FedSLIM: First Federated MDL Framework for Privacy-Preserving Pattern Mining
Cuts search by orders of magnitude while recovering hidden insights across data silos.
Federated learning has excelled at predictive tasks, but descriptive analytics—finding high-level patterns and summaries—remains largely untouched by privacy-preserving methods. Existing federated pattern miners rely on frequency-based rules and lack a principled global objective. A new paper from researchers at Leiden University introduces FedSLIM, a framework that brings Minimum Description Length (MDL) optimization to distributed, privacy-friendly pattern mining.
FedSLIM builds on the SLIM algorithm, which searches for a succinct set of patterns that best compresses the data. It offers two variants tailored to different deployment needs: one prioritizes strict privacy with minimal communication, the other trades some privacy for higher optimization fidelity. On real-world datasets under IID and non-IID splits, both variants preserve centralised-level compression with dramatically less search cost. Crucially, FedSLIM reveals a “local-global discovery gap”—some globally compressive patterns are invisible to isolated local optimization. Its federated design bridges that gap, uncovering patterns absent from all local models. This establishes federated MDL mining as a practical foundation for descriptive analytics across data silos.
- FedSLIM is the first federated MDL-based framework for descriptive pattern mining, extending the SLIM algorithm.
- Two complementary variants balance privacy, communication overhead, and optimization fidelity for different deployment scenarios.
- Experiments show orders-of-magnitude reduction in search cost while recovering globally informative patterns missed by standalone local mining.
Why It Matters
Enables organizations to collaboratively discover data patterns without exposing sensitive raw transactions, unlocking federated descriptive analytics.