sindhri / IMDB

Use IMDB movie reviews to classify the sentiments either negative or positive via deep learning: MLP, CNN, LSTM

IMDB movie review sentiment classification

Use IMDB movie reviews to classify the sentiment, either positive or negative
Credit: Deep Learning with Python by Jason Brownlee

EDA

Shape of the data: 50000 reviews Unique classes of the predicted variable: [0, 1] Total number of unique words: 88585 Review length: Mean 234.76 +/- 173 words

Parameters: Maximum unique words: 5000 Maximum length of review: 500 words Mumber of vectors for word embedding: 32 50/50 split of training and validation

Models

1. MLP

batch size 128, only 2 epochs is able to reach an accuracy of 87.16%. It only took a minute to run it.

2. 1-d CNN

Two epochs reached an accuracy of 88.23%, and it only took a minutes to run it.

3. LSTM

3 epoches, batch size 64
Much slower Final accuracy: 87.62%, comparable to previous, but much much much slower, saved

4. LSTM with dropout layers

3 epoches, batch size 64 Final accuracy: 86.78%, slightly worse

5. LSTM with dropout at the concurrent gate

3 epoches, batch size 64 Final accuracy: Accuracy: 87.73%, not bad, saved

6. LSTM + CNN

Much faster than LSTM along Final accuracy: Accuracy: 87.93%, not bad.

Conclusion:

1-d CNN outperform simple MLP slightly and reached an accuracy of 88.23% with only 2 epoches. Training more epoches will likely to improve the accuracy