Improving active browsing with the negative inference and selective search methods.

Abstract

Active Browsing is a technique whereby a learning appretice assists a designer in locating software artifacts in reusable software libraries by inferring the user's search goal from the user's normal browsing actions. The aim of this research is to improve the response time and success rate of Active Browsing. Two methods are proposed for this. The Negative Inference method improves the success rate of active browsing by producing a more accurate representation of the user's goal. The Selective Search method improves the response time of the learning apprentice by limiting the system's evaluation of the library to a fraction of the library. The Negative Inference method adds finer-grained features to the system's internal representation of the user's goal and rules for negative inference (i.e., inferring features that the user is not interested in). The Selective Search method defines a technique for partitioning the library and a strategy, called a migration policy, which determine which items to evaluate. An implementation of both methods, based around a browser used to explore object oriented code, is described. This implementation is used to validate experimentally both methods. With Negative Inference the active browser's success rate is twice that of the normal active browser, and it ranks the search goal much more accurately at all stages of the search. With selective search, the active browser achieves similar success rate while only evaluating a quarter of the library.