TL;DR
This paper analyzes the interplay between mainstream news and social media (specifically Twitter) during the 2020 US Presidential primaries. By examining sentiment, topic coverage, and temporal influence, the authors reveal how these two information ecosystems interact and shape public perception of candidates.
What Problem Does This Paper Address?
In today's digital age, understanding how citizens form political opinions is complex. Traditionally, mainstream media (newspapers, TV) were the primary information conduits. However, the rise of the internet and social media has created a more dynamic and interconnected landscape. This paper addresses the crucial question of how mainstream news and social media platforms like Twitter influence each other and, in turn, shape public understanding and perception of political candidates during a major event like the US Presidential primaries. The authors aim to quantify these relationships, moving beyond anecdotal observations to provide a data-driven analysis of their coexistence and impact.
Methods
The researchers employed a multi-faceted approach to analyze the data:
- Time Series Analysis: They tracked the volume of news articles and tweets related to candidates over time, looking for correlations and trends.
- Sentiment Analysis: Using the VADER lexicon and rule-based sentiment model, they classified news articles and tweets as positive or negative to gauge public and media sentiment towards candidates.
- Topic Modeling: They identified key topics discussed by candidates on Twitter, in mainstream news, and in their own tweets, using word2vec to represent topics and cosine similarity for matching.
- Co-occurrence Analysis: They measured how often candidates were mentioned together in news articles and tweets to understand how media platforms group candidates.
- Cross-Media Influence Analysis: They used a novel approach involving windowed correlation heatmaps and the Hawkes process to visualize and quantify the temporal influence between news and Twitter. This involved shifting time series data to identify leading and lagging relationships.
- Granger Causality Test: This statistical method was employed to formally test for potential causal relationships between news and Twitter time series data.
- Toxicity Analysis: Google's Perspective API was used to measure the toxicity of tweets mentioning candidates.
Datasets
The study utilized four primary data sources:
- News Articles: Collected via the Currents API, including title, URL, publication time, and excerpts. Full article texts were obtained using the Python Newspaper API.
- Twitter Data: Collected in real-time using the Tweepy Python Streaming API, with a sampling strategy for high-volume candidates like Donald Trump.
- Google Trends: Data was collected directly from Google Trends to measure public interest in search terms.
- Polling Data: Averaged poll data was obtained from RealClearPolitics to provide a baseline.
Key Insights and Ideas
- Intertwined Ecosystems: Mainstream news and Twitter are not isolated but deeply intertwined, with each influencing the other's content and perception.
- Twitter's Leading Role in Influence: While both sources influence each other, the Granger Causality test suggests that mainstream news has a stronger impact on Twitter conversations than vice-versa, though this influence can decay rapidly.
- Candidate Grouping Matters: Media platforms can group candidates together, potentially influencing viewer perceptions by association, and this grouping tendency is not significantly different between news and Twitter.
- Topic Mismatches: Candidates may focus on certain topics in their own communications, but mainstream media and Twitter might associate them with different issues, creating a "topic mismatch."
- Popularity and Toxicity: More popular candidates tend to attract more toxic online discussions, likely due to increased engagement and the polarized political climate.
- Bias in News Sources: Different news sources (liberal, central, conservative) exhibit distinct biases in their coverage and sentiment towards candidates, with conservative media sometimes portraying candidates more negatively than liberal media, even for candidates they ostensibly support.
Key Results
- High Correlation: News and Twitter data showed a strong correlation on average, indicating a shared focus on candidates.
- Sentiment Differences: Liberal-leaning news sources generally showed higher average sentiment compared to conservative and central sources.
- Granger Causality: The p-value for news influencing Twitter was significantly smaller than for Twitter influencing news, suggesting a stronger causal link from news to Twitter.
- Toxicity Levels: Donald Trump and Joe Biden, being the most popular candidates, had the highest toxicity levels in their associated tweets.
- Topic Consistency: Cybersecurity was a topic that received little attention from candidates, mainstream news, or Twitter, suggesting it's a less controversial or less discussed issue.
Limitations and Future Work
- Proposed Explanations: The paper acknowledges that some explanations for observed phenomena (e.g., why liberal/conservative sources have different sentiment) are proposed and not rigorously tested.
- Future Work: The authors suggest further testing of these explanations, exploring if the observed influence patterns hold across different topics or candidates, and expanding the analysis to other platforms and election cycles. They also highlight the need to further investigate the nuances of cross-media influence and topic mismatch.
Summary updated