COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation

Authors: Wenxu Wang, Rui Zhou, Jing Wang, Yun Zhou, Cheng Zhu, Ruichun Tang, Bo Han, Nevin L. Zhang

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our approach achieves state-of-the-art performance. Ablation studies and experiments on synthetic datasets confirm the effectiveness of each proposed module. 4. Experiments.
Researcher Affiliation Academia 1College of Computer Science and Technology, Ocean University of China, Qingdao, China 2National Key Laboratory of Information Systems Engineering, National University of Defense Technology, Changsha, China 3Department of Mechanical Engineering, University of British Columbia, Vancouver, Canada 4Department of Computer Science, Hong Kong Baptist University, Hong Kong, China 5Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong, China. Correspondence to: Ruichun Tang <EMAIL>.
Pseudocode Yes Algorithm 1 Overall Training Process of COSDA
Open Source Code Yes The code is available at https://github.com/ZHOURui6025/COSDA-master.
Open Datasets Yes Extensive experiments are conducted on five benchmarks following the standard setting (Qu et al., 2023; Bucci et al., 2020): Office Home (Venkateswara et al., 2017), Office-31 (Wang et al., 2019), Image-CLEF (Li et al., 2023), Domain Net (Wen & Brbic, 2024), and Vis DA (Li et al., 2021).
Dataset Splits Yes Office Home (Venkateswara et al., 2017), a dataset across four distinct domains: Art (Ar), Clipart (Cl), Product (Pr), and Real World (Rw) with the first 25 categories as the known, while the subsequent 40 classes as the unknown; 2) Office-31 (Wang et al., 2019), another dataset spanning three domains: Amazon (A), Dslr (D), and Webcam (W), with the first 10 classes as the known and the last 11 classes as the unknown; 3) Image-CLEF (Li et al., 2023), a database with four domains and 12 shared common classes. The first 6 classes are utilized as the known, and the rest as the unknown; 4) Vis DA (Li et al., 2021), a synthetic-to-real (S2R) dataset with 12 classes, with first 6 classes as the known and the rest classes as the unknown; 5) Domain Net (Wen & Brbic, 2024) is the largest dataset, including 345 classes and six domains. Following the previous work (Wen & Brbic, 2024), we use three domains: Painting (P), Real (R), and Sketch (S).
Hardware Specification Yes All experiments are conducted on an RTX-4090 GPU with Py Torch-1.10. ... We utilized six 40GB NVIDIA A100 GPUs to execute the new experiments.
Software Dependencies Yes All experiments are conducted on an RTX-4090 GPU with Py Torch-1.10.
Experiment Setup Yes During target model adaptation, we apply the SGD optimizer with a momentum of 0.9 and a batch size of 32 for all benchmark datasets, following (Li et al., 2023). We set the learning rate to 1 × 10−3 for Office-31, Image-CLEF, and Office-Home. For hyperparameters, λs=0.2, λexo=1. ... But the learning rate has been reduced, specifically lr = 5e−4.