Research map / Federated learning and privacy / Communication-efficient FL

Effective and Efficient Federated Tree Learning on Hybrid Data

ICLR 2024 · Federated learning and privacy › Communication-efficient FL · approach: Efficient local training in FL · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HybridTree, a layer-level federated GBDT method for hybrid (feature- and sample-partitioned) data that trains the host's upper tree layers and appends guest-held lower layers, cutting communication and computation versus node-level VFL.

Problem

Federated learning of tree models (GBDT) when parties hold data that differ in both feature and sample spaces (hybrid data), where existing node-level federated GBDT methods are designed only for horizontal or vertical settings and incur high communication and cryptographic overhead.

Approach

Observes that guest-feature split rules often form "meta-rules" (prediction is determined once the rule holds), proves that reordering guest splits to the last tree layers preserves expected predictions, and therefore lets the host train upper layers and guests append lower layers with only two communication rounds per tree.

Key results

• HybridTree reaches AUPRC 0.689 on AD and 0.553 on DEV-AD, versus 0.703 and 0.574 for centralized ALL-IN, and well above SOLO, FedTree, Pivot and TFL.
• Training communication is reduced by 5.4x to 6.3x and training time speedup over FedTree ranges from 4.3x to 8.0x across the four datasets.
• In the multi-host setting HybridTree scores 0.911 on Cod-rna versus 0.931 for ALL-IN and 0.784-0.804 for FedTree.

Measured effects reported in the paper

Surprising findings

• TFL (tree-level aggregation) is sometimes stronger than HybridTree on Cod-rna (0.884 vs 0.927 for HybridTree, i.e. HybridTree is higher, but in the sensitivity table TFL reads 0.773 vs 0.689 at depth 8, suggesting table inconsistencies).

Limitations

• Designed for tabular data; meta-rule knowledge is hard to capture for image, text, and graph data.
• Assumes honest-but-curious parties that do not collude; inference attacks are not addressed.
• Relies on the existence of meta-rules, which is shown empirically on a few datasets rather than guaranteed.

Experimental setup

Datasets: PETs Prize Challenge AD dataset PETs Prize Challenge DEV-AD dataset Adult Cod-rna

Benchmarks: PETs Prize AD PETs Prize DEV-AD Adult Cod-rna

Metrics: AUPRC classification accuracy communication size (GB) training time (s) inference time (s)

Baselines: ALL-IN SOLO FedTree SecureBoost Pivot TFL

Builds on: GBDT SecureBoost FedTree Pivot Diffie-Hellman key exchange

Closest papers by meaning

• Principled Federated Random Forests for Heterogeneous Data (NeurIPS 2026 · similarity 0.87)
• Federated Learning under Periodic Client Participation and Heterogeneous Data: A New Communication-Efficient Algorithm and Analysis (NeurIPS 2024 · similarity 0.86)
• A Primal-Dual Algorithm for Hybrid Federated Learning (AAAI 2024 · similarity 0.85)
• Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.85)
• Fake It Till Make It: Federated Learning with Consensus-Oriented Generation (ICLR 2024 · similarity 0.84)
• Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators (AAAI 2024 · similarity 0.83)