runnerxin/HyFea

HyFea: Winning Solution to Social Media Popularity Prediction for Multimedia Grand Challenge 2020

HyFea

HyFea: Winning Solution to Social Media Popularity Prediction for Multimedia Grand Challenge 2020

The directory tree should look like this:

${ROOT}
├── data
│   ├── train_all_json
│   ├── train
│   ├── test_all_json
│   ├── test
│   ├── none_picture.jpg
│   ├── user_additional.csv
│   ├── alltags_feature.csv
│   ├── feature_data_530.csv
│   └── title_feature.csv
├── save_model
│   ├── KFold_catboost_0.pkl
│   ├── KFold_catboost_1.pkl
│   ├── KFold_catboost_2.pkl
│   ├── KFold_catboost_3.pkl
│   └── KFold_catboost_4.pkl
├── readme.txt
├── run_step2.sh
├── KFold_catboost.json
├── download_img_and_user.py
├── get_data_feature.py
├── test_k_fold_model.py
└── train_k_fold_model.py

1 Dependencies

1. We have been implemented and tested our code on Ubuntu 16.04.5 with python == 3.6.

2. Python packages:

   pip install requests beautifulsoup4 scipy Pillow gensim sklearn pandas catboost lightgbm

2 Quick Start

1. You can test the model saved in folder ./save_mode to get the result KFold_catboost.json base on our processed data (feature_data_530.csv, alltags_feature.csv, title_feature.csv) by the following command:

   python test_k_fold_model.py

2. If you want to reproduce the process to get the processed data, you have to take the following steps:

   2.1 Download the GloVe word embedding, then unzip glove.42B.300d.zip, put the file glove.42B.300d.txt to folder ./data.

   2.2 Download Social Media Prediction Dataset from http://smp-challenge.com/ , unzip the train_all_json.zip to ./data, and put all test data files except test images to ./data, download SMP_test_images.zip and unzip it to folder ./data/test.

   2.3 Then you can run command python download_img_and_user.py to download the train images to saved in ./data/train and crawl user additional information to saved in ./data/user_additional.csv.

   Note: You can also directly use the data user_additional.csv provided by us.

   The new added structure of the data folder should be:

   .
   ├── data
   │   ├── test
   │   │   ├── 1@N18
   │   │   │   ├── 56783.jpg
   │   │   │   └── ...
   │   │   └── ...
   │   ├── test_all_json
   │   │   ├── test_additional.json
   │   │   ├── test_category.json
   │   │   ├── test_imgfile.txt
   │   │   ├── test_tags.json
   │   │   ├── test_temporalspatial.json
   │   ├── train
   │   │   ├── 7107177@N05	
   │   │   │   ├── 23511051534.jpg
   │   │   │   └── ...
   │   │   └── ...
   │   ├── train_all_json
   │   │   ├── train_additional.json
   │   │   ├── train_category.json
   │   │   ├── train_img.txt
   │   │   ├── train_label.txt
   │   │   ├── train_tags.json
   │   │   ├── train_temporalspatial.json
   │   │   └── train_userdata.json
   │   ├── ...
   │   ├── user_additional.csv
   │   └── glove.42B.300d.txt
   └── ...

   2.4 Next you should run the command python get_data_feature.py to get the feature data file feature_data_530.csv, alltags_feature.csv, title_feature.csv, and put all of them in folder ./data.

   2.5 Run command python train_k_fold_model.py to train the model, and model will be saved in the folder ./save_model.

   2.6 Finally you can test the model like step1. run python test_k_fold_model.py to get the submission file.

3. If the Social Media Prediction Dataset and GloVe word embedding have been downloaded, You can do all steps in 2 by running the script sh run_step2.sh .