An application that predicts post popularity from the features of a given Instagram post.

Notes:

12/1

• Command used to scrape Instagram images

instagram-scraper -f ig_users.txt --retain-username --media-types image --maximum 50 --media-metadata --include-location

11/15

• Now the dataset is cached (so it doesn't have to read in a bunch of different JSON files every time).
• There is now a consistent train/test data split.
• The algorithm now runs scikit linear regression model.

11/14

• Just installed scikit-learn (which depends upon numpy and scipy) and I don't understand virtualenv.
• Pipenv just stopped working. Attempted to uninstall everything and reinstall everything. Nothing works.
• Starting to look at the project again for the upcoming progress report.
• Current framework: Linear Regression using squared loss
• Color and other basic photo qualities are important.
  • Source: https://www.curalate.com/blog/6-image-qualities-that-drive-more-instagram-likes/
• Faces are important.
  • Source: http://www.news.gatech.edu/2014/03/20/face-it-instagram-pictures-faces-are-more-popular
• The captions are important.
  • Specifically, the number of emojis and hashtags.
  • Source: https://sproutsocial.com/insights/instagram-stats/
  • Length is not important, but @mentions are.
  • Source: http://get.simplymeasured.com/rs/simplymeasured2/images/InstagramStudy2014Q3.pdf?mkt_tok=3RkMMJWWfF9wsRolua%252FAZKXonjHpfsX57%252BwtX6a2lMI%252F0ER3fOvrPUfGjI4CTsViI%252BSLDwEYGJlv6SgFQrDEMal41bgNWRM%253D
• Timing is important.
  • Specifically, time of day and day of week.
  • Source: https://blog.bufferapp.com/instagram-stats-instagram-tips

10/26

• Downloaded OpenCV!
• Working on the infrastructure to analyze images in bulk.
• Possible cause of skew: new pictures (< 1 week old). New pictures may not be fully saturated. Their like counts could still be in flux.
• Baseline algorithm is implemented as follows:
  • Input: a user's profile and a photo.
  • A mean is calculated across the like counts of the 50 most recent photos.
  • A coefficient is calculated based on the features of the photo. For the baseline, this coefficient is either 1.1 (if face is detected) or 0.9 (if face is not detected).
  • Output: coefficient multiplied by the mean.
• Baseline algorithm has 46.60% average error rate. This means that on average, the algorithm overestimated or underestimated the like count for a given image of a user by 46.60%. The human error rate was ~30%, so there is a large amount of avoidable bias to overcome.
• Possible ideas:
  • Belly button detection
  • Proportion of photo that is face
  • If user's face is in photo

10/25

• Scraped the most recent ~50 photos of the top 40 instagram celebrities using instagram-scraper. Ignored @taylorswift because she only had 25 photos. Sad.
  • Source: https://github.com/rarcega/instagram-scraper
• We say we have approximately 50 photos because we are excluding videos.