Data leakage detection in machine learning code: transfer learning, active learning, or low-shot prompting?

With the increasing reliance on machine learning (ML) across diverse disciplines, ML code has been subject to a number of issues that impact its quality, such as lack of documentation, algorithmic biases, overfitting, lack of reproducibility, inadequate data preprocessing, and potential for data leakage, all of which can significantly affect the performance and reliability of ML models. Data leakage can affect the quality of ML models where sensitive information from the test set inadvertently influences the training process, leading to inflated performance metrics that do not generalize well to new, unseen data. Data leakage can occur at either the dataset-level (i.e., during dataset construction) or at the code-level. Existing studies introduced methods to detect code-level data leakage using manual and code analysis approaches. However, automated tools with advanced ML techniques are increasingly recognized as essential for efficiently identifying quality issues in large and complex codebases, enhancing the overall effectiveness of code review processes. In this article, we aim to explore ML-based approaches for limited annotated datasets to detect code-level data leakage in ML code. We proposed three approaches, namely, transfer learning, active learning, and low-shot prompting. Additionally, we introduced an automated approached to handle the imbalance issues of code data. Our results show that active learning outperformed the other approaches with an F-2 score of 0.72 and reduced the number of needed annotated samples from 1,523 to 698. We conclude that existing ML-based approaches can effectively mitigate the challenges associated with limited data availability.