This work is building on my masters thesis at the Information Systems chair of TU Darmstadt. It applies deep learning techniques to the problem of tweet engagement prediction.
Build increasingly complex deep neural networks for predicting retweet counts. An emphasis is put on the development of models suitable for production settings, i.e., capable of making real-time predictions. In other words, preprocessing should be kept at a minimum and all input data should stem from tweet objects as fetched from the public Twitter API.
Two specific tasks will be examined:
- Regression task: predict positive, real-valued number for eventual retweet count
- Classification task: predict certain range for eventual retweet count
Classes for classification task: 0, 1-9, 10-99, 100-999, >=1,000
Assumption: All examined tweets are at least one month old at the time of collection, s.th. the retweet process can be seen as exhausted (based on observations by Kwak et al. 2010).
Tweets come from a variety of user groups, namely:
- corporations (Fortune 500 and startups)
- politicians (US senators and governors)
- celebrities (athletes, musicians, actors,...)
- journalists (political, pop culture,...)
For the assembled accounts, all tweets from March 2016 to March 2018 were collected, leaving out retweets without a quote. The final data set contains a total of around 1.3 million tweets.
Valid Twitter credentials are needed in order establish an API connection.
Credentials should be stored in an .env file and will be loaded in the
respective modules.
All results can be reproduced using notebooks in the nbs directory.
For further questions regarding this work, contact the repository owner: