lulu0913 / vrkg4rec

This is the PyTorch Implementation for our model VRKG4Rec (WSDM'23)

VRKG4Rec: Virtual Relational Knowledge Graph for Recommendation

This is the PyTorch Implementation for the paper VRKG4Rec (WSDM'23):

Lingyun Lu, Bang Wang, Zizhuo Zhang, Shenghao Liu and Han Xu. VRKG4Rec: Virtual Relational Knowledge Graph for Recommendation.

Introduction

Virtual Relational Knowledge Graph for Recommendation (VRKG4Rec) is a knowledge-aware recommendation framework, which explicitly distinguishes the influence of different relations for item representation learning and design a local weighted smoothing (LWS) mechanism for user and item encoding.

Citation

If you want to use our codes and datasets in your research, please cite:

@inproceedings{10.1145/3539597.3570482,
author = {Lu, Lingyun and Wang, Bang and Zhang, Zizhuo and Liu, Shenghao and Xu, Han},
title = {VRKG4Rec: Virtual Relational Knowledge Graph for Recommendation},
year = {2023},
isbn = {9781450394079},
publisher = {Association for Computing Machinery},
url = {https://doi.org/10.1145/3539597.3570482},
doi = {10.1145/3539597.3570482},
booktitle = {Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining},
series = {WSDM '23}
}

Environment Requirement

The code has been tested running under Python 3.8.0. The required packages are as follows:

• pytorch == 1.10.1
• networkx == 2.5.1
• numpy == 1.22.4
• pandas == 1.4.3
• scikit-learn == 1.1.1
• scipy == 1.7.0
• torch == 1.9.0
• torch-cluster == 1.5.9
• torch-scatter == 2.0.9
• torch-sparse == 0.6.12

Usage

The instruction of commands has been clearly stated in the codes (see the parser function in utils/parser.py).

• Last-fm dataset

python main.py --dataset last-fm --lr 0.0001 --n_virtual 3 --context_hops 2 --n_iter 3

• MovieLens dataset

python main.py --dataset MovieLens --lr 0.0001 --n_virtual 3 --context_hops 2 --n_iter 3

Dataset

We provide three processed datasets: Last-FM and MovieLens.

• You can find the full version of recommendation datasets via Last-FM and MovieLens.
• We follow the previous study to preprocess the datasets.

		Last-FM	MovieLens
User-Item Interaction	#Users	1,872	6,036
	#Items	3,915	2,347
	#Interactions	42,346	753,772
Knowledge Graph	#Entities	9,366	6,729
	#Relations	60	7
	#Triplets	15,518	20,195

vrkg4rec's Issues

数据集

请问有处理数据集的代码或处理完成的数据集吗

请问代码requirement里的环境是对的吗，我跑完的结果和论文里不匹配？

• pytorch == 1.10.1
• networkx == 2.5.1
• numpy == 1.22.4
• pandas == 1.4.3
• scikit-learn == 1.1.1
• scipy == 1.7.0
• torch == 1.9.0
• torch-cluster == 1.5.9
• torch-scatter == 2.0.9
• torch-sparse == 0.6.12

在论文对比实验中KGIN的参数

你好，请问KGIN的实验是如何进行的，我采用了MCCLK论文中固定采样方式8：2分别生成训练集和测试集，我的随机数种子为2023，然后进行训练，其中epoch设置为1000，batch为1024，dim为64，lr为1e-4，test_batch 128这些参数我没有改动，在多轮训练中RKG4Rec在Recall@20中能到达接近30%多的性能，但KGIN我尝试多个参数，在100次的epoch的Recall@20中总是维持在10%-12%之间，甚至会出现性能下降，我很想知道如何进行基线的实验，非常感谢

请教

你好，请教一下对于其他模型的复现，如何使用本论文的评价指标呢，比如我想对比其他模型，该如何得到本论文相同的评价模型呢？发现有些不同模型即使评价指标相同最终结果也不一样

数据

请问有提供数据处理的代码吗

在论文对比实验中KGIN的参数

你好，请问KGIN的实验是如何进行的，我采用了MCCLK论文中固定采样方式8：2分别生成训练集和测试集，我的随机数种子为2023，然后进行训练，其中epoch设置为1000，batch为1024，dim为64，lr为1e-4，test_batch 128这些参数我没有改动，在多轮训练中RKG4Rec在Recall@20中能到达接近30%多的性能，但KGIN我尝试多个参数，在100次的epoch的Recall@20中总是维持在10%-12%之间，甚至会出现性能下降，我很想知道如何进行基线的实验，非常感谢

代码运行结果有问题

你好，我运行了仓库里的代码（没做改动），处理lastfm数据集，在metric@20 (%)的条件下，结果recall=0.0647，但论文里的recall是38.78，这可能是因为什么问题呢？