AskReddit popularity score prediction system with an automatic title refiner

Abstract

In this project, we develop a neural network classifier for popularity prediction of the AskReddit post titles. Moreover, we invent an automatic sentence refiner for suggestions of new candidate sentences that are likely to gain more popularity. To perform sentence modification, the refiner makes a single change to the original sentences by word insertion, deletion, and replacement. Any modified sentences with higher predicted probability of becoming popular are suggested as candidate titles. The results show that our model was able to predict the popularity of AskReddit submissions with an accuracy of 71%. We finally carried out an evaluation by human on all suggested sentences. Of all, 77 % of sentences with one word replaced were successfully acceptable, 71% with a word being deleted and 60% with a word being inserted. Those candidate titles can be used as alternatives to their original sentences for AskReddit users.