Personalized Interpretation on Federated Learning: A Virtual Concepts approach

TL;DR

This work tackles non-IID data in federated learning by making client-specific personalization explicit through Virtual Concepts (VCs). It introduces FedVC, a client-supervised FL framework where each client's preference is represented as a mixture of VC vectors learned jointly with the global model, using a projection head to gauge per-sample relevance. VC distributions are modeled as a Gaussian Mixture Model and can be learned via EM or end-to-end optimization, with mechanisms like moving averages and stop-gradient to integrate updates. Empirical results on MNIST and Digit-5 demonstrate improved robustness to distribution shifts and the ability to deploy a single global model without post-training adaptation, while providing interpretable insights into client properties.

Abstract

Tackling non-IID data is an open challenge in federated learning research. Existing FL methods, including robust FL and personalized FL, are designed to improve model performance without consideration of interpreting non-IID across clients. This paper aims to design a novel FL method to robust and interpret the non-IID data across clients. Specifically, we interpret each client's dataset as a mixture of conceptual vectors that each one represents an interpretable concept to end-users. These conceptual vectors could be pre-defined or refined in a human-in-the-loop process or be learnt via the optimization procedure of the federated learning system. In addition to the interpretability, the clarity of client-specific personalization could also be applied to enhance the robustness of the training process on FL system. The effectiveness of the proposed method have been validated on benchmark datasets.

Figures (10)

• Figure 1: Illustration to FedVC. (a) data distribution in an FL system; (b) virtual concepts (pentagon, plus and triangle) are vectors indicating underlying cluster structures of data, e.g., cluster centres; (c) a client's preference (star) is represented by a combination of virtual concepts; (d) client-supervised loss requires sample representations on the same client (data points within the circle) to be close to each other as they share the identical client preference.
• Figure 2: Projection head
• Figure 3: FedVC architecture.
• Figure 4: Clients in the target shift setting. Each bar denotes a client. Each colour indicates one type of distribution. Samples on each client are split into a training set and a test set.
• Figure 5: Class distributions on clients. Each bar denotes the class distribution on a client. Each colour corresponds to a class and the length indicates its proportion on the client.