Exploring Automated Test Generation with Boundary Value Analysis (BVA) and Equivalence Partitioning (EP) Integration for Machine Learning Testing

Abstract

Machine learning models must be robust and reliable particularly for critical applications like autonomous systems and medical diagnosis. This paper introduces a novel Blackbox testing approach to improve machine learning model testing. It combines automated test generation with Boundary Value Analysis (BVA) and Equivalence Partitioning (EP). Using two well-known datasets the Iris and Titanic datasets we carried out in-depth experiments to train different models such as decision trees, Support vector machines (SVM) and Neural networks. The methodology was centred on the shortcomings of conventional Blackbox testing techniques which often overlook important edge cases and have incomplete coverage. These results were echoed in test coverage and performance metrics such as F1-score, recall and precision. An example of our methodology is that the F1-score for neural networks on Iris dataset improved from 0.89 to 0.95. In contrast, the F1-score for decision trees on Titanic dataset changed from 0.74 to 0.81. Such developments were realized through a successful fusion of BVA and EP, which ensures both automated instruments for comprehensive and effective test case creation along with conducting a complete analysis of input boundaries and partitions. The suggested technique combines Metamorphic testing for machine learning with two Blackbox coverages metrics BVA and EP by generating follow-up dataset from the original datasets using generators created for BVA and EP. As such, it has been confirmed using experiments, that the models we tested using our method are effective than those constructed based on classical approaches. This has been further confirmed using statistical analyses like P-Test and T-tests to establish the significance of these improvements observed. The quality assurance procedure for machine learning models can be improved by developing easy-to-scale Blackbox testing framework that is dependable according to this research work’s outcomes. The results highlight the need for incorporating cutting-edge testing techniques to guarantee machine learning systems dependability in a range of crucial applications.