Multimodal Learning for Just-In-Time Software Defect Prediction in Autonomous Driving Systems

The growing importance of reliable software in autonomous driving technologies underscores the need for effective defect prediction. This paper introduces a multimodal learning approach for just-in-time software defect prediction (JIT-SDP) using transformers to integrate diverse data modalities such as code features, metrics, and context. By employing attention mechanisms, the model combines outputs from text and tabular data via fully connected layers to predict defects. Experiments on datasets from Apollo, Carla, and DonkeyCar demonstrate that the approach outperforms state-of-the-art models like CodeBERT and GraphCodeBERT, with accuracy improvements of 7.55% to 17.31%, showcasing its potential to enhance software reliability.