AI Paper Atlas

Research map / Federated learning and privacy / Personalized and heterogeneous FL

Controlled Collaboration Geometry for Personalized Federated Learning

Hongbo Yin, Jichun Wu, Zhou Yang, Chi Jiang, Yin Zhang, Yan Zhang

ICML 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Personalization with global-local trade-off

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace similarity-feedback dynamic weight updates with a fixed feedback-free affinity template (ablation)pFedCCG · CIFAR-10 (beta=0.2)personalized accuracy (%): 82.64 -> 85.20Improved
add the feedback-free template over SF (ablation)pFedCCG · Yahoo (beta=0.2)personalized accuracy (%): 77.81 -> 79.76Improved
add collaboration geometry scheduling (strong early mixing, weaker later) (ablation)pFedCCG · CIFAR-10 (beta=0.2)personalized accuracy (%): 84.72 -> 85.20Improved
add geometry scheduling (ablation)pFedCCG · Aya LLM-SFT (k=3)perplexity: 6.39 -> 6.35Improved
prescribe stationary distribution q equal to dataset proportions in the projection (ablation)pFedCCG · CIFAR-10 (beta=0.2)personalized accuracy (%): 85.01 -> 85.20Improved
add the q-stationary projection (ablation)pFedCCG · Yahoo (beta=0.2)personalized accuracy (%): 79.27 -> 79.76Improved
build affinity template from cosine similarity of initial model updates instead of JS divergence of data distributions (ablation)pFedCCG · CIFAR-10 (beta=0.2) and MovieLensaccuracy / Recall@20: CIFAR-10 85.20 -> 85.22; RecSys 88.81 -> 88.56No clear effect
use weaker output-based proxy templates (softmax on 256 public samples) (ablation)pFedCCG · CIFAR-10 (beta=1)mean personalized accuracy (%): 78.35 -> 77.49 (softmax-cos); best baseline 76.27Hurt
use any proxy template instead of the JS template (ablation)pFedCCG · CIFAR-10 (beta=0.2, 0.5, 1)mean personalized accuracy (%): all proxies still above best baseline, e.g. 84.59 vs 83.97 at beta=0.2Mixed
use pFedCCG projection instead of adaptive graph learning for collaboration weightscollaboration overhead · 50 clientsextra seconds: pFedGraph 281.60 -> pFedCCG 0.04Improved
pFedCCG vs best baseline on Aya LoRA SFTpersonalized LLM fine-tuning · Aya (k=3) Acc@5Acc@5 (%): 80.04 -> 80.12No clear effect
use clustering-style PFL (CFL, pFedCS) versus FedAvg-FTpersonalized FL baselines · CIFAR-10 (beta=1)personalized accuracy (%): FedAvg-FT 76.20 vs CFL 68.07, pFedCS 72.03Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy