Application of defuzzification-free hierarchical fuzzy inference rule generation method to software fault prediction problems and fuzzy rule discussion

Fuzzy inference systems, a sub-branch of artificial intelligence, are decision support systems based on fuzzy set theory. The knowledge is transmitted to these systems through linguistic fuzzy rules. In general, the rule base is determined by expert opinion. However, the rules are generated by automatic rule generation methods when the problem domain becomes complex and expert knowledge is insufficient. In the studies conducted so far, the dataset used to produce the rules with automatic rule generation methods has not been analyzed in detail. The dataset may mislead the model and the ruleset obtained from the project may not always be the most accurate one. The most accurate and generalizable ruleset may be produced from another project. Thus, the rules can be transferred to another application for the same problem domain. The goal of this study is to solve new problems more effectively and faster way by using the ruleset of a different project without having to re-generate the rules when the dataset is changed. Therefore, unlike the data-driven approaches, the knowledge is transferred from the source project to the target project to make a predictive model in the current project. We investigate the portability of the rules by generating them from five different projects for Software Fault Prediction problem. The rulesets are generated by three automatic rule generation methods, namely "Wang-Mendel", "Interval-Valued Fuzzy Reasoning Method with Tuning and Rule Selection" and rule production approach of "Defuzzification-free Hierarchical Fuzzy System". In addition, automatic rule generation methods were compared with Artificial Neural Networks, which is a data-driven machine learning method to compare the success of the model. The results of the experiments indicate that the rules generated from more consistent datasets instead of own dataset of a project significantly improves the performance of existing fuzzy inference systems.