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

SP-CACW: Convergence-Aware Client Weighting for Selfish Personalized Learning

NeurIPS 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning · 0 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 SP-CACW, a selfish federated learning method where a target client sets peer aggregation weights online by minimizing an upper bound on its convergence error (bias-variance tradeoff), using EMA bias estimates and FTAL, evaluated on MNIST, CIFAR-100 and LEAF Shakespeare.

Problem

Standard federated learning optimizes a global average that underperforms for individual clients; a target client needs to use peer gradients to minimize its own risk while avoiding negative transfer from biased peers.

Approach

Choose per-round aggregation weights over the simplex by minimizing a convergence-derived proxy cost that trades off squared gradient bias (estimated via EMA of gradient differences) against weighted variance; weights are updated online with Follow-The-Approximate-Leader, yielding sparse weights that can zero out harmful peers.

Key results

• SP-CACW reaches 94.85% on MNIST relabel vs 95.31% Oracle and about 92.36% best baseline (Local-only); FedDisco best on CIFAR-100 sigma=10 at 54.03% vs SP-CACW 52.23%.
• On LEAF Shakespeare n=15, SP-CACW achieves 53.64% vs Fed Cluster 53.40%, close margins.
• Theory claims a cluster-setting rate of O(sigma/sqrt(n0 T)) improving over Autarky by sqrt(n0), but the appendix derivation relies on a candidate-weight comparison and a stated FTAL regret term.

Surprising findings

• FedDisco outperforms SP-CACW on CIFAR-100 sigma=10 (54.03 vs 52.23) and on MNIST rotation-related and CIFAR settings, contrary to the claim of consistent top-three placement.
• Local-only training (Only My Data) is the strongest baseline on MNIST rotation (85.24%), beating FedAMP, Ditto and Federated Clustering.

Limitations

• Theory assumes stationary biases and smoothness/bounded-variance; practical experiments use a constant smoothing beta with no guarantee.
• Full communication with all peers assumed in the main formulation; partial participation deferred to appendix.
• Reported gains over baselines are often within one standard deviation; some claims (e.g. 'consistently top three') are not supported by the table.

Experimental setup

Datasets: MNIST CIFAR-100 LEAF Shakespeare

Benchmarks: MNIST label-switching MNIST rotation CIFAR-100 sigma=10 CIFAR-100 sigma=0.01 LEAF Shakespeare n=15 m=7 LEAF Shakespeare n=30 m=3

Metrics: mean test accuracy over five seeds

Baselines: FedAMP Federated Clustering Ditto FedDisco Local-only training Oracle

Builds on: FTAL (online exp-concave optimization) Ditto FedAMP FedDisco Federated Clustering

Closest papers by meaning

• A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.87)
• Decentralized Directed Collaboration for Personalized Federated Learning (CVPR 2024 · similarity 0.87)
• Competitive Advantage Attacks to Decentralized Federated Learning (NeurIPS 2025 · similarity 0.86)
• DePRL: Achieving Linear Convergence Speedup in Personalized Decentralized Learning with Shared Representations (AAAI 2024 · similarity 0.86)
• Interaction-Aware Gaussian Weighting for Clustered Federated Learning (ICML 2025 · similarity 0.86)
• Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning (NeurIPS 2025 · similarity 0.86)