FlatLand: Personalized Graph Federated Learning via Tailored Lorentz Space

AI summary

Overview

• Research area: Federated learning (FL), graph neural networks (GNNs), hyperbolic geometry, and personalized federated learning (PFL).
• Technical level: Advanced. The paper builds on the Lorentz (hyperboloid) model of hyperbolic space, Forman–Ricci curvature, bi-Lipschitz embedding theory, and mutual-information arguments.
• Scope: The paper proposes FlatLand, a personalized graph federated learning framework that assigns each client its own learnable Lorentz space and decouples time-like (personalized) from space-like (shared) parameters so that only the shared parameters are aggregated.

What This Paper Is About

Graph federated learning struggles because different clients hold structurally different graphs, so one shared model rarely fits all. Existing personalized federated learning methods try to fix this inside ordinary flat (Euclidean) geometry, using clustering, similarity estimation, or extra modules. FlatLand instead measures the geometry of each client's graph and gives every client its own tailored hyperbolic Lorentz space, then separates what is client-specific from what is common so that aggregation can happen directly.

Key Contributions

1. A geometric diagnosis of graph FL heterogeneity. Using Forman–Ricci curvature, the authors show empirically that client graphs predominantly exhibit negative curvature (hyperbolic structure) and that curvature values vary substantially across clients, indicating geometric heterogeneity beyond ordinary statistical differences.
2. Theory for tailored curvature and time-like heterogeneity encoding. Theorem 4.1 shows that if client graphs have different average Ricci curvatures, no single Lorentz scale parameter K can yield simultaneously small distortion for all clients; Theorem 4.3 shows the time-like coordinate x_t retains mutual information with the client identity C, while the space-like part carries no additional client-specific information.
3. The FlatLand framework. Each client receives a learnable Lorentz scale parameter K_c (initialized from average Forman–Ricci curvature via sigmoid reparameterization and updated during local training) plus a parameter decoupling strategy that splits each Lorentz linear layer's matrix into personalized time-like parameters m^(l) and shared space-like parameters M^(l).
4. Theory-backed aggregation. Proposition 6.1 proves that after server-side averaging of the shared space-like parameters, transformations still map into the same client-specific Lorentz space L_{K_c}^m; the convergence rate remains O(1/T), matching FedAvg, and the added operations are O(1) (exponential map and curvature estimation).

Main Findings

• Curvature is negative and client-dependent. Averaged Forman–Ricci curvature across Cora, ogbn-arxiv, and Amazon-Photo shows pronounced non-Euclidean characteristics, and per-client curvature values vary considerably across clients.
• Local Lorentz training beats local Euclidean training in several settings. For example, on CiteSeer with 10 clients, Local (L) reaches 72.30 ± 0.04 versus Local (E) at 67.82 ± 0.13; the gap is described as particularly pronounced on heterophilic graphs.
• FlatLand leads on several node classification benchmarks. Best results include CiteSeer with 10 clients (73.90 ± 0.23 versus the next best FED-PUB at 72.35 ± 0.53), CiteSeer with 20 clients (72.24 ± 0.24), ogbn-arxiv with 10 clients (67.52 ± 0.16 versus FED-PUB at 66.58 ± 0.08), ogbn-arxiv with 20 clients (67.64 ± 0.04 versus FED-PUB at 66.64 ± 0.12), Roman-empire with 10 clients (66.10 ± 0.21 versus AdaFGL at 64.55 ± 0.00), Minesweeper with 10 clients (76.34 ± 0.05) and 20 clients (74.72 ± 0.11), Tolokers with 20 clients (72.11 ± 0.12), and Questions with 10 clients (67.71 ± 0.08) and 20 clients (66.25 ± 0.10).
• FlatLand is competitive rather than dominant everywhere. On Cora with 10 clients FED-PUB is statistically significant at 81.54 ± 0.12 (FlatLand 80.46 ± 0.28), and on Photo both FED-PUB (92.73 ± 0.18, 91.92 ± 0.12) and FedGTA (93.50 ± 0.21, 92.61 ± 0.15) are stronger than FlatLand (92.49 ± 0.19, 91.06 ± 0.15).
• Standard FL baselines degrade under heterogeneity. FedAvg, FedProx, FedGNN, and FedPer frequently trail local training; the text notes the paper targets non-overlapping node partitions that many prior methods were not designed for.
• FedHGCN underperforms and hits memory limits. Although FedHGCN operates in hyperbolic space, it omits explicit client-geometry modeling and reports OOM (out-of-memory) on ogbn-arxiv for both 10 and 20 clients.
• Graph classification improves modestly. On CHEM (7 datasets) FlatLand reaches 76.55 ± 2.28 and on BIO-CHEM-SN (13 datasets) 67.31 ± 2.58, both statistically significant at p < 0.05; FedHGCN reports OOM on BIO-CHEM-SN.
• Low-dimensional behavior is highlighted. The paper states FlatLand performs well in low-dimensional settings, shown in a dimension study on CiteSeer with 20 clients (Figure 4). The specific accuracy values for that sweep are not included in the provided text.
• Ablation results exist but are not quantified in the provided text. The ablation study of FlatLand on the Cora

Authors’ abstract

Federated learning enables privacy-preserving collaborative training, but highly heterogeneous client data remain challenging, especially in graph federated learning where clients possess structurally diverse graphs. Existing personalized federated learning (PFL) methods ignore the intrinsic geometric properties of diverse graph structures. We propose FlatLand, a novel personalized federated learning method that embeds different clients' data in tailored Lorentz space of hyperbolic geometry. Our key insight is that hyperbolic geometry naturally accommodates the intrinsic negative curvature prevalent in real-world graphs, while the time-like dimension in Lorentz space provides a principled way to encode client-specific heterogeneity. We develop a parameter decoupling strategy that separates heterogeneous information (captured in time-like parameters) from common knowledge (preserved in space-like parameters), enabling direct aggregation without requiring client similarity estimation and extra calculation modules. Empirical results on diverse federated graph learning tasks demonstrate that FlatLand achieves superior performance, particularly in low-dimensional settings.