About

This repository is a Tensorflow 2 implementation of RippleNet (arXiv):

RippleNet: Propagating User Preferences on the Knowledge Graph for Recommender Systems, Hongwei Wang, Fuzheng Zhang, Jialin Wang, Miao Zhao, Wenjie Li, Xing Xie, Minyi Guo, The 27th ACM International Conference on Information and Knowledge Management (CIKM 2018)

RippleNet is a deep end-to-end model that naturally incorporates the knowledge graph into recommender systems. Ripple Network overcomes the limitations of existing embedding-based and path-based KG-aware recommendation methods by introducing preference propagation, which automatically propagates users' potential preferences and explores their hierarchical interests in the KG.

You can find other implementations below:

• Authors' official Tensorflow 1.x implementation of RippleNet by @hwwang55
• A Tensorflow 2 implementation of RippleNet by @SSSxCCC
• A PyTorch implementation of RippleNet by @qibinc

Environment Setup

Tested using Python 3.7 and Tensorflow 2.2.0. You should setup the following virtual environment and install the required packages:

python3 -m venv venv
source venv/bin/active
pip install -r requirements.txt

Data

Unzip the data.zip file to the project root and your folder structure should look like the following (note that /data folder is gitignored):

.
├── LICENSE
├── README.md
├── data
│   ├── book
│   │   ├── book_ratings.csv
│   │   ├── item_index2entity_id_rehashed.txt
│   │   └── kg_rehashed.txt
│   └── movie
│       ├── item_index2entity_id_rehashed.txt
│       ├── kg_part1_rehashed.txt
│       ├── kg_part2_rehashed.txt
│       └── movie_ratings.dat
├── framework.jpg
├── main.py
├── model
│   ├── __init__.py
│   ├── layers.py
│   ├── model.py
│   └── ripple_net.py
├── preprocess.py
├── requirements.txt
└── tools
    ├── __init__.py
    ├── load_data.py
    └── metrics.py

• model/: implementation of RippleNet
• tools/: data loader and model metrics
• data/book/
  • book_ratings.csv: raw rating file of Book-Crossing dataset
  • item_index2entity_id.txt: the mapping from item indices in the raw rating file to entity IDs in the KG
  • kg.txt: book knowledge graph file
• data/movie/
  • item_index2entity_id.txt: the mapping from item indices in the raw rating file to entity IDs in the KG
  • kg_part1.txt and kg_part2.txt: movie knowledge graph files
  • movie_ratings.dat: raw rating file of MovieLens-1M;

Run

for the movie dataset (for the book dataset, replace movie with book in the commands):

• run python preprocess.py --dataset movie - this will generate two new files kg_final.txt and ratings_final.txt
• run python main.py --dataset movie - this will start the training, create a logs folder, and do a final evaluation

By default, the model is trained using 10 epochs, which takes about 25 minutes on a MacBook Pro (3.1 GHz Dual-Core Intel Core i5 with 8G RAM). One sample evaluation result is as follows:

evaluate model ...
148/148 [==============================] - 13s 88ms/step - loss: 0.3811 - binary_accuracy: 0.8464 - auc: 0.9221 - f1: 0.8486 - precision: 0.8399 - recall: 0.8577
- loss: 0.3810572326183319 - binary_accuracy: 0.8464029431343079 - auc: 0.9220795631408691 - f1: 0.848639190196991 - precision: 0.8399457931518555 - recall: 0.8577455878257751

You can use Tensorboard to check the training results (in real time or after training) by running tensorboard --logdir=logs and then point your browser to http://localhost:6006/