There is no question that accuracy is an important requirement of classification and prediction models used in software engineering management. It is, however, just one of a number of attributes that contribute to a model being 'useful'. Understandably much research has been undertaken with the objective of maximizing model accuracy, but this has often occurred with little regard for these other model attributes, which might include cost-effectiveness, credibility and, for want of a better term, meaningfulness. The research described in this paper addresses both model accuracy and meaningfulness as conveyed by self-organizing maps (SOMs). SOMs are neural-network based representations of data distributions that provide two-dimensional depictions of multi-dimensional relationships. As such they can enable developers and project managers (and researchers) to visualize often complex interactions among and between software measurement data. We illustrate the effectiveness of SOMs by building on two previous empirical studies. Not only are the maps able to portray graphically the distributions of variables and their interrelationships, they also prove to be effective in terms of classification and prediction accuracy. As a result we believe that they could be a useful supplementary tool for researchers and managers concerned with understanding, modeling and controlling complex software projects.
展开▼
