Comparison the Performances for Distributed Machine Learning: Evidence from XGboost and DNN

Abstract. As a matter of fact, distributed machine learning approaches are widely adopted to enhance the training speed. The purpose of this study is to compare the performance of distributed machine learning models, specifically XGBoost and Deep Neural Networks (DNNs), using a telecommunications customer dataset. The dataset consists of 320,000 samples and 25 features and the analysis focuses on model performance under different data heterogeneity conditions. According to the analysis, XGBoost achieves excellent performance, rapidly reaching a high AUC of 97.7% with a minimal number of iterations, proving its effectiveness on structured, small and medium-sized datasets. In contrast, DNN struggled with this dataset and failed to outperform XGBoost due to its low dimensionality and small size. The paper also discusses the main limitations, including the lack of model diversity, the low dimensionality of the dataset, and the problem of model interpretability, especially for DNN. The results suggest that XGBoost is more suited to small and structured datasets, whereas DNN excels at high dimensionality and complex datasets. Future research should focus on improving model diversity, tuning, and addressing interpretability challenges.