Partial Participation

Edge devices such as smartphones, sensors, and IoT hardware are not perpetually online or available for computation. Their participation is governed by charging cycles, sleep schedules, and user activity. A federated server cannot wait for all devices to be simultaneously available, making partial participation a necessity for continuous training.

The uplink communication from clients to the server is typically the bottleneck, with bandwidth orders of magnitude lower than the server's downlink. Transmitting model updates from millions of devices simultaneously is infeasible. Selecting a manageable subset of clients each round is essential to avoid network congestion and training collapse.

• Primary Constraint: Uplink capacity limits concurrent transmissions.
• System Design: Client selection must be communication-aware.

The federated client population is highly heterogeneous in compute power, memory, and battery life. Devices range from powerful laptops to constrained microcontrollers. Partial participation allows the system to select clients capable of completing a local training epoch within a reasonable time window, preventing stragglers from delaying the entire round. This is closely related to adaptive client selection strategies.

Training machine learning models consumes significant energy. Performing local SGD and transmitting updates incurs a real economic cost for end-users in terms of battery drain and data usage. Federated learning protocols must minimize participant burden. Partial participation distributes this cost across the population over time, making the system viable for real-world deployment where user retention is critical.

Training on a new, randomly selected subset of clients each round helps the global model generalize. However, with non-IID data across devices, repeated selection of the same clients could bias the model. Partial participation, when randomized, acts as a form of statistical regularization. The challenge is mitigating client drift, where local models diverge due to multiple local steps on heterogeneous data—a core problem addressed by algorithms like FedProx and SCAFFOLD.

The central server's ability to manage connections, authenticate devices, and aggregate updates has practical limits. Federated learning orchestrators (e.g., TensorFlow Federated, Flower) must handle scheduling, state management, and fault tolerance for potentially vast fleets. Partial participation keeps the operational load on the server's aggregation infrastructure within scalable bounds, preventing bottlenecks in the coordination plane.

The algorithmic methods for determining which subset of clients participates in a given federated learning round. These strategies are the direct implementation mechanism for partial participation. Key approaches include:

• Random Sampling: Uniform selection, providing statistical fairness.
• Resource-Aware Selection: Choosing clients based on available battery, compute, or bandwidth.
• Data-Driven Selection: Prioritizing clients with data that maximizes learning progress or reduces variance.
• Adaptive Client Selection: Dynamically adjusting selection criteria based on system state and learning objectives.

A communication protocol that decouples client updates from synchronized global rounds. Instead of waiting for all selected clients, the server aggregates updates as they arrive. This paradigm:

• Eliminates stragglers as a bottleneck, naturally accommodating partial participation where clients have highly variable completion times.
• Introduces update staleness, where a client's model may be several global iterations behind when its update is applied.
• Requires staleness-aware aggregation techniques to weight older updates appropriately to maintain convergence stability.

A straggler mitigation technique that introduces redundant computation across clients. By designing local tasks with built-in redundancy, the server can reconstruct the full gradient aggregate from a partial set of completed updates. This technique:

• Tolerates client dropouts or slow devices without requiring recomputation, making partial participation more robust.
• Trades off increased local computation for reduced waiting time per communication round.
• Is particularly valuable in heterogeneous environments where device capabilities vary widely.

A fundamental optimization challenge exacerbated by partial participation and local training on non-IID data. Client drift occurs when local client models diverge from the global objective after multiple steps of local Stochastic Gradient Descent. Key implications:

• Partial participation amplifies drift because the global model is updated by a non-representative subset of clients each round.
• Mitigation algorithms like FedProx (which adds a proximal term to local loss) or SCAFFOLD (which uses control variates) are essential to counteract drift and ensure convergence.
• Managing drift is critical for the final performance of models trained under communication constraints.

A multi-tier communication architecture that introduces intermediate aggregators (e.g., edge servers, cluster heads) between end devices and the central cloud. This structure directly addresses the scalability limits of partial participation by:

• Performing local aggregation at the edge tier, reducing the number of long-haul transmissions to the cloud.
• Enabling more frequent partial participation within a cluster without incurring full WAN communication costs.
• Improving latency and bandwidth utilization, which is crucial for cross-silo federated learning in settings like healthcare or finance.

The fundamental iterative unit of federated learning, consisting of server broadcast, local training, and client update aggregation. Partial participation is defined within this construct.

• The total number of rounds is a primary metric for communication efficiency.
• Partial participation directly controls the number of active uplink transmissions per round, a major cost factor.
• System designers must balance round efficiency (progress per round) with client efficiency (utilization of available devices) to minimize total rounds to convergence.