Solution for the Quora Question Pair contest hosted on Kaggle[1]
Big thanks to the authors of all kernels & posts, which were of great inspiration and some features were derived based on them.
Kaggle Profile : Daft Vader
Final submission is the ensemble of 7 LSTM based models and 6 xgboost models with different hyperparameters. LSTM model architecture is based on lystdo[2] model on kaggle kernel. Keras architecture of the model is shown below:
![[model_1.png]](https://github.com/dalamar66/kaggle-question-pairs-quora/blob/master/model_1.png)
Nadam optimizer and binary cross-entropy is used as the loss function, early stopping round by monitoring validation loss is used to avoid overfitting. To create different permutations of hyper-parameter in both xgboost model and lstm, values of hyper-parameter is chosen at random from a defined set.
Dataset is split into two parts in 90:10 ratio. Each LSTM model was trained for approx 19 eochs and optimal weights were saved using model checkpoint by monitoring validation loss, each epoch took around 220s on Nvidia GTX 1070 gpu and each xgboost model is trained using early stoping round of 50. The configration gives a logloss of 0.14672 on private leaderboard.
| Model | LogLoss |
|---|---|
| Avg(7 LSTM + 6 XGB) | 0.14672 |
| LSTM Model | 0.15918 |
| XGB Model | 0.15392 |
- Python 2.7.13
- jupyter 4.3.1
This repository contains 5 notebooks performing different function, it can be used by simply running the notebook server by using standard Jupyter command:
$ jupyter notebook
feature_engineering.ipynb : Basic feature Generation
page_rank.ipynb : Page rank feature generation
glove_840b+leaky_features.ipynb : LSTM model training
xgboost.ipynb : XG Boost model training
averagging.ipynb : Ensemblling the models
This project is licensed under the MIT License - see the LICENSE.md file for details