KPRE

This repository is the implementation of KPRE :

Knowledge Graph Attention Network with Path Rotate Encoding for Recommendation

Required packages

The code has been tested running under Python 3.7.13, with the following packages installed (along with their dependencies):

• torch==1.12.1
• numpy==1.21.5
• tqdm==4.64.1
• scikit-learn==0.24.1

Files in the folder

• data/
  • music/
    • ratings_final.txt: user-item interaction graph of Last.FM dataset;
    • kg_final.txt: knowledge graph file;
  • book/ ( the structure of other datasets is similar )
  • movie/
  • restaurant/
• src/: implementations of KPRE.

Download & preprocess data

We have downloaded and processed the music and book datasets. However, for larger datasets (Movies and Restaurants) you will need to download and pre-process them yourself. The datasets are pre-processed using the same way as CKAN. The following are the original files of the datasets for download:

• Music

$ url http://files.grouplens.org/datasets/hetrec2011/hetrec2011-lastfm-2k.zip

• Book

$ url http://www2.informatik.uni-freiburg.de/~cziegler/BX/BX-CSV-Dump.zip

• movie

$ url http://files.grouplens.org/datasets/movielens/ml-20m.zip

• Restaurant

$ url https://github.com/hwwang55/KGNN-LS/raw/master/data/restaurant/Dianping-Food.zip

Run the code

We set a random seed to facilitate users to observe the effect of the model easily. You can reset the random seed by adding parameters this way: --random_flag False

• music

$ cd src
$ python main.py --dataset music --adj_size 8 --n_layer 3 --aim_num 3 --lr 0.03

• book

$ cd src
$ python main.py --dataset book --adj_size 32 --n_layer 4 --aim_num 4 --lr 0.01

• movie

$ cd src
$ python main.py --dataset movie --adj_size 64 --n_layer 2 --aim_num 3 --lr 0.01

• restaurant

$ cd src
$ python main.py --dataset restaurant --adj_size 8 --n_layer 3 --aim_num 3 --lr 0.01