Jiaju Chen, Yuxuan Lu, Xiaojie Wang, Huimin Zeng, Jing Huang, Jiri Gesi, Ying Xu, Bingsheng Yao, Dakuo Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Ziyi Wang, Yuxuan Lu, Wenbo Li, Amirali Amini, Bo Sun, Yakov Bart, Weimin Lyu, Jiri Gesi, Tian Wang, Jing Huang, Yu Su, Upol Ehsan, Malihe Alikhani, Toby Jia-Jun Li, Lydia Chilton, Dakuo Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Ziyi Wang, Yuxuan Lu, Yimeng Zhang, Pei Chen, Ziwei Dong, Jing Huang, Jiri Gesi, Xianfeng Tang, Chen Luo, Qun Liu, Yisi Sang, Hanqing Lu, Manling Li, Jin Lai, Dakuo Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Interpreting and debugging machine learning models is necessary to ensure the robustness of the machine learning models. Explaining mispredictions can help significantly in doing so. While recent works on misprediction explanation have proven promising in generating interpretable explanations for mispredictions, the state-of-the-art techniques “blindly” deduce misprediction explanation rules from all data features, which may not be scalable depending on the number of features. To alleviate this problem, we propose an efficient misprediction explanation technique named Bias Guided Misprediction Diagnoser (BGMD), which leverages two prior knowledge about data: a) data often exhibit highly-skewed feature distributions and b) trained models in many cases perform poorly on subdataset with under-represented features. Next, we propose a technique named MAPS (Mispredicted Area UPweight Sampling). MAPS increases the weights of subdataset during model retraining that belong to the group that is prone to be mispredicted because of containing under-represented features. Thus, MAPS make retrained model pay more attention to the under-represented features. Our empirical study shows that our proposed BGMD outperformed the state-of-the-art misprediction diagnoser and reduces diagnosis time by 92%. Furthermore, MAPS outperformed two state-of-the-art techniques on fixing the machine learning model's performance on mispredicted data without compromising performance on all data. All the research artifacts (i.e., tools, scripts, and data) of this study are available in the accompanying website [1].

Background: Just-In-Time (JIT) defect prediction models predict if a commit will introduce defects in the future. DeepJIT and CC2Vec are two state-of-the-art JIT Deep Learning (DL) techniques. Usually, defect prediction techniques are evaluated, treating all training data equally. However, data is usually imbalanced not only in terms of the overall class label (e.g., defect and non-defect) but also in terms of characteristics such as File Count, Edit Count, Multiline Comments, Inward Dependency Sum etc. Prior research has investigated the impact of class imbalance on prediction technique's performance but not the impact of imbalance of other characteristics. Aims: We aim to explore the impact of different commit related characteristic's imbalance on DL defect prediction. Method: We investigated different characteristic's impact on the overall performance of DeepJIT and CC2Vec. We also propose a Siamese network based few-shot learning framework for JIT defect prediction (SifterJIT) combining Siamese network and DeepJIT. Results: Our results show that DeepJIT and CC2Vec lose out on the performance by around 20% when trained and tested on imbalanced data. However, SifterJIT can outperform state-of-the-art DL techniques with an average of 8.65% AUC score, 11% precision, and 6% F1-score improvement. Conclusions: Our results highlight that dataset imbalanced in terms of commit characteristics can significantly impact prediction performance, and few-shot learning based techniques can help alleviate the situation.

Static analysis tools attempt to detect faults in code without executing it. Understanding the strengths and weaknesses of such tools, and performing direct comparisons of their ef-fectiveness, is difficult, involving either manual examination of differing warnings on real code, or the bias-prone construction of artificial test cases. This paper proposes a novel automated approach to comparing static analysis tools, based on producing mutants of real code, and comparing detection rates over these mutants. In addition to making tool differences quantitatively observable without extensive manual effort, this approach offers a new way to detect and fix omissions in a static analysis tool's set of detectors. We present an extensive comparison of three smart contract static analysis tools, and show how our approach allowed us to add three effective new detectors to the best of these. We also evaluate popular Java and Python static analysis tools and discuss their strengths and weaknesses.