Contextual recommender system for E-commerce applications
适用于电子商务应用的情境推荐系统
Highlights
- •Hybrid Collaborative filtering model is proposed for recommender system.
为推荐系统提出了混合协同过滤模型。 - •It is aware of both context and semantic of user and item textual details.
它知道用户和项目文本详细信息的上下文和语义。 - •User embeddings are prepared using word2vec(w2v).
用户嵌入是使用 word2vec(w2v) 准备的。 - •Items embeddings are generated using Convolutional Neural Network(CNN).
项目嵌入是使用卷积神经网络 (CNN) 生成的。 - •PMF is used as collaborative filtering techniques.
PMF 用作协作过滤技术。 - •The model is primarily proposed for missing rating prediction.
该模型主要用于缺失评级预测。 - •The model can be used as recommendation model.
该模型可用作推荐模型。 - •The proposed model is tested on three real world dataset.
所提出的模型在三个真实世界数据集上进行了测试。
Abstract 抽象
当今的地球村组织、社交应用程序和商业网站提供了有关产品、个人和活动的大量信息。这导致了大量内容,需要有效处理才能获得所需的信息。推荐系统 (RS) 根据用户所需的偏好向用户推荐相关项目。它处理与用户和项目相关的各种信息。但是,RS 存在数据稀疏的问题 。通常,RS 中使用深度学习技术对项目内容进行深入分析,以创建精确的推荐。然而,在项目评论的同时有效处理用户评论仍然是一个开放的研究领域,可以进一步探索。在本文中,提出了一种同时处理用户和项目元数据的混合模型,旨在解决稀疏性问题。为了证明所提出的方法的可行性,我们在三个真实世界的数据集上进行了一系列实验。结果表明,据我们所知,所提出的模型优于其他最先进的方法。
Keywords 关键字
1. Introduction 1. 引言
在当今社会经济竞赛的时代,推荐系统 (RS) 是众多电子商务协会、电影网站、电子图书馆、文章、新闻以及音乐和社交论坛的中心部分。公司正在使用 RS 来推断用户、粉丝和潜在客户的好恶。商业公司和社交网站组织广泛使用他们自己开发的 RS。此类组织根据用户一致的建议产生可观的收入。这些有效的建议起着至关重要的作用,并增加了他们的业务收入。因此,随着 Internet 的广泛使用,RS 变得越来越重要。互联网上很容易获得单个实体的大量数据。RS 专注于用户需求,并提供用户希望拥有或期望被推荐的内容。RS 是从大量数据中建议必要相关信息的内容检索过程。随着时间的推移,商业组织 收集了如此庞大的数据 [1]。RS 在有效推荐和产生丰厚收入方面对用户和组织都有好处。RS 改善了客户对在线产品的决策过程,并允许他们快速 找到相关信息 [2]。 在这方面,文献中提出了不同的推荐方法,例如协同过滤 (CF)、基于内容的过滤 (CBF) 和混合过滤 (HF) [1]、[3]、[4]。 Netflix、Amazon、Facebook、Google News、YouTube、Twitter、LinkedIn 和许多其他组织已经部署了他们的 RS 来定位他们的潜在客户。他们的目标是根据客户的喜好在他们的网站上显示推荐。RS 性能取决于用户与项目之间的历史交互。这种用户与项目的交互可以以显式评分矩阵的形式进行映射。相同的交互也可以隐式地形成在用户评论和项目描述方面。但是,数据很少,因为大多数用户不会一直对他们购买的商品进行评论或评分。因此,此类映射 (隐式和显式) 保持稀疏。为了通过并发处理用户和项目来处理数据的稀疏性,本文提出了一种混合模型。混合模型将用户和项目文本描述的语义和上下文成分与显式评级数据(显式)相结合。这种混合协同过滤模型旨在解决与用户相关的项目数量非常少的数据稀疏问题。
所提出的模型被称为 RS 的上下文混合模型,它有效地处理了评级矩阵数据的稀疏性。所提出的模型将生成的用户和项目的嵌入集成到概率矩阵分解 (PMF) 的相应潜在因子中。这些用户和项目嵌入由 w2v 和 CNN 同时开发。在这里,w2v 捕获文本的语义,而 CNN 提取文本描述的上下文细节。此外,通过将每个特征向量划分为多个部分来捕获特征位置信息,然后将 max-pool 作应用于 max-over pooling 层 [5] 中的每个部分。PMF 用作一种协作过滤技术,当与辅助信息一起孵化时,该技术在稀疏、不平衡和大型数据集中表现出色。它可以预测针对用户的近似 Nearer 评级。预测的评级越接近,可以提供给用户的商品推荐就越准确。实验证据表明,同时对组成实体的语义和上下文的理解可以提高 RS 的性能。本文的主要贡献总结如下:
- •A hybrid contextual model is proposed, which improves the accuracy of rating prediction by merging semantic and contextual details of both the user and item to collaborative filtering.
