A Clustered Federated Learning Approach for Estimating the Quality of Experience of Web Users

Abstract

In this paper, we present FedCLWAvg, a novel clustered Federated Learning (FL) approach for estimating the Quality of Experience (QoE) of Web users. FedCLWAvg performs clustering on the weights of the trained local models (CLW stands for clustered weights) to identify users with similar data distribution. In the context of QoE modelling, the hypothesis is that personal differences (in terms of perceived QoE for the same stimuli) between groups of users are reflected in different weights of the trained local models. Then, each identified cluster learns its own model using the FedAvg algorithm. To validate our approach, we used the Web QoE dataset including the subjective quality of 3,400 Web browsing sessions identified by the measurement of 9 Web session features. Experimental results have shown that FedCLWAvg achieved greater QoE estimation performance than the classical FedAvg algorithm in terms of mean accuracy and recall, F1-score, and precision computed for the single quality scores.