Blowing up the Twittersphere: Predicting the Optimal Time to Tweet
We can separate our problem into a few different steps. First, we need to model information about a tweet and how successful a given tweet is. Second, given a tweet, user, and post time, we must predict how successful that tweet will be. Finally, we then need to use our predictor to determine the optimal time for a given user to post a specific tweet, i.e. what time maximizes our success prediction for a specific user and tweet. We considered two papers that address similar problems of using Machine Learning to understand interactions in social media and predict success of online content. Lakkaruja, McAuley, and Leskovec consider the connections between title, content and community in social media. From their work, we saw the benefits of breaking features into different models in order to better understand which types of features were having the greatest impact on our final predictions. They also did reasonably extensive language modeling when considering titles, which we considered when designing our own language model to examine the effect of the text of the tweets. (H. Lakkaraju & Leskovec,2013) Tsagkias, Weerkamp, and de Rijke wrote a paper considering a similar problem to ours as they attempted to predict the number of comments an online news article would receive. We were able to draw from their technique of first classifying an article to determine if it would receive any comments before then running a regression to determine how many comments it would receive when we were handling issues of sparsity in our data set.
Retweets and Traffic Over a Day
Research Paper Link: Download Paper