该文提出一种混合上下文模型,通过将用户和项目的语义和上下文细节合并到协同过滤中,提高了评分预测的准确性。 - •User/item latent models are converged for accurate recommendations based on their corresponding semantic and contextual details. In this regard, a convolutional neural network generates semantically enhanced contextual embedding of users and items.
用户/项目潜在模型会融合在一起,以便根据其相应的语义和上下文详细信息提供准确的推荐。在这方面,卷积神经网络生成用户和项目的语义增强上下文嵌入。 - •The feature locality information is captured by dividing each feature vector into sections and then the max-pool operation is applied to each section in the max-over pooling layer [5].
- •Experimentation shows effective handling of data sparsity and enhanced rating prediction and item recommendation accuracy when compared with other state-of-art models to the best of our knowledge.
2. Background and related work
Table 1. Comparative analysis of models.
Model | Inputs | Task | Metric | Feature vector | Semantics | Context |
---|---|---|---|---|---|---|
PMF [14] | Ratings data | Ratings prediction | RMSE. | No | No | No |
CDL [30] | Ratings data, User documents. | Ratings prediction | mAP, Recall | LDA | No | No |
CDL [31] | Ratings data, User documents. | Ratings prediction | Recall. | LDA | No | No |
ConvMF [44] | Ratings data, User documents. | Ratings prediction | RMSE | TF–IDF | No | User only |
CERMF [45] | Ratings data, User & item documents. | Ratings prediction | RMSE, MAE | TF–IDF | No | User & item |
DRMF [37] | Ratings data, User & item documents | Ratings prediction & Items Rec | RMSE, MAE, Precision, Recall | Glove | No | User & item |
CapsMF [46] | Ratings data, User & item documents | Rating prediction & Items Rec | RMSE, MAE, Precision, Recall | Glove | No | User & item |
CMF-HRS | Ratings data, User & item documents | Ratings prediction & Items Rec | RMSE, MAE, Precision, Recall, F1 score | w2v | User & item | User & item |
3. Contextual matrix factorization-hybrid recommendation system (CMF-HRS)

Fig. 1. Creation of user and item embeddings using w2v.

Fig. 2. Contextual matrix factorization for hybrid recommendation system.
3.1. Mathematical illustration of the proposed model
3.1.1. Embeddings generation
- •: Item cleaned Corpus
- •: User cleaned Corpus
- •N : Total number of Items
- •M : Total number of Users
3.1.2. Probabilistic matrix factorization
- •: User Latent Factor
- •: Item Latent Factor
- •: Concept latent factor
3.2. Optimization

4. Experimentation and results
4.1. Dataset preprocessing
Table 2. Datasets statistical data.
Datasets | Users | Items | Ratings | Range | Density |
---|---|---|---|---|---|
AIV | 5130 | 1685 | 37 126 | [1 – 5] | 0.429% |
AA | 87 271 | 13 209 | 752 937 | [1 – 5] | 0.065% |
Yelp | 12 146 | 27 774 | 408 410 | [1 – 5] | 0.121% |
4.2. Evaluation matrices
- •PMF [14] is a basic collaborative filtering technique in the field of RS that applies random initialization of both user and item latent factors. It only uses user-to-item explicit rating matrix as input to the collaborative filtering technique.
- •ConvMF [44] exploited movie plots and formulated item contextual embeddings from CNN for further integration into PMF. It takes into account both explicit rating matrix and implicit movie plots (item’s content). This approach improved the performance of RS; however, the user’s profile was not made part of the model.
