Transfer and Alignment Network (TAN)

Data and code for paper titled Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 paper)

Generalized Category Discovery (GCD) is a crucial task that aims to recognize both known and novel categories from an unlabeled dataset by leveraging another labeled dataset with only known categories. In this paper, we propose a Transfer and Alignment Network (TAN), which incorporates two knowledge transfer mechanisms (Prototype-to-Prototype Transfer and Prototype-to-Instance Transfer) to calibrate the biased knowledge and two feature alignment mechanisms (Instance-to-Prototype Alignment and Instance-to-Instance Alignment) to learn discriminative features.

Contents

1. Data

2. Model

3. Requirements

4. Running

5. Results

6. Thanks

7. Citation

Data

We performed experiments on three public datasets: clinc, banking and stackoverflow, which have been included in our repository in the data folder ' ./data '.

Model

An overview of our model is shown in the figure.

Requirements

• python==3.8
• pytorch==1.11.0
• transformers==4.19.2
• scipy==1.8.0
• numpy==1.21.6
• scikit-learn==1.1.1
• pytorch-pretrained-bert==0.6.2

Running

Pre-training, training and testing our model through the bash scripts:

sh run.sh

You can also add or change parameters in run.sh (More parameters are listed in init_parameter.py)

Results

It should be noted that the experimental results may be different because of the randomness of clustering when testing even though we fixed the random seeds.

Thanks

Some code references the following repositories:

• DPN
• DeepAligned

Citation

If our paper or code is helpful to you, please consider citing our paper:

@inproceedings{an2024transfer,
  title={Transfer and alignment network for generalized category discovery},
  author={An, Wenbin and Tian, Feng and Shi, Wenkai and Chen, Yan and Wu, Yaqiang and Wang, Qianying and Chen, Ping},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={10},
  pages={10856--10864},
  year={2024}
}