A Data Analysis of TED
Data Science Introduction
There are plenty of videos uploaded online each day, the views reflect, at least, the success of videos’ first impressions. What makes videos more attractive than others, and what makes viewers click the video, are important for people who want to do self-marketing and business that try to make an advertisement. A talk views analysis with data scraped from TED.com will help find the key factors contributed to high video views.
TED devoted to spreading ideas through powerful talks began in 1984. Its topics first covered Technology, Entertainment and Design, which built up the ‘TED’ with the initial letter of the three words. Now, TED covers almost all topics - from Science to Business to Global Issues - in more than 100 languages.
Data Collection & Analysis
The data was scraped using Scrapy from the TED website. There are about 4500 videos in total from June 2006 until now, information includes:
and English Transcript collected.
With data scraped, the talk topics’ distribution and correlation between number of talk views and different features will be defined.
General Talk Views Distribution
The histogram shows the right skewed data with most data falling in the range from 0 to 2 million, with extremely large outlines that over 10 million. By ranking the data based on view number, the most viewed talk is “Do Schools Kill Creativity?”, with almost 65 million views. The box plot contains data within 1 standard deviation (about 68% of data falling in this range), with a median at 1.2 million, which means that there are at least 50% of the talks having 1.2 million views.
Talk Categories Vs. Views
Definitely, it is necessary to check out how the views change across different categories. However, each talk might have more than one category, thus a new dataframe grouped by categories is created. First, check the number of talks in each category. Among the top 20 categories shown in the graph, technology and science are well ahead of the rest, the following are culture, society, global issues etc. It is reasonable that these topics are extensive that contain more specific and narrowed topics.
To reflect the popularity of talks in different categories, the project values the rank by comparing the average number of views. At this time, categories such as technology, science and culture, that with more videos do not take top place, whereas, psychology(introvert, mindfulness, mind, happiness, depression, etc) and success science(success, productivity, motivation) prompt more clicks. Therefore, compared to academic talks, people hold more interest in emotional and abstract topics.
How titles will influence people’s decisions of whether to watch the talk? To analyze the contribution of wording, a package called Word Cloud is used. Word clouds is a method for visually presenting word frequencies, the more frequent the word is used, the larger and bolder it is displayed. Furthermore, when mining for the word used in talk titles, stopwords, words hold no meaning, such as ‘is’, ‘the’, are removed. Finally, the word clouds indicates ‘life’, ‘world’ are the two words that appear most frequently in talk titles as they are closely related to culture, society and global issues.
Nonetheless, frequent use of the words does not mean high views of the video. From the barplot, it is interesting to see that ‘spam’, ‘replay’,’email’ share the same amount of views. With the assumption that the three words in the same title, it is not hard to check that there is a video titled ‘This is what happens when you reply to spam email’, holding 56.7million views. Besides, viewers are also interested in words close to daily actions, such as ‘procrastinator’, ‘orgasm’, and ‘unsubscribe’, or words that describe characters of humans, such as ‘vulnerability’ and ‘perseverance’.
The further step is using sentiment analysis to test if the subjectivity and polarity of titles will influence the user’s selections. Sentiment analysis refers to the use of natural language processing, text analysis, and computational linguistics to identify emotion and subjective information. The project calls the sentiment analysis API from TextBlob to perform the NLP tasks.
Polarity is the emotion expressed in the text, as a float lies in the range of -1 to +1. -1 means a negative statement, on the other hand, +1 is a positive statement, 0 is neutral. Obviously, the barplot and the scatter plot show the majority of the titles are neutral and there is no clear relationship between views and title polarity.
Subjectivity generally refers to the measure of how subjective or objective the text is. Subjectivity is also a float which lies in the range of [0,1]. As subjectivity gets closer to 1, the text is more subjective, otherwise, more objective. In the bar chart, most talk titles are in the range of 0 to 0.1, which shows little subjectivity. Similarly to the title polarity, the scatter plot shows a weak relationship between views and title subjectivity, either.
Using Data to Analyze Talk Length vs. Views
Last but not least, the project also tries to figure out if the length of talk shown on the title page of the video will influence the number of views. The assumption is people would not like to click the long videos which take too much time. The graph indicates an inversely proportional relationship between talk length and views, the longer the talk is, the less the number of views is. Unless it is an extremely interesting video, people would like to spend time on other shorter videos.
TED is a special video collection which focuses on speech and conference. As a sample of video platforms, the data scraped from TED.com shows the attractiveness of emotional topics and simultaneously, titles which are close to life. However, the polarity and subjectivity of title do not have the influence in viewers selection of talks, this is due to particularity of speech title. The result may be different on other video platforms, based on the goal of videos.
Finally, the video length takes an important role in view number, the overlong videos will reduce the viewers’ desire of watching. As a further step, with the allowance of time and date, the contribution of speakers’ backgrounds and number of transcripts to views could be studied.
The skills I demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.