- •CERMF [45] deployed two convolutional neural networks with embedding layers initialized randomly for both users and items. The probabilistic collaborative filtering technique was used by the model for rating prediction.
- •DRMF [37] used a dual regularized deep neural network and simultaneously generated both user and item latent factors. DNN was employed for a deep understanding of auxiliary details and extracted meaningful outcomes.
- •
- •CMF-HRS learned semantics and context details of users and items through integrated function of w2v and convolutional neural network. Subsequently, the learned embeddings are fed to latent factors of the collaborative filtering technique for rating predictions. Moreover, the ratings matrix binary form is used by the model for top-n item recommendation.
4.3. Results and analysis
4.4. Effect of CNN embedding layer dimension d
Table 3. RMSE of CMF-HRS for ‘d’ 100, ‘d’ 200, and ‘d’ 300.
Model | Dataset | Root Mean Square Error (RMSE) | ||
---|---|---|---|---|
Empty Cell | Empty Cell | @d 100 | @d 200 | @d 300 |
AIV | 0.712 | 0.651 | 0.719 | |
CMF-HRS | AA | 1.113 | 0.943 | 1.092 |
Yelp | 0.976 | 0.957 | 0.967 |
Table 4. MAE of CMF-HRS for ‘d’ 100, ‘d’ 200, and ‘d’ 300.
Model | Dataset | Mean Absolute Error (MAE) | ||
---|---|---|---|---|
Empty Cell | Empty Cell | @d 100 | @d 200 | @d 300 |
AIV | 0.709 | 0.685 | 0.694 | |
CMF-HRS | AA | 0.916 | 0.863 | 0.893 |
Yelp | 0.793 | 0.725 | 0.768 |
4.5. Effect of user and item latent factors dimension D
Table 5. RMSE of CMF-HRS for D 25, D 50 and D 100.
Model | Dataset | Root Mean Square Error (RMSE) | ||
---|---|---|---|---|
Empty Cell | Empty Cell | |||
AIV | 0.696 | 0.685 | 0.692 | |
CMF-HRS | AA | 1.102 | 0.9368 | 1.115 |
Yelp | 1.058 | 0.9870 | 1.019 |
Table 6. MAE of CMF-HRS for D 25, D 50 and D 100.
Model | Dataset | Mean Absolute Error (MAE) | ||
---|---|---|---|---|
Empty Cell | Empty Cell | |||
AIV | 0.7051 | 0.6851 | 0.6913 | |
CMF-HRS | AA | 0.9045 | 0.8623 | 0.9168 |
Yelp | 0.7685 | 0.7251 | 0.7831 |
4.6. Item’s recommendation performance

Fig. 4. Performance of CMF-HRS (a): Dataset AIV, (b): Dataset AA, (c): Dataset Yelp..
Table 7. Precision values of CMF-HRS for top n-recommendation.
Yelp | ||||||
---|---|---|---|---|---|---|
Th 3 | Th 2 | Th 3 | Th 2 | Th 3 | Th 2 | |
0.00308 | 0.00327 | 0.00163 | 0.00236 | 0.01023 | 0.01036 | |
0.00337 | 0.00356 | 0.00201 | 0.00319 | 0.01032 | 0.01047 | |
0.00377 | 0.00382 | 0.00293 | 0.00383 | 0.01042 | 0.01054 | |
0.00413 | 0.00429 | 0.00364 | 0.00476 | 0.01056 | 0.01066 | |
0.00441 | 0.00462 | 0.00502 | 0.00589 | 0.01065 | 0.01071 | |
0.00493 | 0.00510 | 0.00593 | 0.00603 | 0.01069 | 0.01079 |
Table 8. Recall values of CMF-HRS for top n-recommendation.
Yelp | ||||||
---|---|---|---|---|---|---|
Th 3 | Th 2 | Th 3 | Th 2 | Th 3 | Th 2 | |
0.399 | 0.427 | 0.379 | 0.415 | 0.228 | 0.329 | |
0.421 | 0.463 | 0.395 | 0.427 | 0.309 | 0.331 | |
0.493 | 0.539 | 0.430 | 0.486 | 0.328 | 0.353 | |
0.521 | 0.616 | 0.461 | 0.513 | 0.365 | 0.384 | |
0.639 | 0.708 | 0.492 | 0.533 | 0.393 | 0.406 | |
0.727 | 0.802 | 0.539 | 0.565 | 0.410 | 0.423 |
Table 9. F1-score of CMF-HRS for top n-recommendation.
