Distributed Strategies

Partitioning Strategies

The project now supports a broader range of client-data topologies.

Strategy	Best for	Notes
uniform	balanced client splits	Fast default for smoke tests
stratified	class-balance preservation	Good baseline for fair comparisons
feature	rule-based site segmentation	Useful when client boundaries depend on a feature
sized	capacity planning	Simulate known site volumes
dirichlet	non-IID research	Standard benchmark for heterogeneous federated learning
label_skew	site specialization	Simulate clients that only observe a subset of labels

Aggregation Strategies

Legacy paper-style strategies

Strategy	Behavior
rf_s_dts_a	top-K trees per client by accuracy
rf_s_dts_wa	top-K trees per client by weighted accuracy
rf_s_dts_a_all	global top-K trees by accuracy
rf_s_dts_wa_all	global top-K trees by weighted accuracy

Extended strategies

Strategy	Behavior
top_k_global_balanced_accuracy	global top-K by balanced accuracy
top_k_global_f1	global top-K by macro F1
proportional_weighted_accuracy	allocate tree budget by client data volume, then rank by weighted accuracy
proportional_balanced_accuracy	proportional budget with balanced-accuracy ranking
threshold_weighted_accuracy	include trees above a quality floor, with fallback selection

When To Use Which Strategy

• Use rf_s_dts_wa_all when you want a strong general-purpose baseline.
• Use top_k_global_balanced_accuracy when class imbalance matters.
• Use proportional_weighted_accuracy when larger clients should contribute more trees.
• Use aggregation_strategy="auto" in FederatedRandomForest when you want validation-driven strategy selection.

Example

from distributed_random_forest import FederatedRandomForest

model = FederatedRandomForest(
    n_clients=6,
    partition_strategy="dirichlet",
    partition_kwargs={"alpha": 0.3},
    aggregation_strategy="auto",
    selection_metric="weighted_accuracy",
    execution_backend="thread",
)

Metric Bundle

Validation and test reports include:

• accuracy
• weighted_accuracy
• balanced_accuracy
• f1_score