基于左归一化图卷积网络的推荐模型

doi:10.11772/j.issn.1001-9081.2023010030

摘要/Abstract

摘要：

图神经网络（GNN）应用于推荐领域之后取得了巨大的成功，通过堆叠多层图神经网络层聚合邻居的信息，使节点可以获取更加广阔的协同信息，可以有效解决推荐系统的数据稀疏问题。目前大部分将图卷积网络应用于推荐的工作都遵循对称归一化的设计，对称归一化加强了冷门商品对于用户嵌入的构建，但是会分配给流行度高的商品很低的权重，导致热门商品对于用户节点嵌入影响微乎其微。针对这一问题，提出一种左归一化图卷积网络模型，模型使用了更加灵活的归一化处理方式，加入衰减因子，并且设计了两种针对各图卷积层的衰减机制，相互配合，大幅提高了推荐的效果。在数据集Alibaba、Amazon-book、Yelp2018上与基准模型LightGCN、NGCF（Neural Graph Collaborative Filtering）、PinSage、Bprmf进行了对比实验，结果表明，与LightGCN相比，所提模型的召回率（recall）分别提高9.9%、20.2%、7.5%，归一化折损累计增益（NDCG）分别提高12.7%、24.1%、8.3%，验证了所提模型的有效性。

关键词: 协同过滤, 图卷积网络, 左归一化, 推荐系统, 层衰减

Abstract:

The application of Graph Neural Network （GNN） in the field of recommendation has achieved great success. By stacking multi-layer graph neural network layers to aggregate the information of neighbors， nodes can obtain broader cooperative information， which can alleviate the problem of data sparsity in the recommendation system. At present， most of the works applying graph convolutional network to recommendation follow the design of symmetric normalization. Symmetric normalization strengthens the construction of user embedding for unpopular items， but it assigns very low weight to popular items， so that popular products have little influence on user node embedding. To solve this problem， a new left-normalized graph convolution network model was proposed. The model uses a more flexible normalization processing method， adding attenuation factors， and designing two attenuation mechanisms for each graph convolution layer， which cooperate with each other to greatly improve the effect of recommendation. Comparative experiments with LightGCN， NGCF （Neural Graph Collaborative Filtering）， PinSage and Bprmf were conducted on data sets Alibaba， Amazon-book， Yelp2018. Experimental results show that compared with LightGCN， the proposed model improves the index recall by 9.9%， 20.2%， 7.5%， and NDCG （Normalized Discounted Cummulative Gain） by 12.7%， 24.1%， and 8.3%， respectively， verifying the effectiveness of the proposed model.

Key words: collaborative filtering, graph convolutional network, left normalization, recommendation system, layer attenuation

中图分类号:

TP391

马汉达, 梁文德. 基于左归一化图卷积网络的推荐模型[J]. 计算机应用, 2023, 43(S2): 111-116.

Handa MA, Wende LIANG. Recommendation model based on left-normalized graph convolution network[J]. Journal of Computer Applications, 2023, 43(S2): 111-116.