Yelp | ||||||
---|---|---|---|---|---|---|
Th 3 | Th 2 | Th 3 | Th 2 | Th 3 | Th 2 | |
0.0061 | 0.0064 | 0.0032 | 0.0046 | 0.0195 | 0.0200 | |
0.0066 | 0.0070 | 0.0039 | 0.0063 | 0.0199 | 0.0202 | |
0.0074 | 0.0075 | 0.0058 | 0.0076 | 0.0201 | 0.0205 | |
0.0082 | 0.0085 | 0.0072 | 0.0094 | 0.0203 | 0.0207 | |
0.0087 | 0.0091 | 0.0099 | 0.0115 | 0.0206 | 0.0209 | |
0.0096 | 0.0097 | 0.0113 | 0.0117 | 0.0208 | 0.0210 |
4.7. Comparative analysis
4.7.1. Model performance on rating prediction
Table 10. Ratings prediction comparison.
Yelp | ||||||
---|---|---|---|---|---|---|
RMSE | MAE | RMSE | MAE | RMSE | MAE | |
1.2080 | 0.9493 | 1.4087 | 1.1805 | 1.2194 | 0.9697 | |
1.0057 | 0.7508 | 1.2461 | 0.9718 | 1.0038 | 0.7863 | |
0.9868 | 0.7259 | 1.2187 | 0.9275 | 0.9988 | 0.7755 | |
0.9722 | 0.7184 | 1.1863 | 0.9088 | 1.0022 | 0.7864 | |
0.9621 | 0.7353 | 1.2091 | 0.9101 | 0.9843 | 0.7697 | |
0.9426 | 0.6982 | 0.1789 | 0.9000 | 0.9865 | 0.7615 | |
0.9593 | 0.7064 | 1.1570 | 0.8878 | – | – | |
0.7010 | 0.6850 | 0.8631 | 0.8632 | 0.9532 | 0.7257 |

Fig. 5. Performance of CMF-HRS on AIV Dataset.

Fig. 6. Performance of CMF-HRS on AA Dataset.

Fig. 7. Performance of CMF-HRS on Yelp Dataset.
4.7.2. Model performance on top-n recommendations
Table 11. Top-n item recommendation performance comparison.
Yelp | ||||||
---|---|---|---|---|---|---|
Precision | Recall | Precision | Recall | Precision | Recall | |
0.00506 | 0.772 | 0.00583 | 0.549 | 0.01067 | 0.421 | |
0.00505 | 0.771 | 0.00583 | 0.546 | 0.01063 | 0.423 | |
0.00510 | 0.780 | 0.00591 | 0.551 | 0.01084 | 0.429 | |
0.00522 | 0.810 | 0.00611 | 0.589 | 0.01064 | 0.432 | |
0.00518 | 0.801 | 0.00588 | 0.550 | 0.01025 | 0.410 | |
0.00519 | 0.802 | 0.00604 | 0.583 | 0.01084 | 0.430 | |
0.00515 | 0.810 | 0.00451 | 0.520 | – | – | |
0.00527 | 0.819 | 0.00615 | 0.586 | 0.01087 | 0.433 |
5. Conclusion and future work
CRediT authorship contribution statement
Declaration of Competing Interest
Acknowledgments
References
- [1]Topicmf: Simultaneously exploiting ratings and reviews for recommendationAAAI, AAAI Press (2014), pp. 2-8
- [2]Recommendation systems: Principles, methods and evaluationEgypt. Inform. J., 16 (2015), pp. 261-273
- [3]Computationally efficient face spoofing detection with motion magnificationIEEE Conference on Computer Vision and Pattern Recognition Workshops (2013), pp. 105-110
- [4]Fusion of fragmentary classifier decisions for affective state recognitionInternational Conference on Multimodal Pattern Recognition of Social Signals in Human-Computer-Interaction, Vol. 7742 (2012), pp. 116-130
- [5]A content-based deep hybrid approach with segmented max-poolingIntelligent Information Processing X (2020), pp. 299-309, 10.1007/978-3-030-46931-328
- [6]A collaborative filtering method based on Artificial Immune NetworkExpert Syst. Appl., 36 (4) (2009), pp. 8324-8332
- [7]Probabilistic memory-based collaborative filteringIEEE Trans. Knowl. Data Eng. (2013)
- [8]P. Massa, Paolo Avesani, Trust-Aware Collaborative Filtering for Recommender Systems Via Sommarive 14 - I-38050 Povo (TN) - Italy.
