/20newsgroups_GBDT

Primary LanguagePython

20newspaper_GBDT

This is a implemetation of GBDT multi-classification with sklearn，the dataset is 20news-19997.tar.gz

Usage

Requirements

  • python 3.6
    • nltk
    • scikit-learn
    • numpy
    • pickle

Run

1. Generate train data & test data

need download and unzip 20news-19997.tar.gz to root dir(eg. './20_newsgroups'), and download crawl-300d-2M.vec, put it in './vectors/'.

python data_process.py

you will get 2 dirs, each with 4 pickle file：

  • ./data_tfidf: word used TD-IDF
  • ./data: word embedding using fasttext, pretrained file: crawl-300d-2M.vec

2. Start training GBDT

python main.py

Result

1. GBDT + fasttext:

image

2. GBDT + TF-IDF:

image

3. GBDT(n_estimators: 60, max_depth: 6) + TF-IDF

image