Coping with Feature Increment and Label Scarcity: A Two-Stage Adaptive Learning Framework
Ruichao Wang , Yucong Dai , Chenping Hou
Feature increment learning is essential for adapting models to dynamic data streams where the feature space dynamically expands over time. While existing methods address feature evolution, they typically rely on the idealized assumption of complete supervision. In high-stakes domains like medical diagnosis, introducing complex heterogeneous features requires expert interpretation, inevitably resulting in severe label scarcity. Consequently, the dual challenge of learning under simultaneous feature increment and label scarcity remains largely unexplored, constituting a significant bottleneck in open-environment machine learning. To bridge this crucial research gap, we propose an Adaptive Learning Framework (ALFL). To address the heterogeneous gap caused by feature increment, we derive a cross-space mapping for heterogeneous knowledge transfer to directly adapt the prior model to the current stage, functioning as a bridge for knowledge preservation. Furthermore, to overcome severe label scarcity, we introduce a manifold graph regularizer to exploit the intrinsic geometry of abundant unlabeled data. Rather than addressing this coupled challenge through isolated paradigms, our framework establishes a structural synergy between the prior model and the graph regularizer under a unified learning objective. Crucially, we provide rigorous theoretical guarantees for our framework. By leveraging Rademacher complexity analysis, we establish a strict generalization error bound and prove an convergence rate for the proposed algorithms. Extensive experimental results validate the effectiveness of our method in handling feature increment and label scarcity.
Adaptive Learning / Dynamic Feature / Label Scarcity / Generalization Bound
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(7513KB)
