We all know that having several experts give their opinion is better than having just one give it. This paper proves it. The focus is on Bayesian classifiers and the method for developing a prediction algorithm.

Often, the paper is hard to comprehend. Consider, for example, this passage from section 2: “By reusing existing models, we reuse the common domain knowledge represented by versatile contexts, and by guiding the adaptation via the context of specific data, we consider the specific knowledge represented by D(c).” This seems to mean that we can use existing estimation models such as COCOMO and then weight the results based on our confidence in them, as compared to other estimation models. Their algorithm in section 2 leaves out the important assumption of conditional independence. This is mentioned in section 3 after we struggle to understand the algorithm. The authors do not discuss the implications of this assumption except to say that without it the math is very hard. You bet it is.

The punch line is to use many experts, and always more than one, for every domain, and weight your confidence in each. Then, iterate your predictions to convergence. This is quite similar to Barry Boehm’s famous wideband Delphi estimation method [1].

The math in section 4 seems correct, but I did not and could not follow it completely. The 19 metrics listed in table 3 as attributes in their experiments are insightful and useful. Table 4 can give the reader leads to tools to be used as expert predictors. I wish the authors would have defined stress. They define stability as “changes to source code from release to release,” which contrasts with the definition I prefer: a measure of performance.

My advice is to use the paper as a reference if and when you need to make profound predictions about a software project. Skip the math and employ the algorithm.