learning model [31] along with a bounded sampling technique that aims at removing the maximum expected number of disjunctive terms from QCS. The refinement involves removing irrelevant conjunctive terms (clauses) from the CCS based on negative examples and removing of irrelevant disjunctive terms from the QCS based on positive examples. The MEGA approach is applicable only in situations where user-specific initial queries are not supplied. MEGA does not provide methods to learn query values and weights. The approach also requires handling an exponential number of terms in queries that results in high learning complexity (the authors suggest a dimensionality reduction method to tackle this). Moreover, it is not clear how CCS and QCS are formed for continuous valued attributes.

Beyond applications in multimedia retrieval, relevance feedback has general applicability to database searches. [28] explores an interactive browser for databases that uses relevance feedback. In contrast, [16] describes a general way to incorporate relevance feedback on arbitrary data types in a database server.

Relevance feedback for multimedia search is a valuable tool that every retrieval system should include. The opportunities for a retrieval system to help the user find her information are tremendous. The retrieval performance improvement that is achievable through relevance feedback frequently improves results so much that users barely believe that the system was able to "outsmart" them and construct better queries. Current developments in relevance feedback techniques will improve retrieval results even more.