Keywords: News, media, journalism and public opinion

In response to the prevalence of bias in media content and its potential effects, scholars have made various attempts to automatically detect bias in media content in the hope to expose media consumers to more balanced viewpoints (Park et al., 2009; Munson & Resnick, 2010). However, these attempts are mostly limited to textual data (Groseclose & Milyo, 2005; Gentzkow & Shapiro, 2010). Yet, visual factors like facial appearances and emotional expressions can shape how we perceive political candidates (Sullivan & Masters, 1988; Marcus, 2000). This study draws inspiration from an emerging field, computer vision. Computer vision aims to imitate the human vision’s ability to perceive and understand visuals, being applied in a variety of contexts, such as medical imaging, facial recognition and automated driving (Szeliski, 2016). Applying latest techniques in facial recognition and emotion detection, this study provides a timely look into the visual representation of the two candidates in the 2016 U.S. presidential election across different media outlets.

Method
This research preselected a list of news websites that were diverse in their ideological positions based on the various studies on measuring media bias (Mitchell et al., 2014). The seven websites included Slate, the New York Times, Huffington Post, CNN, USA Today, Fox News, and TheBlaze. Next, with Google image search, the study searched for images of the two political candidates—Hillary Clinton and Donald Trump—that were limited to a specific news website (e.g., “Hillary Clinton site:nytimes.com”). A total of 14,955 images were retrieved at the end of November in 2016. Next, the analysis used an online facial detection tool, Microsoft Face API, to identify the faces belonging to the two candidates. Then, this study used Microsoft Emotion API to detect the emotions associated with these faces. At last, we also instructed a group of human coders (N = 244) from Amazon Mturk to rate their impressions of a subset of these images.

Results
First, Hillary Clinton displayed more happiness than Donald Trump, which led her to be rated as more friendly, competent, and attractive. In contrast, Donald Trump expressed more anger, sadness, disgust, and more surprise, which made him look negative, but also more dominant. There is also a significant interaction between media outlets and political candidates for detected happiness, anger, sadness, fear, disgust, surprise. In terms of happiness, sadness, anger, and disgust, the gaps between the two candidates were narrower in conservative media like Fox News and TheBlaze than the gaps in liberal media like Slate, Huffington Post, and CNN. It was more striking in liberal media outlets than in conservative media outlets that Clinton had more positive emotions and Trump expressed more negative emotions, predominantly anger.

Discussion
The results demonstrated that candidates’ emotional expressions are an extreme salient form of visual bias: They substantially shaped viewers‘ judgment of media slant and evaluations of candidates and successfully distinguished liberal from conservative media. The study also demonstrated the potential of applying computer vision methods in analyzing visual media on a large scale.