Adaptive PCA-based feature drift detection using statistical measure

摘要

Abstract The plethora of existing methods in the streaming environment is sensitive to extensive and high-dimensional data. The distribution of these streaming data may change concerning time, known as concept drift. Several drift detectors are built to identify the drift near its occurrence point. Still, they lack proper attention to determine the feature relevance change over time, known as feature drift. Over time, the distribution change of the relevant features subset or the change in the relevant features subset itself may cause feature drift in the data stream. The paper proposes an adaptive principal component analysis based feature drift detection method (PCA-FDD) using the statistical measure to determine the feature drift. The proposed work presents a framework for identifying the most important features subset, feature drift, and incremental adaptation of the prediction model. The proposed method finds the relevant features subset by utilizing the incremental PCA and detects feature drift by observing the change in the percentage similarities among the most important features subset with respect to time. It also helps to forecast the prediction error of the base learning model. The proposed method is compared with state-of-the-art methods using synthetic and real-time datasets. The evaluation results exhibit that the proposed work performs better than the existing compared methods in terms of classification accuracy.