Here is one of my submissions to Kaggle challenge 'Bag of Words meets Bags of Popcorn'.

It is based on the idea of combining pre-trained word2vec embeddings with convolutional networks proposed by Yoon Kim [http://arxiv.org/abs/1408.5882].

The code consists of two IPython Notebooks:

1. Process Kaggle Dataset Train+Test.ipynb contains data pre-processing.

2. Train CNN IMDB.ipynb implements convolutional network with one convolutional layer.

This model (trained for 3 epochs) yields AUC = 0.96823 (on test data).

Ensemble of three convolutional networks (having different number of convolutional layers and feature maps) gives AUC = 0.97310.

• Python 2.7.11
• IPython / Jupyter Notebook
• NumPy
• Theano 0.8.2
• Keras 1.2.0
• scikit-learn 0.17.1
• Pandas