Transferring Structure Knowledge: A New Task to Fake news Detection Towards Cold-Start Propagation

Lingwei Wei,Dou Hu,Wei Zhou,Songlin Hu

2024-07-13

Abstract:Many fake news detection studies have achieved promising performance by extracting effective semantic and structure features from both content and propagation trees. However, it is challenging to apply them to practical situations, especially when using the trained propagation-based models to detect news with no propagation data. Towards this scenario, we study a new task named cold-start fake news detection, which aims to detect content-only samples with missing propagation. To achieve the task, we design a simple but effective Structure Adversarial Net (SAN) framework to learn transferable features from available propagation to boost the detection of content-only samples. SAN introduces a structure discriminator to estimate dissimilarities among learned features with and without propagation, and further learns structure-invariant features to enhance the generalization of existing propagation-based methods for content-only samples. We conduct qualitative and quantitative experiments on three datasets. Results show the challenge of the new task and the effectiveness of our SAN framework.

Social and Information Networks,Artificial Intelligence,Computation and Language

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the challenge of detecting fake news in the cold - start propagation scenario. Specifically, existing fake news detection methods usually rely on extracting effective semantic and structural features from content and propagation trees, but in practical applications, when there is no propagation data, the performance of these methods will be greatly reduced. Therefore, this paper proposes a new task - cold - start fake news detection, aiming to use previously available propagation and content information to identify news that only contains content but lacks propagation data.

### Main problems and background

1. **Limitations of existing methods**:
- Most of the existing fake news detection models rely on the structural information in the propagation tree. These models learn the features of content and propagation structure simultaneously during training.
- When facing news without propagation data (i.e., cold - start news), these models cannot generalize well, resulting in a decline in performance.

2. **Challenges in the cold - start scenario**:
- Cold - start propagation refers to the situation where there is a lack of propagation data during detection, which is very common in practical applications because obtaining propagation data usually requires a lot of time and resources.
- In this case, directly applying existing propagation - based detection models will lead to a significant decline in performance because these models cannot handle data lacking propagation structure.

### Solution

To solve the above problems, the author designed a simple but effective framework - Structure Adversarial Net (SAN) to transfer the propagation structure knowledge to samples that only contain content. SAN estimates the differences between features with and without propagation structure by introducing a structure discriminator, and further learns structure - invariant features to enhance the generalization ability of existing propagation - based methods for samples that only contain content.

### Model architecture

- **Input**: It includes the content \(x\) of the source news and the propagation tree \(G\).
- **Feature extraction**: Use the encoder \(f_{\text{enc}}(x, G; \Theta)\) to calculate the hidden representation \(h\), where \(\Theta\) is the trainable parameter.
- **Classifier**: Predict the label probability \(\hat{y} = f_{\text{cls}}(h; \theta_f)\) through the fully - connected layer and softmax function, where \(\theta_f\) is the learnable parameter of the classifier.
- **Structure Adversarial Network**: Introduce a structure discriminator \(f_d(h; \theta_d)\) to predict whether the high - dimensional representation contains structural attributes, where \(\theta_d\) is the learnable parameter of the discriminator.

### Objective function

- **Classification loss**: \(L_{\text{CLS}}(\Theta, \theta_f)= -y\log(\hat{y})-(1 - y)\log(1-\hat{y})\)
- **Discriminator loss**: \(L_d(\Theta, \theta_d)= -y_d\log(\hat{y}_d)+(1 - y_d)\log(1-\hat{y}_d)\)
- **Final optimization objective**: \(L_{\text{SAN}} = L_{\text{CLS}}(\Theta, \theta_f)-L_d(\Theta, \theta_d)\)

### Experimental results

The author conducted experiments on three public datasets and constructed two different cold - start propagation settings (general cold - start and event - aware cold - start). The experimental results show that the performance of existing propagation - based detection methods decreases in the cold - start scenario, while the SAN framework can significantly improve the detection performance of these methods in the cold - start scenario.

### Conclusion

This paper proposes a new task - cold - start fake news detection, and designs a simple but effective Structure Adversarial Net (SAN) framework.