Diversity in news consumption is essential for fostering a democratic society. However, the rise of social media platforms and recommendation systems has given rise to social phenomena such as filter bubbles and selective exposure, limiting users’ a...
Diversity in news consumption is essential for fostering a democratic society. However, the rise of social media platforms and recommendation systems has given rise to social phenomena such as filter bubbles and selective exposure, limiting users’ access to a diverse range of news perspectives. Additionally, the increasingly polarized media landscape, characterized by the segregation of political viewpoints, poses a significant threat to public discourse. Understanding the biases embedded within news articles and leveraging this knowledge to promote diverse news consumption is therefore of critical importance. Motivated by these challenges, this dissertation proposes two approaches to advance political bias analysis and recommendation systems, improving the performance and fairness of deep learning models in these areas.
First, I seek to propose a novel bias prediction model that can mitigate the news outlet’s influence in prediction. I first identify distinct text patterns associated with specific news sources or publishers that are minimally relevant to predicting the political bias of a news article. I then conduct comprehensive experiments to investigate (i) whether existing models trained to predict political bias can also accurately predict the source, and (ii) whether these models change their predictions when a distinct pattern from a source with a different political stance is introduced into a news article. The experimental results reveal that all existing models tend to predict the source, even when trained solely to predict bias. Based on these findings, I propose deep learning models for political bias prediction that avoids learning source-indicative patterns that are only used in the given news source.
Then, I study the diversity in the recommendation systems, a crucial component in various platforms from shopping to online social networks. A key challenge in recommendation systems is to leverage diversity, exposing or recommending diverse items to individuals. Despite much effort on studying diversity in the recommendation systems, little work has focused on estimating how much an item will potentially affect user’s diversity experiences by contributing to consecutive recommendations in a session. To this end, I propose a deep learning model that can predict diversification scores, which is a degree of potential contribution to users’ diversity experiences of an item. The proposed model adopts multiple graph neural network layers with a novel attention mechanism that can capture the features of a given item and its related items in terms of recommendation. To prove the effectiveness of my approach, I collect a large dataset of video recommendations from YouTube and conduct random-walk experiments to simulate user traces. The evaluation results on the dataset shows that the proposed model accurately predicts each item’s contribution on user diversity experiences.