Deep multi-metrics learning for mobile app defect prediction using code and process metrics

초록

The rise in mobile apps necessitates precise defect prediction to aid developers in resource allocation. The efficacy of defect prediction relies on data representation and model selection. However, existing research relies on isolated data sources, limiting models' ability to capture the complex interplay of code and process metrics in app development. This paper addresses this limitation by proposing a Deep Multi-Metrics Learning Model (DMLM), which leverages code metrics from the current code version and process metrics from previous releases. A deep convolutional neural network (CNN) is employed to capture intricate patterns within these metrics, enabling more accurate predictions. Experimental evaluations using nine real-world Android apps from Git Android repositories demonstrate that DMLM outperforms state-of-the-art approaches under non-effort-aware conditions regarding Area Under the Curve (AUC), F1 scores, and Matthews correlation coefficient (MCC). The experimental evaluation also demonstrates that the DMLM model outperforms existing baseline methods in PofB20 under effort-aware conditions. The findings highlight the efficacy of DMLM in enhancing mobile app defect prediction performance across both non-effort-aware and effort-aware contexts.

키워드