HYBRID DEEP LEARNING FRAMEWORK FOR ACCURATE DETECTION OF BIASED NEWS TO PROMOTE TRUSTWORTHY INFORMATION CONSUMPTION

Abstract

In the digital era, news disseminates quickly on social media, online, and news collecting platform. This helps users to stay informed, but the problem of biased reporting distorts public sentiment, opinion and spreads misinformation.  Existing Machine learning Deep Learning model find it challenging with the manipulative and context-dependent nature of bias in text, and detect biased data from an imbalanced dataset, achieving low accuracy. This paper proposes a hybrid approach named “CHL” that combines convolutional layers (CNN), a hierarchical attention network (HAN), and a gradient boosting approach (LightGBM). CNN picks up local language features, and HAN is used for word-level and sentence-level text to capture long-distance dependencies, and both capture local and global text. LightGBM enhances classification accuracy with computationally efficient decision boundaries and handles the minority class better. For this paper, we used a dataset collected from Kaggle with 147111 data, where biased data is labeled 1 and non-biased data is labeled as 0. Previous studies detected biased news using a hybrid model and achieved accuracies ranging from 68\% to 82\%. Our proposed model outperforms these results and gets train accuracy 0.8688 and test accuracy 0.8607. Average interface time per sample is 0.4924ms, Precision 0.8607, Recall 0.9823, F1-Score 0.9223, MCC 0.3331, MAE 0.1393.