Lecture topic: Federated Learning - Data Imbalance Handling Techniques

speaker: Prof. Ligang He

Date: December 7th, 2021 (Tuesday) 4:00 PM

Lecture format and location: Online video, Tencent Meeting ID: 662 608 005

Tencent Meeting Link https://meeting.tencent.com/dm/plYdBJgANcRg

Welcome all teachers and students to actively participate!

Lecture Introduction:

Title: Federated Learning – Part II Dealing with Data Imbalance

Abstract: Machine learning typically assume that the samples in the training dataset are well balanced in terms of the embedded features and sample values. However, it may not be the case in realistic scenarios. Especially in the scenario of federated learning, the clients may collect data in different locations and at different times, the data may well be imbalanced. In this talk, we will first discuss the impact of data imbalance on learning and then introduce the existing approaches on dealing with data imbalance. Next, we will present our work in dealing with the data imbalance for regression tasks. In our work, we propose a method to measure the uniqueness in terms of the samples’ feature space and the level of abnormality in terms of the samples’ values. We then associate the samples’ uniqueness and abnormality with the samples’ learning value and integrate them into the loss function. We will also present the evaluation results in the talk.