Introduction

This framerwork is meant to make working with the MIND dataset easier and allow benchmarking and comparision of different models.
The base of the project is this repo which was updated to use the latest version of pytroch-lightning.
Fastformer was borrowed from this repo.
Implemented models:

• NRMS

• NRMS+Fastformer : the only description provided in the paper -as far as I could figure out- on how this is supposed to be implemented is this line

  In addition, in the news recommendation task, following (Wu et al., 2019) we use Fastformer in a hierarchical way to first learn news embeddings from news titles and then learn user embeddings from the embeddings of historical clicked news. We use Adam (Bengio and LeCun, 2015) for model optimization.

  ¯\_(ツ)_/¯

• Fastformer+PLM-NR : I still have no clue what they mean by this, I still need to decipher this one out...

Running the models

• Create a conda environment and activate it
• Install the requirements

  pip install -r requirements.txt
  

• Install and login to wandb https://docs.wandb.ai/quickstart
• Run the following command to download and parse the tsv files

  python src/dataset_utils.py --action download --size large
  

• Run the following command to preprocess the news for inference

  python src/dataset_utils.py --action preprocess --size large
  

• Run the following command to train

  python src/main.py
  

• Run the following command to generate the prediction.txt file

  python src/evaluate.py --model paht_to_ckpt