One criteria for assessing the software quality is ensuring that there is no defect in the software which is being developed. Software defect classification can be used to prevent software defects. More earlier software defects are detected in the software life cycle, it will minimize the software development costs. This study proposes a software defect classification using Fuzzy Association Rule Mining (FARM) based on complexity metrics. However, not all complexity metrics affect on software defect, therefore it requires metrics selection process using Correlation-based Feature Selection (CFS) so it can increase the classification performance. This study will conduct experiments on the NASA MDP open source dataset that is publicly accessible on the PROMISE repository. This datasets contain history log of software defects based on software complexity metric. In NASA MDP dataset the data distribution between defective and not defective modules are not balanced. It is called class imbalanced problem. Class imbalance problem can affect on classification performance. It needs a technique to solve this problem using oversampling method. Synthetic Minority Oversampling Technique (SMOTE) is used in this study as oversampling method. With the advantages possessed by FARM in learning on dataset which has quantitative data attribute and combined with the software complexity metrics selection process using CFS and oversampling using SMOTE, this method is expected has a better performance than the previous methods.