A Repo for Benchmarking News Recommendation on MIND

This repository is ready-to-use for designing->training->testing a News Recommendation Model on MIND dataset. It offers comprehensive classes and utilities to load data and train/evaluate the model.

Environment

python=3.8.11
torch==1.9.1

Instruction

# manager will email you the training result, so you should add your gmail and password as follows:
mkdir data/configs
echo "email = '[email protected]'" >> data/configs/email.py
echo "password = 'your password'" >> data/configs/email.py

# train a basic twotower model with CNN news encoder and LSTM user encoder
python twotower.py

# test
python twotower.py -m test

• specify -p=/your/path/to/MIND/root
• specify -d=x if you want to train/test on the x-th gpu
• specify -ws=x if you want to train/test with x gpus
• for more hyper parameters, run python twotower.py --help

How to define your own model

1. design your model in models/ referring to TwoTower.py
2. write a script in ./ referring to twotower.py

Basic Introduction

Preprocess

in utils/MIND.py

• tokenize
• generate cache

Manager

in utils/Manager.py

• handle hyper parameters
• prepare dataloaders
• handle the training/evaluation process

News Encoder

in models/Enocders

• CNN
• RNN
• Multi Head Attention
• Transformer
• Pooling

User Encoder

in models/Enocders

• LSTM
• MHA
• LSTUR
• Pooling