基于张量填补和用户偏好的联合推荐算法

doi:10.11959/j.issn.1000-436x.2019231

通信学报 ›› 2019, Vol. 40 ›› Issue (12): 155-166.doi: 10.11959/j.issn.1000-436x.2019231

• 学术通信 • 上一篇

基于张量填补和用户偏好的联合推荐算法

熊智^1,²,徐恺¹,蔡玲如^1,²,蔡伟鸿^1,²

¹ 汕头大学计算机科学与技术系，广东汕头 515063
² 汕头大学智能制造技术教育部重点实验室，广东汕头 515063

修回日期:2019-10-04 出版日期:2019-12-25 发布日期:2020-01-16
作者简介:熊智（1978- ），男，湖北黄冈人，博士，汕头大学副教授，主要研究方向为服务器集群、大数据应用、机器学习|徐恺（1992- ），男，河南信阳人，汕头大学硕士生，主要研究方向为推荐算法、机器学习、大数据应用|蔡玲如（1979- ），女，广东汕头人，博士，汕头大学副教授，主要研究方向为系统建模与仿真、大数据应用、博弈论|蔡伟鸿（1963- ），男，广东潮州人，博士，汕头大学教授，主要研究方向为云计算、信息安全、网络通信
基金资助:
国家自然科学基金资助项目(61202366);广东省自然科学基金资助项目(2018A030313438);广东省自然科学基金资助项目(2018A030313889);广东省科技计划资基金助项目(2019B010116001)

Joint recommendation algorithm based on tensor completion and user preference

Zhi XIONG^1,²,Kai XU¹,Lingru CAI^1,²,Weihong CAI^1,²

¹ Department of Computer Science and Technology,Shantou University,Shantou 515063,China
² Key Laboratory of Intelligent Manufacturing Technology (Shantou University),Ministry of Education,Shantou University,Shantou 515063,China

Revised:2019-10-04 Online:2019-12-25 Published:2020-01-16
Supported by:
The National Natural Science Foundation of China(61202366);Science and Technology Planning Project of Guangdong Province(2018A030313438);Science and Technology Planning Project of Guangdong Province(2018A030313889);Science and Technology Planning Project of Guangdong Province(2019B010116001)

摘要/Abstract

摘要：

针对现有推荐算法缺乏对用户偏好的考虑，推荐效果不理想的问题，提出了一种联合张量填补和用户偏好的推荐算法。首先，基于评分矩阵和项目所属类别矩阵构建用户–项目–类别的三维张量；然后，利用Frank-Wolfe算法进行迭代计算，填补缺失数据，同时基于张量数据构建用户类别偏好矩阵和评分偏好矩阵；最后，基于填补后的张量以及2个偏好矩阵设计联合推荐算法，并采用差分进化算法进行参数调优。实验结果表明，与一些常用算法和新近提出的算法相比，所提算法的推荐效果优于对比算法，其精度平均提升了1.96%~3.44%，召回率平均提升了1.35%~2.40%。

关键词: 推荐算法, 联合推荐, 张量填补, 用户偏好

Abstract:

Aiming at the problem that existing recommendation algorithms have little regard for user preference,and the recommendation result is not satisfactory,a joint recommendation algorithm based on tensor completion and user preference was proposed.First,a user-item-category 3-dimensional tensor was built based on user-item scoring matrix and item-category matrix.Then,the Frank-Wolfe algorithm was used for iterative calculation to fill in the missing data of the tensor.At the same time,a user category preference matrix and a scoring preference matrix were built based on the 3-dimensional tensor.Finally,a joint recommendation algorithm was designed based on the completed tensor and the two preference matrices,and the differential evolution algorithm was used for parameter tuning.The experimental results show that compared with some typical and newly proposed recommendation algorithms,the proposed algorithm is superior to the compare algorithms,the precision is improved by 1.96% ~ 3.44% on average,and the recall rate is improved by 1.35%~2.40% on average.

Key words: recommendation algorithm, joint recommendation, tensor completion, user preference

中图分类号:

TP391.3

熊智,徐恺,蔡玲如,蔡伟鸿. 基于张量填补和用户偏好的联合推荐算法[J]. 通信学报, 2019, 40(12): 155-166.

Zhi XIONG,Kai XU,Lingru CAI,Weihong CAI. Joint recommendation algorithm based on tensor completion and user preference[J]. Journal on Communications, 2019, 40(12): 155-166.