- [9]X. Su, T.M. Khoshgoftaar, A Survey of Collaborative Filtering Techniques Advances in Artificial Intelligence, Volume 2009, Article ID 421425, 19 pages.
- [10]J.S. Breese, D. Heckerman, C. Kadie, Empirical analysis of predictive algorithms for collaborative filtering, in: Proceedings of the Fourteenth Conference on Uncertainty in Artifical Intelligence, 1998.
- [11]Ratings Meet Reviews, a Combined Approach to Recommend, Computer(2014), pp. 6-10
- [12]Hidden factors and hidden topics: understanding rating dimensions with review textProceedings of the 7th ACM Conference on Recommender Systems, ACM (2013), pp. 165-172
- [13]Improved neighborhood-based collaborative filteringKDD’13 CUP (2007)
- [14]Probabilistic matrix factorizationAdv. NeuralInformation Processing Systems (2008), pp. 1257-1264
- [15]Merging trust in collaborative filtering toalleviate data sparsity and cold startKnowl.-Based Syst., 57 (2014), pp. 57-68
- [16]Bayesian probabilistic matrix factorization using markov chain montecarloICML (2008), pp. 880-887
- [17]Incorporating side information in probabilistic matrix factorization with gaussian processesUAI (2010), pp. 1-9
- [18]Generalized probabilistic matrix factorizations forcollaborative filtering2010 IEEE 10th International Conference on DataMining (ICDM) (2010), pp. 1025-1030
- [19]Eigentaste: A constant time collaborative filtering algorithmInformation Retrieval (2001)
- [20]Robin van Meteren, Maarten van Someren, Using Content-Based Filtering for Recommendation. NetlinQ Group, Gerard Brandtstraat 26-28, 1054 JK, Amsterdam.
- [21]Meta-Prod2Vec: Product embeddings using side-information for recommendationRecSys ’16 Proceedings of the 10th ACM Conference on Recommender Systems (2016), pp. 225-232
- [22]Asela Gunawardana, Christopher Meek, Tied boltzmann machines for cold start recommendations, in: RECSYS’08, pp. 19–26.
- [23]Transfer learning in collaborative filtering for sparsity reductionAAAI’10 (2010)
- [24]Bin Li, Qiang Yang, Xiangyang Xue, Transfer learning for collaborative filtering via a rating-matrix generative model, in: ICML’09, pp. 617–62.
- [25]A multi-view deep learning approach for cross domain user modeling in recommendation systemsProceedings of the 24th International Conference on World Wide Web, International World Wide Web Conferences Steering Committee (2015), pp. 278-288
- [26]Y. Wu, C. DuBois, A.X. Zheng, M. Ester, Collaborative denoising auto-encoders for top-n recommender systems.
- [27]Latent Dirichlet allocationJ. Mach. Learn. Res., 3 (2003), pp. 993-1022
- [28]Efficient estimation ofword representations in vector space(2008)Available online: https://arxiv.org/pdf/1301.3781.pdf (accessed on 4 September 2008)
- [29]Topical word embeddingsProceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, Austin, TX, USA (2015)Available online: https://pdfs.semanticscholar.org/9a0a/f9e48aad89512ce3e24b6a1853ed3d5d9142.pdf (accessed on 4 September 2018)
- [30]Collaborative deep learning for recommender systems(2015)arXiv:1409.2944v2 [cs.LG] 18 Jun 2015
- [31]Collaborative topic modeling for recommending scientific articlesKDD (2011), pp. 448-456
- [32]Deep collaborative filtering via marginalized denoising auto-encoderProceedings of the 24th ACM International on Conference on Information and Knowledge Management, ACM (2015), pp. 811-820
- [33]Improving content-based and hybrid music recommendation using deep learningProceedings of the ACM International Conference on Multimedia, ACM (2014), pp. 627-636
- [34]Lei Zheng, Vahid Noroozi, Philip S. Yu, Joint deep modeling of users and items using reviews for recommendation, in: Proceedings of the Tenth ACM International Conference on Web Search and Data Mining (WSDM’17), New York, NY, USA, ACM, pp. 425–434.
