The idea behind this dataset is to try to predict whether a particular comment would be highly up-voted or down-voted given the parent comment.

Presently, I have a model that does fairly well (70+% accuracy on several validation sets) using neural network with a combination of BOW and word embeddings inputs. The BOW input feeds into a dense feed-forward layer, and the word embeddings feed into an LSTM layer.

The code is mostly in Python using Keras and will be posted shortly.

Content

This dataset is a subset of a much larger dataset of 380+ million reddit comments. This dataset contains 4 million of the reddit comments, 2 million of which are the lowest scored (highly downvoted), and 2 million of which are the highest scored (highly upvoted).

For convenience, the dataset includes the text and other metadata of the parent comment.