Application of K-Medoids with Kd-Tree for Software Fault Prediction

Software fault prediction area is subject to problems like non availability of fault data which makes the application of supervised techniques difficult. In such cases unsupervised approaches like clustering are helpful. In this paper, K-Medoids clustering approach has been applied for software fault prediction. To overcome the inherent computational complexity of KMedoids algorithm a data structure called Kd-Tree has been used to identify data agents in the datasets. Partitioning Around Medoids is applied on these data agents and this results in a set of medoids. All the remaining data points are assigned to the nearest medoids thus obtained to get the final clusters. Software fault prediction error analysis results show that our approach outperforms all unsupervised approaches in the case of one given real dataset and gives best values for the evaluation parameters. For other real datasets, our results are comparable to other techniques. Performance evaluation of our technique with other techniques has been done. Results show that our technique reduces the total number of distance calculations drastically since the number of data agents is much less than the number of data points.