CoWPiRec

This is the official implementation for Collaborative Word-based Pre-trained Item Representation for Transferable Recommendation.

<div align=center> <img src="CoWPiRec.png" alt="CoWPiRec.png" width="100%" /> </div>

Requirements

python==3.9.7
recbole==1.0.1
torch==1.10.0
cudatoolkit==11.3
transformers==4.21.2

Quick Start

We provide a preprocessed dataset of Scientific dataset in dataset/Scientific. You can download the pretrained model from Google Drive and place it in saved/model. Then You can repreduce our experiment results following the steps below.

Git clone this repository.

git clone https://github.com/ysh-1998/CoWPiRec.git

Get the text-based item embedding with CoWPiRec.

python get_emb.py --gpu_id=0 --dataset Scientific

Train and evaluate on downstream datasets Scientific.

python downstream/finetune.py --gpu_id=0 -d Scientific

Pipeline

The pipeline to repreduce CoWPiRec is shown below.

Datasets

To preprocessing datasets, you should prepare tow raw_data files:

1. user interaction data data_name.csv, format: user_id,item_id,time
2. item metadata meta_data_name.csv, format: item_id,item_text

Place these two files in ./dataset/raw_data/interaction and ./dataset/raw_data/metadata, respectively, then run script below:

python ./dataset/preprocessing/process_dataset.py --dataset dataset_name
python ./dataset/preprocessing/tokenizer.py --dataset dataset_name

The dataset path generated contains:

dataset/
  dataset_name/
    data_name.train.inter
    data_name.dev.inter
    data_name.test.inter
    data_name.text
    data_name.tokenize.json
    data_name.item2index
    data_name.user2index

Word Graph

The word graph is a key component in pre-training of CoWPiRec and is constructed based on co-click items. Obtain the co-click item pairs first.

python ./WordGraph/get_coclick_item_pairs.py --dataset dataset_name

Process co-click item pairs to get co-click word graph.

python ./WordGraph/get_coclick_word_graph.py \
  --dataset dataset_name --num_workers 40 --max_len 64

You can perform multiprocessing by modifing the argument --number_workers. The max length of item text is controlled by the argument --max_len.

Process co-click word graph, filter neighbors based on tf-idf.

python ./WordGraph/get_word_graph.py --dataset dataset_name --topN 30

argument --topN denotes the number of neighbors after filtering.

Pretraining

Pre-train on single GPU.

python pretrain.py --gpu_id=0 -d pretrain_dataset_name

Pre-train with multi GPUs.

CUDA_VISIBLE_DEVICES=0,1,2,3 python ddp_pretrain.py

Downstream

Get item embedding using CoWPiRec.

python get_emb.py --gpu_id=0 --dataset dataset_name \
    --load_pretrain_model --pretrain_model_path saved/xxx.pth

Downstream recommendation.

python downstream/finetune.py --gpu_id=0 -d dataset_name