图/表 13

图1

图2

图3

表1

表2

表3

表4

图4

图5

表5

表6

表7

表8

参考文献 27

[1]	JUGOVAC M , JANNACH D . Interacting with recommenders—overview and research directions[J]. ACM Transactions on Interactive Intelligent Systems, 2017,7(3): 1-46.
[2]	SHAH K , SALUNKE A , DONGARE S ,et al. Recommender systems:an overview of different approaches to recommendations[C]// The 4th International Conference on Innovations in Information,Embedded and Communication Systems. IEEE, 2017.
[3]	FORMOSO V , FERNáNDEZ D , CACHEDA F ,et al. Distributed architecture for k-nearest neighbors recommender systems[J]. World Wide Web, 2015,18(4): 997-1017.
[4]	PONNAM L T , PUNYASAMUDRAM S D , NALLAGULLA S N ,et al. Movie recommender system using item based collaborative filtering technique[C]// International Conference on Emerging Trends in Engineering,Technology and Science. IEEE, 2016.
[5]	吴毅涛, 张兴明, 王兴茂 ,等. 基于用户模糊相似度的协同过滤算法[J]. 通信学报, 2016,37(1): 198-206.
	WU Y T , ZHANG X M , WANG X M ,et al. User fuzzy similari-ty-based collaborative filtering recommendation algorithm[J]. Journal on Communications, 2016,37(1): 198-206.
[6]	KOREN Y , . Factorization meets the neighborhood:A multifaceted collaborative filtering model[C]// The 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2008: 426-434.
[7]	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.
	WU B , LOU Z Z , YE Y D . Co-regularized matrix factorization rec-ommendation algorithm[J]. Journal of Software, 2018,29(9): 2681-2696.
[8]	WANG Q , MAO Z , WANG B ,et al. Knowledge graph embedding:a survey of approaches and applications[J]. IEEE Transactions on Knowledge and Data Engineering, 2017,29(12): 2724-2743.
[9]	黄美灵 . Spark MLlib机器学习:算法、源码及实战详解[M]. 北京: 电子工业出版社, 2016.
	HUANG M L . Spark MLlib machine learning:algorithm,source code and detailed practice[M]. Beijing: Publishing House of Electronics IndustryPress, 2016.
[10]	TAY Y , TUAN L A , HUI S C . Latent relational metric learning via memory-based attention for collaborative ranking[C]// The 27th International World Wide Web Conference. 2018: 729-739.
[11]	ZHANG S , YAO L , SUN A ,et al. DeepRec:an open-source toolkit for deep learning based recommendation[J]. ACM Computing Surveys, 2019,52(1): 1-38.
[12]	KUMAR V , PUJARI A K , SAHU S K ,et al. Collaborative filtering using multiple binary maximum margin matrix factorizations[J]. Information Sciences, 2017,380(C): 1-11.
[13]	XU Y , HAO R , YIN W ,et al. Parallel matrix factorization for low-rank tensor completion[J]. Inverse Problems and Imaging, 2015,9(2): 601-624.
[14]	YU W , ZHANG H , HE X ,et al. Aesthetic-based clothing recommendation[C]// The 27th International World Wide Web Conference. 2018: 649-658.
[15]	CAI Y , ZHANG M , LUO D ,et al. Low-order tensor decompositions for social tagging recommendation[C]// The Forth ACM International Conference on Web Search and Web Data Mining. ACM, 2011: 695-704.
[16]	KOLDA T G , BADER B W . Tensor decompositions and applications[J]. Siam Review, 2009,51(3): 455-500.
[17]	NANOPOULOS A , RAFAILIDIS D , SYMEONIDIS P ,et al. MusicBox:personalized music recommendation based on cubic analysis of social tags[J]. IEEE Transactions on Audio,Speech,and Language Processing, 2010,18(2): 407-412.
[18]	SYMEONIDIS P . ClustHOSVD:Item recommendation by combining semantically enhanced tag clustering with tensor HOSVD[J]. IEEE Transactions on Systems,Man,and Cybernetics:Systems, 2016,46(9): 1240-1251.
[19]	JAIN P , OH S . Provable tensor factorization with missing data[C]// International Conference on Neural Information Processing Systems. NIPS, 2014: 1431-1439.
[20]	KASAI H , MISHRA B . Low-rank tensor completion:a riemannian manifold preconditioning approach[C]// The 33rd International Conference on Machine Learning. 2016: 1012-1021.
[21]	LIU W , WU C , FENG B ,et al. Conditional preference in recommender systems[J]. Expert Systems with Applications, 2015,42(2): 774-788.
[22]	HARPER F M , KONSTAN J A . The movielens datasets:history and context[J]. ACM Transactions on Interactive Intelligent Systems, 2015,5(4): 1-19.
[23]	JAGGI M , . Revisiting frank-wolfe:projection-free sparse convex optimization[C]// The 30th International Conference on Machine Learning. IMLS, 2013: 427-435.
[24]	LIU J , MUSIALSKI P , WONKA P ,et al. Tensor completion for estimating missing values in visual data[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013,35(1): 208-220.
[25]	WIMALAWARNE K , SUGIYAMA M , TOMIOKA R . Multitask learning meets tensor factorization:task imputation via convex optimization[C]// International Conference on Neural Information Processing Systems. NIPS, 2014: 2825-2833.
[26]	GUO X , YAO Q , KWOK J T . Efficient sparse low-rank tensor completion using the Frank-Wolfe algorithm[C]// The 31th AAAI Conference on Artificial Intelligence. 2017: 1948-1954.
[27]	PRIC K , STORN R M , LAMPINEN J A . Differential evolution:a practical approach to global optimization[M]. Berlin: SpringerPress, 2005.

