A Federated Transfer Learning Framework Based on Heterogeneous Domain Adaptation for Students’ Grades Classification

Abstract

In the field of educational data mining, the classification of students’ grades is a subject that receives widespread attention. However, solving this problem based on machine learning algorithms and deep learning algorithms is usually limited by large datasets. The privacy problem of educational data platforms also limits the possibility of building an extensive dataset of students’ information and behavior by gathering small datasets and then carrying out the federated training model. Therefore, the balance of educational data and the inconsistency of feature distribution are the critical problems that need to be solved urgently in educational data mining. Federated learning technology enables multiple participants to continue machine learning and deep learning in protecting data privacy and meeting legal compliance requirements to solve the data island problem. However, these methods are only applicable to the data environment with common characteristics or common samples under the alliance. This results in domain transfer between nodes. Therefore, in this paper, we propose a framework based on federated transfer learning for student classification with privacy protection. This framework introduces the domain adaptation method and extends the domain adaptation to the constraint of federated learning. Through the feature extractor, this method matches the feature distribution of each party in the feature space. Then, labels and domains are classified on each side, the model is trained, and the target model is updated by gradient aggregation. The federated learning framework based on this method can effectively solve the federated transfer learning on heterogeneous datasets. We evaluated the performance of the proposed framework for student classification on the datasets of two courses. We simulated four scenarios according to different situations in reality. Then, the results of only source domain training, only target domain training, and federated migration training are compared. The experimental results show that the heterogeneous federated transfer framework based on domain adaptation can solve federated learning and knowledge transfer problems when there are little data at the data source and can be used for students’ grades classification in small datasets.