<p>Online learning has been widely adopted for data stream classification due to its capability for real-time processing. However, in real-world scenarios, data streams often exhibit feature evolution in which features dynamically appear or disappear, leading to a mismatch between the trained model and the current feature space. Moreover, label scarcity further degrades model performance. These challenges pose significant difficulties for traditional online learning methods. To address these challenges, a novel online learning framework for feature-evolvable data streams with label scarcity, termed OL-FELS, is proposed. Within this framework, to accommodate continuously evolving feature spaces, each feature is assumed to follow an independent Gaussian distribution, and classification is performed using a Gaussian Naive Bayes model. Furthermore, a feature-weighting strategy based on Welch’s t-test is introduced to dynamically quantify the contribution of each feature during prediction. To enhance model stability, an adaptive learning rate with a decay mechanism is employed to accommodate evolving data distributions. Additionally, to address label scarcity, a distance-based pseudo-label selection strategy is proposed to identify high-confidence instances for model updates. Experimental results on both synthetic data streams (MOA) and real-world datasets (UCI) demonstrate that OL-FELS consistently achieves superior performance under conditions of feature evolution and label scarcity.</p>