Comparision of Models Built Using AutoML and Data Fusion

Automated machine learning (AutoML) has made life easier for data analysts or scientists by providing quick insights into data by building machine learning (ML) models. AutoML techniques are applied to vast areas from image processing, speech recognition, natural language processing reinforcement learning, and more. However, there is still room for many improvements. AutoML techniques focus only on problems related to predictive modeling, and most of them are designed to work with structured data. AutoML techniques are also time-consuming as they require time to select the appropriate ML pipeline. This paper presents an alternative time-efficient approach for mixed data (both categorical and numerical features obtained from UCI and Kaggle repository) using a data fusion process, which provides high macro average accuracy in less time as compared to AutoML. The AutoML tool considered here is autoscikit-learn (auto-sklearn). This specific library is built in Python using scikit-learn. The implementation of data fusion is also done in Python using scikit-learn. We conclude from the experimental analysis that the pipeline constructed provides better results than the auto-sklearn. This obtained conclusion is supported by a statistical test (Wilcoxon signed ranks test) based on macro average accuracy obtained for both approaches.