联合成对排序的物品推荐模型

doi:10.11959/j.issn.1000-436x.2019137

摘要/Abstract

摘要：

现有的推荐模型大多仅从用户角度进行建模，忽略了物品的功能关系对用户购买决策的影响。从用户和物品这2个角度，同时考虑用户-物品之间的交互关系和物品-物品之间的功能关系，提出了联合成对排序的推荐模型。考虑正样本的排名位置和负采样策略直接影响模型收敛速度，构建一种排序感知的学习算法，用于求解所提模型的参数。实验结果表明，与当前主流推荐算法相比，该算法在多个评价指标上具有明显的性能优势。

关键词: 物品推荐, 成对排序, 协同过滤, 隐式反馈, 矩阵分解

Abstract:

Most of existing recommendation models constructed pairwise samples only from a user’s perspective.Nevertheless,they overlooked the functional relationships among items--A key factor that could significantly influence user purchase decision-making process.To this end,a co-pairwise ranking model was proposed,which modeled a user’s preference for a given item as the combination of user-item interactions and item-item complementarity relationships.Considering that the rank position of positive sample and the negative sampler had a direct impact on the rate of convergence,a rank-aware learning algorithm was devised for optimizing the proposed model.Extensive experiments on four real-word datasets are conducted to evaluate of the proposed model.The experimental results demonstrate that the devised algorithm significantly outperforms a series of state-of-the-art recommendation algorithms in terms of multiple evaluation metrics.

Key words: item recommendation, pairwise ranking, collaborative filtering, implicit feedback, matrix factorization

中图分类号:

TP181

吴宾,陈允,孙中川,叶阳东. 联合成对排序的物品推荐模型[J]. 通信学报, 2019, 40(9): 193-206.

Bin WU,Yun CHEN,Zhongchuan SUN,Yangdong YE. Co-pairwise ranking model for item recommendation[J]. Journal on Communications, 2019, 40(9): 193-206.

图/表 11

图1

表1

主要符号及其含义"

符号	含义
$U$	用户集合 $(\| U \| = m)$
$ℐ$	物品集合 $(\| ℐ \| = n)$
$ℐ_{u}$	用户u购买的物品集合
$ℐ_{k}$	与物品k联合购买的物品集合
$T$	用户购买反馈信息集合
$S$	联合购买反馈信息集合
$D_{T}$	用户-物品成对样本集合
$D_{S}$	物品-物品成对样本集合
$d$	潜在特征向量维度
$P \in ℝ^{m \times d}$	用户潜在特征矩阵
$Q \in ℝ^{n \times d}$	物品潜在特征矩阵
$R \in ℝ^{m \times n}$	用户-物品交互矩阵
$C \in ℝ^{n \times n}$	物品-物品互补矩阵

表1

图2

表2

表3

在4个数据集上的实验结果"

