📰 How to spot Fake News with AI

From: Early Detection of Misinformation for Infodemic Management: A Domain Adaptation Approach

The Problem

Imagine you're in the middle of an infodemic, an overwhelming flood of information where you're bombarded with both truths and falsehoods, initially, you can't tell which is which.
Our mission is to predict whether each piece of information is true or false, but all we have at the start are many unlabeled pieces of information.

The Solution

Let’s dive deeper into DACA (Domain Adaptation with Concept Alignment) and how it tackles the challenge of misinformation detection.

Classifier
The first step is to build classifiers using labelled data. Those classifiers learn to predict whether a piece of information is true or false.

These networks are pre-trained on source data (data that is typically from a domain where labelled data is abundant, for example, political news) and fine-tuned on target data (less abundant, for example, data related to health information during an epidemic)

Covariate Shift Mitigation
The method adjusts for differences in feature (the characteristics of the data) distributions between the source and target domains.

It uses techniques such as re-weighting the source domain. Suppose in the source domain (political news), the word "policy" appears frequently, but in the target domain (health news), the word "vaccine" is more common. If "vaccine" appears less frequently in the source domain, re-weighting can assign higher importance to instances where "vaccine" appears, helping the model recognize its importance in the target domain.

Mapping the source domain features to the target domain feature space: this technique involves transforming the feature representations of the source domain so that they resemble the feature representations in the target domain.

Contrastive Learning
This involves training the model to identify pairs of similar and dissimilar data points.

In simple terms, contrastive learning works by making sure that pieces of information that are similar (should have the same label) are represented closely in the feature space, and those that are different are represented far apart.

Conclusions

The proposed method was evaluated using two real-world datasets and demonstrated superior performance compared to existing state-of-the-art misinformation detection and domain adaptation methods.