数据集名称	用户数/个	电影数/个	类别/种	评分次/次	稠度
300u	300	4 667	19	66 887	0.251 4%
1200u	1 200	9 253	19	328 135	0.155 5%
2400u	2 400	11 292	19	611 198	0.118 7%
4000u	4 000	13 370	19	1 028 536	0.101 2%

算法	简介	参数
UserCF^[3]	UCF（user collaborative filtering，基于用户的协同过滤），它根据用户之间的相似度和最近邻的评分进行推荐	近邻数量k
ALS^[9]	ALS（alternating least square，交替最小二乘法），它相当于一种用户和项目协同过滤的混合推荐算法	隐语义因子个数r
HOSVD^[17]	HOSVD（high order singular value decomposition，高阶奇异值分解），它将一个张量近似表示成核张量和因子矩阵乘积的形式，并根据近似张量中的填补值进行推荐	核张量大小R
SVD++^[6]	SVD（singular value decomposition，奇异值分解）的一种改进方式，它在考虑了用户偏置和项目偏置的情况下进行推荐	学习率 b、正则化参数u和步长v
Precon^[20]	利用 Tucker 分解的对称结构提出一种基于黎曼流形处理的张量填补方法，并应用于电影评分数据的填补	核张量大小R
LRML^[10]	LRML（latent relational metric learning），采用一个增强的记忆模块，通过记忆块来学习构建用户和项目之间的潜在关系	嵌入层维度 d和切片数量N
TCUP	TCUP（tensor completion and user preference），本文提出的基于张量填补和用户偏好的联合推荐算法	张量秩的限制 τ，偏好权重w_p和w_q

算法	300u	1200u	2400u	4000u
UserCF	k=12	k=19	k=26	k=28
ALS	r=2	r=2	r=2	r=2
HOSVD	R=20	R=45	R=55	R=70
SVD++	b=u=v=0.15	b=u=v=0.1	b=u=v=0.1	b=u=v=0.1
Precon	R=2	R=2	R=2	R=2
LRML	d=100	d=100	d=100	d=100
	N=25	N=50	N=50	N=100
	τ=65	τ=95	τ=120	τ=140
TCUP	w_p=0.49	w_p=0	w_p=0.32	w_p=0
	w_q=0.98	w_q=0.33	w_q=0.51	w_q=0.31

算法	300u		1200u		2400u		4000u
算法	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
UserCF	0.783 2	1.011 7	0.709 9	0.927 4	0.699 4	0.921 3	0.687 0	0.903 3
ALS	0.979 6	1.373 4	0.698 0	0.941 6	0.681 9	0.923 3	0.658 3	0.878 7
HOSVD	0.821 4	1.018 2	0.771 8	0.973 7	0.770 3	0.975 2	0.769 8	0.974 9
SVD++	0.708 5	0.911 0	0.670 2	0.873 7	0.664 3	0.865 7	0.655 0	0.854 5
Precon	0.719 9	0.944 5	0.668 0	0.875 8	0.658 7	0.862 9	0.648 1	0.851 7
TCUP	0.715 6	0.914 0	0.672 3	0.868 5	0.658 6	0.856 2	0.644 6	0.840 4

