FedAvg for HAR - Exploring the Tradeoff Between Personalized and Generalization Accuracy - Abstract

Sep 21, 2026

Uploaded by Liam

Federated learning (FL) enables distributed training on clients’ local devices, sharing model parameters with a centralized server to improve data usage and address privacy concerns. A key challenge remains balancing personalization with generalization accuracy across heterogeneous clients and imbalanced data. This paper designs testing scenarios comparing centralized, local, and federated training, including a client-data change emulation, and evaluates FedAvg on Human Activity Recognition (HAR).

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Abstract

Introduction

Methodology

FedAvg for HAR: Exploring the Tradeoff Between Personalized and Generalization Accuracy Andrea De Luna DiSPeA University of Urbino [a.deluna1@campus.uniurb.it](a.deluna1@campus.uniurb.it) Susanna Peretti DiSPeA University of Urbino [s.peretti@campus.uniurb.it](s.peretti@campus.uniurb.it) Chiara Contoli DiSPeA University of Urbino [chiara.contoli@uniurb.it](chiara.contoli@uniurb.it) Alessandro Bogliolo DiSPeA University of Urbino [alessandro.bogliolo@uniurb.it](alessandro.bogliolo@uniurb.it) arXiv :2607 .03334v 1 [ cs .LG] 3 Jul 2026 Abstract—The federated learning (FL) paradigm fosters distributed pervasive computing combined with artificial intelligence techniques, allowing for optimized data usage and improved mitigation of privacy concerns. Indeed, model training occurs on the client’s local devices, and model parameters are subsequently shared with a centralized server. However, there is a need to find a tradeoff between models’ personalization and generalization capabilities. In this paper, we design and implement several testing scenarios devoted to evaluating and comparing the centralized, local, and federated paradigm performances. We also design and implement a scenario that emulates a change in clients’ data. We then present experimental results of the FedAvg algorithm applied to the Human Activity Recognition (HAR) domain to understand the trade-off between personalized and generalized accuracy. Results show that, although FedAvg confirms a higher degree of personalization capabilities while keeping a high degree of generalization with respect to the traditional centralized learning, this result is not so obvious under stressful conditions, such as when varying class distribution over clients. Index Terms—Federated Learning, Human Activity Recognition, FedAvg, Performance Analysis, Personalized and Generalization Accuracy I. Introduction Federated learning (FL) gained momentum for its ability to enhance data privacy and promote the use of devices in a distributed, pervasive computing context. Indeed, in such a scenario, clients participating in the federation collect data from their local devices, train a model locally, and share only model parameters with a global (centralized) server. The server aggregates the model’s parameters according to a specific aggregation algorithm strategy and subsequently shares parameter updates with all the clients until a (global) convergence is met. There is a need to find a tradeoff between models’ personalization and generalization capabilities. Such a tradeoff is challenged by multiple factors, such as, but not limited to, device heterogeneity and imbalanced data distribution. In this paper, we delve into FedAvg federated learning mechanisms applied to the Human Activity Recognition (HAR) domain to understand the ©2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses. This manuscript has been accepted for publication and oral presentation at the 2026 11th IEEE International Conference on Machine Learning Technologies (ICMLT 2026) . The final version will appear in IEEE Xplore. trade-off between personalized and generalized accuracy. The idea is to examine the extent to which clients benefit from the FL engagement procedure compared to a nonfederated scenario. Ek et al. were the pioneers in exploring the performance of federated learning model aggregationsin the human activity recognition domain, where they investigated and compared FedAvg, FedPer, and FedMa [1] . This work was subsequently extended in [2], [3], where the authors proposed a new aggregation algorithm and compared their performance against other approaches. Many other works propose new aggregation algorithmsand compare performance [4]–[6], some also explore how to balance accuracy and eﬀiciency with privacy [7], othershow to handle nonindependent and identically distributed (Non-IID) data [8] . Our work is inspired by Ek et al.

FAQs

What is the main goal of using FedAvg for Human Activity Recognition (HAR) in this paper?

The paper studies how FedAvg balances clients’ personalization benefits with the ability to generalize accuracy in the HAR domain, especially when client conditions change.

Which training paradigms are compared, and what extra testing scenario is added?

The authors compare centralized, local, and federated paradigms, and they also implement a scenario that emulates a change in clients’ data to test robustness.

What experimental finding affects the personalization-vs-generalization tradeoff under stressful conditions?

FedAvg shows stronger personalization while maintaining high generalization compared with centralized learning, but this advantage becomes less clear when class distributions vary across clients.

Sep 21, 2026

Uploaded by Liam