数据集	算法		d=50				d=100
数据集	算法	Precision	Recall	MAP	NDCG	Precision	Recall	MAP	NDCG
Health and Personal Care	POP	0.499 3	1.070 1	0.552 9	0.989 5	0.499 3	1.070 1	0.552 9	0.989 5
	BPR	1.211 2	3.268 7	1.328 4	2.408 2	1.268 5	3.413 5	1.439 5	2.581 0
	GBPR	1.336 2	3.439 7	1.563 5	2.703 9	1.446 3	3.746 0	1.703 7	2.970 3
	AOBPR	1.428 8	3.736 1	1.629 0	2.888 2	1.533 7	3.993 4	1.818 9	3.157 1
	WARP	1.607 9	3.994 5	1.861 0	3.208 7	1.671 1	4.158 5	1.979 5	3.363 9
	UPR	1.282 8	3.512 7	1.427 2	2.583 0	1.366 2	3.836 7	1.609 6	2.877 8
	APR	1.591 0	4.407 3	1.948 2	3.399 4	1.723 2	4.798 7	2.148 7	3.730 2
	CPR	1.862 0	4.933 6	2.307 7	3.904 2	1.962 3	5.123 6	2.407 5	4.077 8
Video Games	POP	0.587 7	1.599 8	0.608 7	1.171 3	0.587 7	1.599 8	0.608 7	1.171 3
	BPR	2.445 6	6.649 6	2.594 6	4.859 8	2.479 5	6.626 0	2.575 9	4.879 4
	GBPR	3.337 2	9.106 0	3.881 0	6.898 2	3.522 4	9.614 0	4.164 6	7.357 5
	AOBPR	3.414 2	9.471 9	4.164 9	7.264 2	3.500 2	9.677 8	4.226 7	7.382 2
	WARP	3.532 8	9.743 5	4.202 7	7.399 3	3.581 0	10.02 6	4.355 2	7.622 2
	UPR	2.501 6	6.684 7	2.687 2	4.963 2	2.562 9	7.071 7	2.650 1	5.023 3
	APR	3.543 2	9.753 6	4.179 5	7.407 6	3.706 2	10.31 2	4.307 4	7.677 8
	CPR	3.737 5	10.44 0	4.579 5	7.994 6	3.909 5	11.01 3	4.795 3	8.330 6
Pet Supplies	POP	0.670 2	1.979 4	0.652 3	1.352 3	0.670 2	1.979 4	0.652 3	1.352 3
	BPR	1.647 1	4.862 7	1.836 2	3.485 7	1.678 6	4.954 6	1.862 1	3.531 7
	GBPR	1.732 9	5.069 4	2.040 9	3.756 5	1.925 7	5.648 3	2.298 8	4.214 0
	AOBPR	1.864 3	5.482 4	2.215 7	4.081 8	1.968 6	5.845 6	2.264 0	4.240 6
	WARP	1.971 4	5.847 8	2.388 0	4.345 3	1.992 9	5.822 2	2.453 5	4.437 9
	UPR	1.694 3	5.123 9	1.996 1	3.695 5	1.751 4	5.356 8	1.887 6	3.657 2
	APR	2.101 4	6.453 1	2.691 6	4.854 8	2.264 3	6.914 1	2.777 3	5.071 9
	CPR	2.232 9	6.778 9	2.856 6	5.111 9	2.360 0	7.101 6	2.972 2	5.325 7
Cell Phones and Accessories	POP	0.312 7	0.969 5	0.360 6	0.688 5	0.312 7	0.969 5	0.360 6	0.688 5
	BPR	1.194 9	4.093 2	1.600 0	2.845 2	1.296 9	4.406 8	1.771 8	3.111 4
	GBPR	1.211 7	4.070 2	1.780 2	3.010 7	1.376 5	4.576 3	2.001 2	3.392 9
	AOBPR	1.340 7	4.533 8	1.863 1	3.235 1	1.443 8	4.966 8	2.107 7	3.588 9
	WARP	1.462 8	4.942 2	2.186 1	3.653 2	1.498 7	5.042 7	2.154 5	3.658 5
	UPR	1.333 9	4.531 6	1.762 3	3.154 9	1.397 8	4.757 4	1.928 7	3.378 7
	APR	1.615 3	5.478 4	2.306 3	3.990 0	1.776 7	6.011 8	2.565 7	4.392 3
	CPR	1.780 1	5.968 9	2.495 9	4.307 3	1.811 5	6.091 7	2.714 6	4.543 0