Top@N	负正率							F1分数							准确率
Top@N	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP
1	0.02%	0.02%	0.02%	0.02%	0.02%	0.02%	0.02%	0.00%	0.04%	0.42%	0.31%	0.00%	1.29%	1.34%	99.62%	99.62%	99.63%	99.63%	99.63%	99.60%	99.63%
2	0.04%	0.04%	0.04%	0.04%	0.04%	0.04%	0.04%	0.07%	0.04%	0.76%	0.29%	0.00%	2.16%	3.10%	99.60%	99.60%	99.61%	99.60%	99.60%	99.58%	99.6%2
5	0.11%	0.11%	0.10%	0.11%	0.11%	0.10%	0.09%	0.16%	0.09%	1.68%	0.43%	0.00%	3.97%	5.75%	99.54%	99.54%	99.55%	99.54%	99.54%	99.52%	99.56%
10	0.21%	0.22%	0.21%	0.21%	0.21%	0.20%	0.19%	0.43%	0.15%	2.64%	0.60%	0.10%	5.70%	8.51%	99.43%	99.43%	99.45%	99.43%	99.43%	99.43%	99.48%
20	0.43%	0.43%	0.42%	0.43%	0.43%	0.41%	0.39%	0.80%	0.13%	3.66%	0.71%	0.13%	6.89%	9.89%	99.22%	99.22%	99.24%	99.22%	99.22%	99.23%	99.29%
50	1.07%	1.08%	1.05%	1.07%	1.07%	1.04%	1.02%	0.76%	0.18%	4.55%	1.01%	0.31%	7.20%	8.79%	98.58%	98.57%	98.64%	98.59%	98.58%	98.63%	98.70%
100	2.15%	2.16%	2.10%	2.14%	2.14%	2.10%	2.06%	0.67%	0.22%	4.24%	1.12%	0.45%	6.42%	7.40%	97.51%	97.50%	97.60%	97.52%	97.52%	97.60%	97.68%

基于张量填补和用户偏好的联合推荐算法

Joint recommendation algorithm based on tensor completion and user preference

在线阅读

PDF下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 13

参考文献 27

相关文章 3

Metrics

推荐阅读 0

[1]	任佳智,田辉,范绍帅,林远卓,聂高峰,李继龙. 基于用户偏好预测的无人机部署和缓存策略[J]. 通信学报, 2020, 41(6): 1-13.
[2]	王兴茂,张兴明,邬江兴. 基于一跳信任模型的协同过滤推荐算法[J]. 通信学报, 2015, 36(6): 193-200.
[3]	孟祥武,史艳翠,王立才,张玉洁. 用户对移动网络服务偏好学习技术综述[J]. 通信学报, 2013, 34(2): 147-155.

Top@N	负正率							F1分数							准确率
Top@N	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP
1	0.01%	0.01%	0.01%	0.01%	0.01%	0.01%	0.01%	0.03%	0.00%	0.36%	0.17%	0.01%	1.20%	1.38%	99.77%	99.77%	99.77%	99.77%	99.77%	99.76%	99.78%
2	0.02%	0.02%	0.02%	0.02%	0.02%	0.02%	0.02%	0.07%	0.00%	0.82%	0.22%	0.01%	2.11%	3.28%	99.76%	99.76%	99.76%	99.76%	99.76%	99.75%	99.77%
5	0.05%	0.05%	0.05%	0.05%	0.05%	0.05%	0.05%	0.11%	0.00%	2.34%	0.35%	0.01%	4.15%	6.02%	99.73%	99.73%	99.74%	99.73%	99.73%	99.73%	99.75%
10	0.11%	0.11%	0.10%	0.11%	0.11%	0.10%	0.09%	0.20%	0.00%	3.69%	0.36%	0.05%	6.17%	8.72%	99.68%	99.67%	99.69%	99.68%	99.68%	99.68%	99.70%
20	0.22%	0.22%	0.21%	0.22%	0.22%	0.20%	0.19%	0.31%	0.09%	4.90%	0.56%	0.06%	8.09%	10.40%	99.57%	99.57%	99.59%	99.57%	99.57%	99.59%	99.61%
50	0.54%	0.54%	0.52%	0.54%	0.54%	0.51%	0.51%	0.44%	0.01%	5.86%	0.74%	0.07%	9.15%	9.70%	99.25%	99.24%	99.29%	99.25%	99.24%	99.30%	99.32%
100	1.08%	1.09%	1.05%	1.08%	1.08%	1.04%	1.03%	0.33%	0.03%	5.47%	1.00%	0.09%	8.41%	8.53%	98.70%	98.70%	98.77%	98.71%	98.70%	98.79%	98.81%

