Computer analysis of nonverbal cues in video blog entries (vlogs) can predict human perceptions of personality, according to research publish in the latest IEEE Transactions on Multimedia.
The researchers gathered human perceptions of videos using Amazon’s Mechanical Turk crowdsourcing platform. Mechanical Turk (MTurk) participants were paid to watch videos and evaluate the personality traits they observed in the videos.
The researchers compared participants’ evaluation of the same video and looked at their completion time to confirm the participants were honestly evaluating the videos. I asked an author of the paper, Joan-Isaac Biel, how they could be sure people were not cheating. He explained that the participants would not get paid if they were caught cheating. When he interviewed them after the work, they seemed to have found the work of watching and evaluating videos for money appealing.
The MTurk evaluations turned out to be correlated to YouTube “metadata” such as number of views, number of times favorited, number of raters, and number of comments. Number of views and average rating turned out to be positively correlated with MTurk ratings of Extraversion, Openness to Experience, and Conscientiousness. The relationship between Agreeableness, and number of views, however, showed a U-shaped curve. The extremes of this personality dimension, “pleasant” and “disagreeable”, appear to cause a vlog to get more views. The “pleasant” side of Agreeableness turned out to be associated with a greater average number of “likes”, suggesting “liking” captures viewers’ impression of Agreeableness. Emotional stability appeared to have no correlation at all with number of views. The researchers were able to fit curves to the data to predict Youtube metadata values based on MTurk participants’ personality trait annotations. This was true even for the U-shaped correlations with Agreeableness, which can be represented by a parabola.
When they compared the MTurk participants’ annotations to software analysis, they found Extraversion to be correlated to audio and visual cues. Conscientiousness and Agreeableness were correlated with visual cues. Agreeableness was not highly correlated to most of the cues the software detected. The researchers suspect that if the software could detect smiling or eye gaze, they would have been more correlated to Agreeableness. Studying computer-detected cues may shed light on the cues humans use in making first-impression assessments of personality traits.
My first thought about this is that this research is that this research could eventually lead to a way to predict if a video will go viral. I told him how my city’s mayor heard that Grand Rapids, MI city leaders made a Youtube video that got 5 million views and thought our city ought to be able to do the same thing. He apparently did not realize how elusive virality is. Biel told me there is a lot of interest in predicting which videos could go virual. His research, he says, cannot be used to predict virality. His work looks at averages of Youtube metadata. Virality has to do with a video being an extreme outlier, which is different from the phenomenon responsible for the correlation between cues, human personality assessments, and average Youtube statistics.
I asked Biel if perhaps we are in a golden age of authenticity when people produce videos with no idea how viewers will respond. He rejects the notion that his research lends itself to manipulating people. Rather, he says this “technology should be meant to understand people’s interests.”
If people do find a way to use computer-detectable cues to make videos more likely to get more views or other favorable statistics, eventually viewers will become aware of the techniques and they won’t work. Perhaps in future decades people will look at historic videos and see cues that the average contemporary viewer would not be aware of.