表3

图3

图4

图5

图6

图7

表4

参考文献 27

[1]	BOBADILLA J , ORTEGA F , HERNANDO A ,et al. Recommender systems survey[J]. Knowledge-Based Systems, 2013(46): 109-132.
[2]	LIU Q , CHEN E , XIONG H ,et al. Enhancing collaborative filtering by user interest expansion via personalized ranking[J]. IEEE Transactions on Systems,Man,and Cybernetics,Part B (Cybernetics), 2012,42(1): 218-233.
[3]	何明, 刘伟世, 张江 . 支持推荐非空率的关联规则推荐算法[J]. 通信学报, 2017,38(10): 18-25.
	HE M , LIU S W , ZHANG J . Association rules recommendation algo-rithm supporting recommendation nonempty[J]. Journal on Commu-nications, 2017,38(10): 18-25.
[4]	SMITH B , LINDEN G . Two decades of recommender systems at amazoncom[J]. IEEE Internet Computing, 2017,21(3): 12-18.
[5]	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.
[6]	ADOMAVICIUS G , TUZHILIN A . Toward the next generation of recommender systems:a survey of the state-of-the-art and possible extensions[J]. IEEE Transactions on Knowledge and Data Engineering, 2005,17(16): 734-749.
[7]	PAN W , CHEN L . GBPR:group preference based Bayesian personalized ranking for one-class collaborative filtering[C]// The International Joint Conference on Artificial Intelligence. AAAI Press, 2013: 2691-2697.
[8]	HU Y , KOREN Y , VOLINSKY C . Collaborative filtering for implicit feedback datasets[C]// The Eighth IEEE International Conference on Data Mining. IEEE, 2008: 263-272.
[9]	HE X , ZHANG H , KAN M Y ,et al. Fast matrix factorization for online recommendation with implicit feedback[C]// The International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 2016: 549-558.
[10]	RENDLE S , FREUDENTHALER C , GANTNER Z ,et al. BPR:Bayesian personalized ranking from implicit feedback[C]// The Conference on Uncertainty in Artificial Intelligence. AUAI Press, 2009: 452-461.
[11]	DING J , FENG F , HE X ,et al. An improved sampler for Bayesian personalized ranking by leveraging view data[C]// Companion Proceedings of the Web Conference 2018. ACM, 2018: 13-14.
[12]	SUN Z , WU B , WU Y ,et al. APL:adversarial pairwise learning for recommender systems[J]. Expert Systems with Applications, 2019(118): 573-584.
[13]	ZHAO T , MCAULEY J , KING I . Leveraging social connections to improve personalized ranking for collaborative filtering[C]// The ACM International Conference on Conference on Information and Knowledge Management. ACM, 2014: 261-270.
[14]	LERCHE L , JANNACH D . Using graded implicit feedback for Bayesian personalized ranking[C]// The ACM Conference on Recommender Systems. ACM, 2014: 353-356.
[15]	PARK Y J , TUZHILIN A . The long tail of recommender systems and how to leverage it[C]// The 2008 ACM Conference on Recommender Systems. ACM, 2008: 11-18.
[16]	RENDLE S , FREUDENTHALER C . Improving pairwise learning for item recommendation from implicit feedback[C]// The ACM International Conference on Web Search and Data Mining. ACM, 2014: 273-282.
[17]	HE X , LIAO L , ZHANG H ,et al. Neural collaborative filtering[C]// The International Conference on World Wide Web. ACM, 2017: 173-182.
[18]	CHEN X , WANG P , QIN Z ,et al. HLBPR:a hybrid local Bayesian personal ranking method[C]// The International Conference Companion on World Wide Web. ACM, 2016: 21-22.
[19]	MCAULEY J , PANDEY R , LESKOVEC J . Inferring networks of substitutable and complementary products[C]// The ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2015: 785-794.
[20]	WANG Z , JIANG Z , REN Z ,et al. A path-constrained framework for discriminating substitutable and complementary products in e-commerce[C]// The ACM International Conference on Web Search and Data Mining. ACM, 2018: 619-627.
[21]	OTUNBA R , RUFAI R A , LIN J . MPR:multi-objective pairwise ranking[C]// The ACM Conference on Recommender Systems. ACM, 2017: 170-178.
[22]	WESTON J , BENGIO S , USUNIER N . WSABIE:scaling up to large vocabulary image annotation[C]// The International Joint Conference on Artificial Intelligence. AAAI Press, 2011: 2764-2770.
[23]	HE R , MCAULEY J . VBPR:visual Bayesian personalized ranking from implicit feedback[C]// The AAAI Conference on Artificial Intelligence. AAAI Press, 2016: 144-150.
[24]	WU B , YE Y , CHEN Y . Visual appearance or functional complementarity:which aspect affects your decision making?[J]. Information Sciences, 2019(476): 19-37.
[25]	彭舰, 王屯屯, 陈瑜 ,等. 基于跨平台的在线社交网络用户推荐研究[J]. 通信学报, 2018,39(3): 147-158.
	PENG J , WANG T T , CHEN Y ,et al. User recommendation based on cross-platform online social networks[J]. Journal on Communications, 2018,39(3): 147-158.
[26]	朱敬华, 明骞 . LBSN 中融合信任与不信任关系的兴趣点推荐[J]. 通信学报, 2018,39(7): 157-165.
	ZHU J H , MING Q . POI recommendation by incorporating trust-distrust relationship in LBSN[J]. Journal on Communications, 2018,39(7): 157-165.
[27]	孙光福, 吴乐, 刘淇 ,等. 基于时序行为的协同过滤推荐算法[J]. 软件学报, 2013,24(11): 2721-2733.
	SUN G F , WU L , LIU Q ,et al. Recommendations based on collabora-tive filtering by exploiting sequential behaviors[J]. Journal of Software, 2013,24(11): 2721-2733.

数据集	用户数量	物品数量	历史购买记录	交互关系稠密度	联合购买记录	互补关系稠密度
Health and Personal Care	16 201	75 431	289 965	2.37×10^-4	48 603	8.5×10^-6
Video Games	8 057	26 729	157 494	7.31×10^-4	19 232	2.7×10^-5
Pet Supplies	7 417	33 798	117 385	4.68×10^-4	22 980	2.0×10^-5
Cell Phones and Accessories	9 534	53 479	139 141	2.73×10^-4	28 578	0.9×10^-5

算法	Health and Personal Care	Video Games	Pet Supplies	Cell Phones and Accessories
BPR	00:05:21	00:02:55	00:02:08	00:02:37
AOBPR	04:06:06	00:47:30	00:48:36	01:32:36
WARP	04:50:55	01:26:49	01:38:23	02:08:46
GBPR	01:39:21	00:45:46	00:41:39	00:49:36
UPR	00:06:18	00:04:11	00:03:16	00:03:42
APR	05:18:46	00:56:17	00:58:19	01:46:41
CPR	04:37:01	00:48:44	00:51:42	01:27:53