This is the 17th place out of 1388 teams (top 2%) solution to the Kaggle machine learning competition What's Cooking: https://www.kaggle.com/c/whats-cooking
It employs a fairly simple neural network classifier, giving a classification accuracy of around 81.818% : https://www.kaggle.com/c/whats-cooking/leaderboard
Code is tested to work with Python 2.7 under Ubuntu 14.04 with GeForce GT 755M (cuDNN v2 was employed as well).
Assuming train.json and test.json exist in the directory, from command line:
python cook_it_up.py
A submission file will be autimatically created in the correct format.
Many things were not tried so there is room for performance improvement. Couple of ideas:
- Advanced text processing features, e.g., tf-idf
- Mapping certain words to eachother with normalized Levenstein distance, e.g. mayonnais & mayonnaise
- Hyperparameter optimization of the neural network e.g. number of layers, number of units, activations, batch_size, regularization etc.
oguzhan.gencoglu@topdatascience.com, oguzhan.gencoglu@tut.fi
License: MIT