In this paper, we present Aligned Adaptation Networks (AAN) to match both the marginal and conditional distributions across domains for UDA.

Two datasets Amazon-Review and Amazon-Text are provided in dataset/ with .json.

Maximum mean discrepancy (MMD) and Conditional MMD (CMMD) are implemented in model/criterion.py

AANs and AAN-As are implmented in model/models.py with three different types of extractors $T(\cdot;\theta_T)$: MLP, TextCNN and BertGRU.

Model trainings are represented in aan_mlp.ipynb, aan_cnn.ipynb and aan_bert.ipynb.

1. Clone this repsitory.

git clone https://github.com/gregbuaa/aan_model.git
cd aan_model

2. Install the dependencies. The code runs with PyTorch-1.2.0 in our experiments.

pip install -r requirements.txt 

3. Play with the Jupyter notebooks.

jupyter notebook

Run all the notebooks to reproduce the experiments on Amazon_Feature with MLP extractor, Amazon_Text_CNN with TextCNN extractor and Amazon_Text_BERT presented in the paper.

AANs can be extended to other different tasks such as Image Classification easily.

@inproceedings{zhang2021discriminative,
  title={Discriminative Feature Adaptation via Conditional Mean Discrepancy for Cross-domain Text Classification},
  author={Zhang, Bo and Zhang, Xiaoming and Liu, Yun and Cheng, Lei},
  booktitle={International Conference on Database Systems for Advanced Applications (DASFAA)},
  year={2021}
}