janzenliu / tencentalgo2018

A machine learning competition held by Tencent Social Ads to predict similarity between users: http://algo.qq.com/home/home/index.html

Data Source

.
├──bin
│   ├── ffm-train                   # executable to train an FFM model, see https://github.com/guestwalk/libffm to build it
│   └── ffm-predict                 # executable to make prediction with a trained FFM model, same as above
└── data                            # Toy Datasets
    ├── dog_breed                   # downloaded from https://www.kaggle.com/c/dog-breed-identification/data
    │   ├── Test                    # (folder) unzipped from test.zip
    |   ├── Train                   # (folder) unzipped from train.zip
    |   ├── labels.csv              # unzipped from labels.csv.zip
    |   └── sample_submission.csv   # unzipped from sample_submission.csv.zip
    └── house_pricing               # downloaded from https://www.kaggle.com/c/house-prices-advanced-regression-techniques/data
        ├── data_description.txt    # directly downloaded
        ├── sample_submission.csv   # directly downloaded or unzipped from sample_submission.csv.gz                 
        ├── test.csv                # directly downloaded or unzipped from test.csv.gz
        └── train.csv               # directly downloaded or unzipped from train.csv.gz

TODO

• Try TensorFlow
• Try PyTorch
• Try MXNet
• Try FM
• Try FFM
• Try PNN (Polynominal Neural Networks)
• Try matrix decomposition (MF, SVD/SVD++, SVD for RecSys etc.)
• Try sklearn FeatureUnion and Pipeline
• Try tuning hyperparameters with hyperopt
• Try stacking
• Try ensemble with raw features involved (my idea)
• Try synthetic oversampling
• Try cost-sensitive learning for imbalanced data (not only with imbalenced labels, but also features)
• Try ensembling for imbalanced data
• Try cython
• Try TSNE
• Add simple data visualization
• Add simple testing
• Implement (or just copy from somewhere) Bayesian smoothing for CTR/CVR
• Implement simple A/B test
• Implement a full tunable and testable pipeline (it's fine to be simple, but it must be complete)
• Fix globals problem in info_utils