Top@N	负正率							F1分数							准确率
Top@N	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP
1	0.01%	0.01%	0.01%	0.01%	0.01%	0.01%	0.01%	0.29%	0.00%	0.34%	0.06%	0.00%	1.28%	1.45%	99.83%	99.83%	99.83%	99.83%	99.83%	99.82%	99.83%
2	0.02%	0.02%	0.02%	0.02%	0.02%	0.02%	0.01%	0.29%	0.01%	0.81%	0.07%	0.00%	2.28%	3.57%	99.82%	99.82%	99.82%	99.82%	99.82%	99.81%	99.82%
5	0.04%	0.04%	0.04%	0.04%	0.04%	0.04%	0.04%	0.35%	0.01%	2.24%	0.15%	0.00%	4.49%	6.44%	99.79%	99.79%	99.80%	99.79%	99.79%	99.79%	99.81%
10	0.09%	0.09%	0.08%	0.09%	0.09%	0.08%	0.08%	0.40%	0.01%	3.77%	0.31%	0.01%	6.57%	9.21%	99.75%	99.75%	99.76%	99.75%	99.75%	99.75%	99.77%
20	0.18%	0.18%	0.17%	0.18%	0.18%	0.16%	0.16%	0.37%	0.02%	5.21%	0.69%	0.02%	8.43%	11.06%	99.66%	99.66%	99.68%	99.67%	99.66%	99.67%	99.70%
50	0.44%	0.44%	0.43%	0.44%	0.44%	0.42%	0.41%	0.38%	0.03%	6.02%	1.07%	0.03%	9.19%	10.09%	99.39%	99.39%	99.43%	99.41%	99.39%	99.43%	99.47%
100	0.89%	0.89%	0.86%	0.88%	0.89%	0.85%	0.84%	0.31%	0.04%	5.59%	1.16%	0.05%	8.29%	8.68%	98.95%	98.95%	99.01%	98.99%	98.95%	99.02%	99.07%

Top@N	负正率							F1分数							准确率
Top@N	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP	UserCF	ALS	HOSVD	SVD++	Precon	LRML	TCUP
1	0.01%	0.01%	0.01%	0.01%	0.01%	0.01%	0.01%	0.26%	0.00%	0.10%	0.02%	0.00%	1.14%	1.57%	99.85%	99.85%	99.85%	99.85%	99.85%	99.84%	99.85%
2	0.01%	0.01%	0.01%	0.01%	0.01%	0.01%	0.01%	0.30%	0.00%	0.30%	0.03%	0.00%	2.09%	3.59%	99.84%	99.84%	99.84%	99.84%	99.84%	99.84%	99.85%
5	0.04%	0.04%	0.04%	0.04%	0.04%	0.03%	0.03%	0.35%	0.00%	1.27%	0.11%	0.01%	4.19%	6.57%	99.82%	99.82%	99.82%	99.82%	99.82%	99.82%	99.83%
10	0.07%	0.08%	0.07%	0.07%	0.07%	0.07%	0.06%	0.40%	0.01%	2.85%	0.18%	0.02%	6.24%	9.52%	99.78%	99.78%	99.79%	99.78%	99.78%	99.79%	99.80%
20	0.15%	0.15%	0.14%	0.15%	0.15%	0.14%	0.13%	0.41%	0.01%	4.70%	0.27%	0.02%	8.04%	11.25%	99.71%	99.71%	99.72%	99.71%	99.71%	99.72%	99.74%
50	0.37%	0.37%	0.36%	0.37%	0.37%	0.35%	0.35%	0.31%	0.02%	5.86%	0.51%	0.03%	8.87%	10.17%	99.49%	99.48%	99.51%	99.49%	99.49%	99.52%	99.54%
100	0.75%	0.75%	0.73%	0.75%	0.75%	0.72%	0.71%	0.21%	0.02%	5.60%	0.59%	0.05%	8.09%	8.72%	99.11%	99.11%	99.16%	99.12%	99.11%	99.17%	99.19%