Meta-feature-based data preprocessing for machine learning through correlation structures and cost-aware benchmarking: A case study in cybersecurity

One of the current challenges in applying machine learning is optimizing the constructed models to achieve the best possible performance. This article proposes a novel approach to determine which data preprocessing strategy may be most beneficial for enhancing the evaluation fit metrics, based on studying the correlation between the dataset's meta-features and the performance response variables. Additionally, the meta-features were categorized in terms of modification cost and control over the model's fit to determine which strategies appeared to be the most optimal. Also, we studied if these transformation can improve the results obtained by several training and testing proportions. The study was conducted with 42 different configuration datasets derived from the multiclass classification problem of IP address maliciousness. Three machine learning algorithms and tools were evaluated: Autosklearn, Gaussian Mixture Models, and Extreme Gradient Boosting; all of them have been studied on the same problem in previous works. Also, we applied five different types of data transformations, such as scaling, dimensionality reduction techniques, and quantile transformation. The results show that meta-feature correlation analysis significantly improves machine learning performance by guiding data preprocessing and transformation strategies, sometimes even surpassing the best original performance.