In this paper, the authors predict software failure using their Markov Bayesian network model (MBN) when the parameters in the related distributions are not available. This paper is an extension of another paper [1].
The authors talk extensively about various research works, from the classic probability models to the latest neural network models, conducted in the areas of software failure prediction and software reliability. They stress the various restrictive assumptions made in these research models as their major limitations. Since Bayesian networks have been increasingly applied in many fields, the authors make an attempt to apply them to software failure prediction.
Section 4 (“Software failure prediction”) is well written, with neat equations and explanations. The authors, in their two distribution assumptions, have made use of proven and existing theories, such as Gibbs sampling theory and the expectation-maximization algorithm.
The concept flow in the paper is generally good. However, certain sections are not correctly sized. Targeting beginners, a detailed explanation of Bayesian network concepts is provided, with an example (section 2); the case study description (section 5), however, which is supposed to be the core of the paper, is short and a bit confusing.
Even though the authors criticize other research models for making restrictive assumptions, they themselves are forced to assume a few parameters to follow the proposed MBN model. Also, the authors admit that their model is complex, and claim that the time consumed by their algorithm is manageable. I consider this paper to be a good reference for researchers working in the software failure prediction and software reliability areas.