SLeDGe: Semi-Supervised Learning on Data Streams with Graph Structure Learning

Researchers introduce SLeDGe, a semi-supervised learning method designed for streaming data that dynamically learns graph structures to capture evolving relationships between samples. The approach achieves significant accuracy improvements (31.7% relative gain with 0.1% labels) by balancing memory constraints with adaptive graph learning, addressing a key limitation in existing SSL methods that rely on static similarity measures.

SLeDGe addresses a fundamental challenge in machine learning: extracting useful patterns from continuous data streams where labeled examples are scarce. Traditional semi-supervised learning assumes stable data distributions and fixed relationships between samples, assumptions that break down in real-world streaming scenarios where data characteristics shift continuously. This research advances the field by introducing mechanisms to learn graph structures dynamically, allowing the model to update its understanding of sample relationships as new data arrives.

The breakthrough lies in SLeDGe's dual-memory architecture, which maintains separate strategies for labeled and unlabeled data while respecting strict computational and storage constraints typical of streaming environments. By encouraging sparsity in the relational graph, the method eliminates noise and focuses computational effort on meaningful connections between samples. This design choice directly improves label propagation efficiency, the core mechanism through which semi-supervised learning leverages unlabeled data.

For practitioners building machine learning systems on streaming platforms, SLeDGe offers substantial practical value. The 31.7% accuracy improvement under extreme label scarcity (0.1% labeled data) translates to lower annotation costs and faster model deployment in production settings. Industries managing continuous data flows—from IoT sensor networks to financial market analysis—could reduce their labeling burden significantly while maintaining model quality.

The research establishes new benchmarks across 12 datasets, creating a foundation for future work in streaming SSL. Developers should monitor whether these results transfer to production environments and whether the memory constraints prove adequate for high-velocity data streams. The focus on adaptive graph learning may also inspire similar approaches in other domains facing streaming data challenges.

• →SLeDGe learns dynamic graph structures for semi-supervised learning on data streams, improving accuracy by 31.7% with minimal labeled data
• →The method uses dual-memory strategies to balance adaptation and historical consistency under strict computational constraints
• →Graph sparsity encourages the model to filter spurious connections and strengthen label propagation efficiency
• →Performance gains tested across 12 datasets demonstrate broad applicability to streaming scenarios
• →The approach reduces annotation costs for real-world applications managing continuous, high-volume data