Online learning from capricious data streams via shared and new feature spaces

Data streams refer to data sequences generated at a high rate over a continuous period, such as social media analysis, financial transaction monitoring, and sensor data processing. Most existing data stream mining methods make assumptions about the feature space, assuming it is either fixed or undergoes regular changes, such as trapezoidal or evolving data streams. However, these restrictions do not hold for real-world applications where data streams may exhibit arbitrary missing features. To address the issue of arbitrary missing features in the feature space, we propose the Online Learning from Capricious Data Streams (OLCDS) algorithm and its variant, OLCDS-I. Specifically, OLCDS first identifies the higher uncertainty features that can provide more information for the optimization model. Then, based on the shared and new feature space, we formulate the constrained optimization problem using the soft margin technique. We deduce the update rules and use model sparsity to retain the essential features for classifier learning. Compared to existing online learning approaches, our new method eliminates the need for feature space assumptions and avoids generating missing features. Extensive experiments compared with five state-of-the-art methods on ten real-world datasets demonstrate the effectiveness and efficiency of our new algorithms