Mikezz1 / NewsRecommendations

Implementation of several news recommendation methods in Pytorch

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

News Recommendation

Implementation and modification of NRMS model for MIND dataset. Props to yflyl613 for initial implementation.

Requirements

  • python
  • pytorch
  • numpy
  • scikit-learn
  • nltk
  • tqdm
  • recbole (optional)

Usage

  • Clone this repository

    git clone https://github.com/Mikezz1/NewsRecommendation.git
    cd NewsRecommendation
  • Prepare data

    A scirpt download_data.sh is provided to download and unzip all the required data. It will create a new folder data/ under NewsRecommendation/.

    # In NewsRecommendation/
    chmod +x download_data.sh
    ./download_data.sh
  • Start training

    Script run.sh and demo_local2.sh are provied for model training and testing, in which you can modify parameters for the experiment. Please refer to parameters.py for more details.

    # In NewsRecommendation/data/
    cd ../src
    chmod +x run.sh
    
    # train
    ./run.sh train
    
    # test
    ./run.sh test <checkpoint name>
    # E.g. ./run.sh test epoch-1.pt
  • To run baseline models using Recbole

    First, download MIND-small and convert it to atomic files

    cd baselines
    sh prepare_data.sh

    Then, you can modify configs in configs directory as you wish and run training with the following command (if you want to use default models)

    sh run_experiments.sh

    or this command if you wish to use custom model

    python train_custom.py

Results on MIND-small validation set[1]

  • NRMS[3]

    News information AUC MRR nDCG@10 Configuration
    title 66.64 31.90 41.50 batch size 128 (32*4)
    4 epochs
    lr 3e-4

Reference

[1] Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu and Ming Zhou. MIND: A Large-scale Dataset for News Recommendation. ACL 2020.

[2] Chuhan Wu, Fangzhao Wu, Suyu Ge, Tao Qi, Yongfeng Huang, and Xing Xie. Neural News Recommendation with Multi-Head Self-Attention. EMNLP-IJCNLP. 2019.

About

Implementation of several news recommendation methods in Pytorch

License:MIT License


Languages

Language:Python 86.5%Language:Shell 8.4%Language:Jupyter Notebook 5.1%