Experiment pipeline

The repo ships scripts that follow a four-stage EXP 1 → EXP 4 flow. Stages build on the best settings from the previous step where applicable. For a fuller account of each partition style (EXP 2) and each aggregation style (EXP 3), read Supported distributed RF patterns.

EXP 1 — Hyperparameters

Runs before simulating federated split data.

• Search: number of trees (odd counts in a configured range, e.g. 1–100), gini vs entropy, SV vs WV.
• Output: One best configuration reused in later experiments.

Script: run_exp1_hparams.py

EXP 2 — Independent client forests

Each client trains with the EXP 1 configuration. Data split style matters:

Sub-exp	Split idea
2.1	Feature-based partitions; test on the client’s slice and on the full test set.
2.2	Uniform random shards of equal size; evaluate on the full test set.
2.3	Same sizes as 2.1 but randomize which samples go to which client.

Script: run_exp2_clients.py

EXP 3 — Global RF via federation

Merge client RFs with RF_S_DTs_A, RF_S_DTs_WA, RF_S_DTs_A_All, or RF_S_DTs_WA_All; compare the global model to independent RFs, best single client, and (if run) a centralized baseline.

Script: run_exp3_federation.py

EXP 4 — Differential privacy

• Clients train DP random forests (example ε values: 0.1, 0.5, 1, 5).
• Each DP model is scored on the full test set; then trees are merged using the best strategy from EXP 3.
• Compare: DP per-client, federated DP global, and non-DP global.

Script: run_exp4_dp_federation.py

---

See Getting started for one-line commands to run each script.