图/表 8

参考文献 23

1	赵俊逸，庄福振，敖翔，等.协同过滤推荐系统综述［J］.信息安全学报， 2021， 6（5）： 17-34. 10.19363/J.cnki.cn10-1380/tn.2021.09.02
2	吴宾，娄铮铮，叶阳东.联合正则化的矩阵分解推荐算法［J］.软件学报， 2018， 29（9）： 2681-2696. 10.13328/j.cnki.jos.005274
3	黄立威，江碧涛，吕守业，等.基于深度学习的推荐系统研究综述［J］.计算机学报， 2018， 41（7）： 1619-1647. 10.11897/SP.J.1016.2018.01619
4	CHENG H T， KOC L， HARMSEN J， et al. Wide & deep learning for recommender systems ［C］// Proceedings of the 1st Workshop on Deep Learning For Recommender Systems. New York： ACM， 2016： 7-10. 10.1145/2988450.2988454
5	GUO H， TANG R， YE Y， et al. DeepFM： a factorization-machine based neural network for CTR prediction ［EB/OL］. ［2022-10-03］. . 10.24963/ijcai.2017/239
6	徐冰冰，岑科廷，黄俊杰，等.图卷积神经网络综述［J］.计算机学报， 2020， 43（5）： 755-780. 10.11897/SP.J.1016.2020.00755
7	马帅，刘建伟，左信.图神经网络综述［J］.计算机研究与发展， 2022， 59（1）： 47-80. 10.7544/issn1000-1239.20201055
8	GAO C， WANG X， HE X， et al. Graph neural networks for recommender system ［C］// Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. New York： ACM， 2022： 1623-1625. 10.1145/3488560.3501396
9	MICHELI A. Neural network for graphs： a contextual constructive approach［J］. IEEE Transactions on Neural Networks， 2009， 20（3）： 498-511. 10.1109/tnn.2008.2010350
10	VELIČKOVIĆ P， CUCURULL G， CASANOVA A， et al. Graph attention networks ［EB/OL］. ［2022-10-03］. .
11	HAMILTON W L， YING R， LESKOVEC J. Inductive representation learning on large graphs ［EB/OL］. ［2022-10-03］. . 10.7551/mitpress/11474.003.0014
12	BERG R， KIPF T N， WELLING M. Graph convolutional matrix completion ［EB/OL］. ［2022-10-03］. .
13	WANG X， HE X， WANG M， et al. Neural graph collaborative filtering ［C］// Proceedings of the 42nd International ACM SIGIR Conference on Research and development in Information Retrieval. New York： ACM， 2019： 165-174. 10.1145/3331184.3331267
14	HE X， DENG K， WANG X， et al. LightGCN： Simplifying and powering graph convolution network for recommendation ［C］// Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. New York： ACM， 2020： 639-648. 10.1145/3397271.3401063
15	WANG X， JIN H， ZHANG A， et al. Disentangled graph collaborative filtering ［C］// Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. New York： ACM， 2020： 1001-1010. 10.1145/3397271.3401137
16	HE X， LIAO L， ZHANG H， et al. Neural collaborative filtering ［C］// Proceedings of the 26th International Conference on World Wide Web. New York： ACM， 2017： 173-182. 10.1145/3038912.3052569
17	CHEN J， ZHANG H， HE X， et al. Attentive collaborative filtering： multimedia recommendation with item- and component-level attention ［C］// Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York： ACM， 2017： 335-344. 10.1145/3077136.3080797
18	DEFFERRARD M， BRESSON X， VANDERGHEYNST P. Convolutional neural networks on graphs with fast candidates spectral filtering ［C］// Proceedings of the 30th International Conference on Neural Information Processing Systems. New York： ACM， 2016： 3844-3852.
19	KIPF T N， WELLING M. Semi-supervised classification with graph convolutional networks ［C］// Proceedings of the 5th International Conference on Learning Representations. Piscataway： IEEE， 2017： 14-28.
20	CHEN L， WU L， HONG R， et al. Revisiting graph based collaborative filtering： a linear residual graph convolutional network approach ［C］// Proceedings of the 2020 AAAI Conference on Artificial Intelligence. Menlo Park， CA： AAAI Press， 2020， 34（1）： 27-34. 10.1609/aaai.v34i01.5330
21	GLOROT X， BENGIO Y. Understanding the difficulty of training deep feedforward neural networks ［C］// Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics. New York： JMLR.org， 2010： 249-256.
22	YING R， HE R N， CHEN K F， et al. Graph convolutional neural networks for web-scale recommender systems ［C］// Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York： ACM， 2018： 974-983. 10.1145/3219819.3219890
23	RENDLE S， FREUDENTHALER C， GANTNER Z， et al. Bayesian personalized ranking from implicit feedback ［C］// Proceedings of the Twenty-fifth Conference on Uncertainty in Artificial Intelligence. ［S. l.］： AUAI Press， 2009： 452-461.

数据集	用户数	项目数	交互数	稀疏度
Yelp2018	31 668	38 048	1 561 406	0.001 30
Amazon-book	52 643	91 599	2 984 108	0.000 62
Alibaba	106 042	53 591	907 407	0.000 16

数据集	用户数	项目数	交互数	稀疏度
Yelp2018	31 668	38 048	1 561 406	0.001 30
Amazon-book	52 643	91 599	2 984 108	0.000 62
Alibaba	106 042	53 591	907 407	0.000 16

模型	Yelp2018		Amazon-book		Alibaba
模型	recall	NDCG	recall	NDCG	recall	NDCG
LightGCN^［14］	0.064 9	0.053 0	0.041 1	0.031 5	0.040 5	0.020 5
NGCF^［13］	0.057 9	0.047 7	0.034 4	0.026 3	0.033 2	0.016 6
PinSage^［22］	0.047 1	0.039 3	0.028 2	0.021 9	0.019 6	0.008 5
Bprmf^［23］	0.043 3	0.035 4	0.025 0	0.019 6	0.019 2	0.008 9
本文模型	0.0698	0.0574	0.0494	0.0391	0.0445	0.0231

模型	Yelp2018		Amazon-book		Alibaba
模型	recall	NDCG	recall	NDCG	recall	NDCG
LightGCN^［14］	0.064 9	0.053 0	0.041 1	0.031 5	0.040 5	0.020 5
NGCF^［13］	0.057 9	0.047 7	0.034 4	0.026 3	0.033 2	0.016 6
PinSage^［22］	0.047 1	0.039 3	0.028 2	0.021 9	0.019 6	0.008 5
Bprmf^［23］	0.043 3	0.035 4	0.025 0	0.019 6	0.019 2	0.008 9
本文模型	0.0698	0.0574	0.0494	0.0391	0.0445	0.0231

组合	Yelp2018		Amazon-book		Alibaba
组合	recall	NDCG	recall	NDCG	recall	NDCG
LightGCN-L₁-L	0.063 0	0.051 1	0.041 9	0.032 0	0.040 5	0.020 6
LightGCN-L₁-R	0.058 7	0.047 7	0.033 4	0.025 9	0.031 5	0.015 5
LightGCN-L₁	0.057 3	0.046 5	0.036 1	0.027 5	0.040 6	0.020 5
LightGCN-L	0.061 9	0.050 9	0.038 3	0.029 9	0.039 6	0.020 7
LightGCN-R	0.052 1	0.040 1	0.025 2	0.019 6	0.038 5	0.019 8
LightGCN	0.064 9	0.053 0	0.041 1	0.031 5	0.040 5	0.020 5