CADA

Reproduced code of ACM-MM 2023 <Improving Rumor Detection by Class-based Adversarial Domain Adaptation>

Model Architecture

In the paper Improving Rumor Detection by Class-based Adversarial Domain Adaptation, there are two components: The rumour detection model and the class-based adversarial domain adaptation module.

In our codes, they are implemented by three modules: the feature extractor, the label classifier and the domain classifier, as shown in CADA.py.

Dataset

To experiment the model's cross-domain detection ability, we use the Twitter dataset and Twitter-COVID19 dataset as the source domain and target domain datasets. They can be downloaded from NAACL-2022 ACLR. Download the Twitter and Twitter-COVID19 datasets, unzip them into directories ./data/in-domain and ./data/out-of-domain. The finalised directory architecture should be like this:

├── project_root/
│   ├── data/
│   │   ├── in-domain/
|   |   |   |   ├──Twitter/
|   |   |   |   ├──Twittergraph/
│   │   ├── out-of-domain/
|   |   |   |   ├──Twitter/
|   |   |   |   ├──Twittergraph/
│   ├── ...
│   ├── README.md

Implementation setting.

In the reproduce, we design the CADA as an individual object which takes feature extractor, label classifier and domain classifier as parameters. You can substitude the three components with any modules (eg. BERT, GCN, MLP). The original paper uses GACL and BERT as the framework of these three components. Instead, for the conveniece of implemetation, we use the BiGCN.

Loss Calculation Controversy.

In the paper of the repository, the loss in the 2nd round of training is defined as

L = label_loss - domain_loss

However, this will make negative optimisation on the domain discriminator. Therefore, we follow the definition in the original paper of adversarial domain adaptation (DANN), and set the loss function as:

L = label_loss + domain_loss

The Gradient Reverse Layer will control the gradient flow and only optimise the domain discriminiator, and make adversarial learning on the feature extrator.

To Run

requirement:

numpy
pytorch
torch_geometric
torch_scatter
sklearn
random

A detailed requirements file with version on my device is shown in environment.yml

Run the code

python main.py

gszswork/CADA