- [35]Enriching non-negative matrix factorization with contextual embeddings for recommender systemsNeurocomputing, 380 (2020), pp. 246-258, 10.1016/j.neucom.2019.09.080
- [36]DST-HRS: A topic driven hybrid recommender system based on deep semanticsComput. Commun., 156 (2020), pp. 183-191
- [37]Dual-regularized matrix factorization with deep neural networks for recommender systemsKnowledge-Based Systems, 145(J. Mach. Learn. Res. 3 2003) (2018), pp. 46-58
- [38]Deep matrix factorization approach for collaborative filtering recommender systemsAppl. Sci., 10 (14) (2020), p. 4926
- [39]Improved recurrent neural networks for session-based recommendationsACM International Conference Proceeding Series, Vol. 15-Septemb (2016), pp. 17-22
- [40]Parallel recurrent neural network architectures for feature rich session-based recommendationsRecSys 2016 - Proceedings of the 10th ACM Conference on Recommender Systems (2016), pp. 241-248
- [41]Deep context-aware recommender system utilizing sequential latent context(2019)arXiv preprint arXiv:1909.03999 (2020)
- [42]J. Wang, H. Xie, O.T.S. Au, D. Zou, F.L. Wang, Attention-based CNN for personalized course recommendations for MOOC learners, in: Proc. Int. Symp. Educ. Technol. (ISET), Bangkok, Thailand, Aug. 2020, pp. 180–184.
- [43]Rating prediction based on merge-CNN and concise attention review miningIEEE Access, Digital Object Identifier (2020)
- [44]Convolutional Matrix Factorization for Document Context-Aware Recommendation(2016)978-1-4503-4035-9
- [45]H. Wu, Z. Zhang, K. Yue, B. Zhang, R. Zhu, Content embedding regularized ma- trix factorization for recommender systems, in: 2017 IEEE 6th International Congress on Big Data, IEEE, June 25-June 30, 2017, Honolulu, Hawaii, USA, 2017, pp. 209–215.
- [46]A novel product recommender system using deep learning based text analysis modelMultimed Tools Appl (2020) (2020), 10.1007/s11042-020-09199-5
- [47]Modeling trust for rating prediction in recommender systemsSIAM Workshop on Machine Learning Methods for Recommender Systems, SIAM (2016), p. 18
- [48]HRS-CE: A hybrid framework to integrate content embeddings in recommender systems for cold start itemsJ. Comput. Sci., 29 (2018), pp. 9-18Cited 19 times (2018)
Cited by (25)
CNNRec: Convolutional Neural Network based recommender systems - A survey
2024, Engineering Applications of Artificial IntelligenceCitation Excerpt :The CNN’s design simplicity makes the model good in learning the context of information provided. Khan et al. (2021a) proposed a contextual hybrid model to handle data sparsity in the rating matrix. The embedding developed by Word2Vec (Mikolov et al., 2013) and a CNN of user and items are integrated into a corresponding latent factor of PMF.
DCARS: Deep context-aware recommendation system based on session latent context
2023, Applied Soft ComputingCitation Excerpt :The importance of the user’s context to increase the quality of suggestions has been studied by researchers [3]. Adding user contexts in the recommendations systems (RSs) extends the applications of CARSs [4]. In comparison to other RSs that apply user past interests, CARSs create appropriate suggestions that are close to the current context of the user.
Recommendation Systems for e-Shopping: Review of Techniques for Retail and Sustainable Marketing
2023, Sustainability SwitzerlandImpact of word embedding models on text analytics in deep learning environment: a review
2023, Artificial Intelligence Review