Client Selection and Sampling
What you'll learn: How federated learning systems strategically choose which clients participate in each training round to balance efficiency, fairness, and model performance.
The Challenge of Scale
In federated learning, you rarely train on all clients simultaneously. Imagine orchestrating 10 million smartphones to participate every round—network congestion, device availability, and computational costs make this infeasible. Instead, the server selects a subset of clients each round, similar to polling a representative sample rather than surveying an entire population.
Selection Strategies
Random Sampling
The simplest approach: randomly select a fraction of clients (e.g., 10%) each round. This works when clients are relatively homogeneous and equally available.
Weighted Selection
Clients can be sampled with probabilities proportional to their dataset size, ensuring larger data contributors participate more often. This accelerates convergence but may bias toward certain data distributions.
Availability-Aware Selection
Real-world devices aren't always online or idle. Systems filter for clients that are charging, on WiFi, and idle before random selection—prioritizing those most likely to complete training without interruption.
Strategic Sampling
Advanced approaches intentionally sample to reduce bias: over-sample underrepresented populations, rotate through all clients systematically (cyclic sampling), or select clients with high gradient diversity to improve model generalization.
Why It Matters
Poor sampling strategies can lead to biased models (if certain populations are consistently excluded), slow convergence (if selected clients have redundant data), or fairness issues (if some devices never contribute). Client selection is where federated learning's privacy promises meet practical constraints.
Key Takeaway: Client selection determines which devices train your model each round—balancing statistical efficiency, system constraints, and fairness requires thoughtful sampling strategies beyond simple randomization.