AXGFS: A hybrid alpha-investing and XGBoost framework for streaming feature selection

Abstract

In real-time applications like predictive maintenance in industry, patient health monitoring, and video surveillance, a large volume of data is being generated, and new features are becoming available over time. Existing feature selection methods assume that all features are available at the same instance of time and that the complete feature space should be known in advance to perform feature selection. Therefore, they are not suitable for environments where new features arrive dynamically. To overcome this problem, the paper presents a feature stream selection framework that incrementally processes incoming features and identifies a subset of relevant features that has higher predictive power for classification. This is a hybrid framework that combines the strength of statistical feature selection using Alpha-Investing with model-based feature importance ranking using XGBoost to select a relevant and non-redundant feature set. The proposed framework was evaluated using eight benchmark datasets, and classification accuracy was evaluated using three different classifiers. The experimental results show that the proposed hybrid framework increases classification accuracy by up to 5–7% while selecting fewer features than existing feature selection methods. These results indicate that the proposed hybrid approach of integrating Alpha-Investing with XGBoost selects more relevant and non-redundant features, thereby improving the predictive accuracy of the classifier. Thus, the framework can be used in applications where features arrive dynamically, and feature stream selection needs to be performed using the features available up to the current time, such as healthcare monitoring and predictive maintenance.