Semi-Supervised Acoustic Scene Classification under Domain Shift with MixMatch and Information Bottleneck Optimization

Domain shift is a prevalent issue in acoustic scene classification (ASC) tasks, leading to reduced classification accuracy in the target domain. This paper tackles the domain shift problem in ASC caused by differing recording environments across cities for the ICME2024 Grand Challenge. The proposed approach employs a ResNet backbone trained using the MixMatch semi-supervised learning framework. To address domain shift problem, data augmentation techniques and a task-specific 3-class classification loss are integrated. Furthermore, domain generalization is enhanced by optimizing the information bottleneck problem and maximizing feature entropy through the addition of two loss functions: the class-wise instance discrimination (CID) loss and the feature dimension correlation (FDC) loss. Experimental result shows that the proposed method achieves a final score of 0.699 on the ICME2024 Grand Challenge private test set and secures the 4th rank, demonstrating its effectiveness in mitigating domain shift in ASC tasks.