EARS: Emotion-aware recommender system based on hybrid information fusion

Information Fusion 46 (2019) 141–146

Contents lists available at ScienceDirect

Information Fusion journal homepage: www.elsevier.com/locate/inffus

EARS: Emotion-aware recommender system based on hybrid information fusion

T

⁎

Yongfeng Qiana,b, Yin Zhang ,c, Xiao Mac, Han Yuc, Limei Pengd a

School of Computer Science, China University of Geosciences, Wuhan, China School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China c School of Information and Safety Engineering, Zhongnan University of Economics and Law, Wuhan, China d School of Computer Science and Engineering, Kyungpook National University, South Korea b

A R T I C LE I N FO

A B S T R A C T

Keywords: Hybrid information fusion Emotion-aware intelligent system Recommender systems Matrix factorization

Recommender systems suggest items that users might like according to their explicit and implicit feedback information, such as ratings, reviews, and clicks. However, most recommender systems focus mainly on the relationships between items and the user’s final purchasing behavior while ignoring the user’s emotional changes, which play an essential role in consumption activity. To address the challenge of improving the quality of recommender services, this paper proposes an emotion-aware recommender system based on hybrid information fusion in which three representative types of information are fused to comprehensively analyze the user’s features: user rating data as explicit information, user social network data as implicit information and sentiment from user reviews as emotional information. The experimental results verify that the proposed approach provides a higher prediction rating and significantly increases the recommendation accuracy.

1. Introduction Information overload (information overload) is an increasing problem that cannot be ignored. Recommendation systems were developed to reduce the time that users spend browsing useless information. A recommendation system recommends interests and merchandise to the user by observing the user’s interest characteristics and selection behavior, and even provides personalized services [1]. In recent years, research on recommendation systems has expanded, and the huge amount of accompanying data has brought new challenges for recommendation systems. With the development of big data, cloud computing, mobile computing and other advanced information technologies [2], many types of data are used in recommender systems, and the information that is advantageous to the system must be identified. Therefore, researchers have begun to integrate all types of information and make recommendations based on the fused information. However, the fused information comes from different dimensions, such as personal information, social information, emotional information and whether the activity area of users is explicit [3,4]. Various works have proved that fusion information can significantly improve the availability of the recommendation system [5,6]. The core resources that support the recommendation system are the

⁎

user’s historical behavior data, including explicit feedback and implicit feedback [7]. Most explicit feedback-based collaborative filtering recommendation systems are based on user ratings and trust information to improve the accuracy of recommendation [8]. However, this will miss a lot of implicit feedback data. Implicit feedback data is more common than the additional inputs required for explicit feedback data, and its collection costs are low and do not affect the user experience. To address the shortcomings of explicit feedback data recommendation systems, recommendation systems based on implicit feedback data have been introduced. However, implicit feedback data can express the user’s positive feedback but are less able to express the user’s negative feedback. Solving the problem of the lack of negative feedback is thus very important. At present, recommendation systems based on implicit feedback data face the following three challenges: 1. Sample imbalance. In implicit feedback data, there are usually only positive feedback and no negative feedback. In contrast to explicit feedback data, which directly reflect the tendency of the user’s likes and dislikes, implicit data include “selected” and “non-selected” categories. Although the & quot; selected & quot; may indicate a user & apos; s tendency, the & quot; unselected & quot; cannot directly represent a user & apos; s negative tendency because “not selected” includes not only products in which the user is not really

Corresponding author. E-mail addresses: [email protected] (Y. Qian), [email protected] (Y. Zhang), [email protected] (X. Ma), [email protected] (H. Yu), [email protected] (L. Peng).

https://doi.org/10.1016/j.inffus.2018.06.004 Received 13 February 2018; Received in revised form 26 May 2018; Accepted 17 June 2018 Available online 19 June 2018 1566-2535/ © 2018 Elsevier B.V. All rights reserved.

Information Fusion 46 (2019) 141–146

Y. Qian et al.

user. However, poor scalability and a user data transferring method that is based on filters are problems. To solve these problems, another collaborative filtering method is based on content of the recommendation. This method mainly uses the Bayesian probability model, genetic algorithm and other machine learning methods [15]. With the expansion of this application, collaborative filtering also encounters some problems, which are mainly confined to the following three aspects: sparsity, scalability and synonymy. In general, the collaborative filtering algorithm searches a large group of people and finds a smaller set who are similar to the target user’s preferences. The basic mechanism of collaborative filtering is as follows [16]: (i) based on behavior, preferences and other factors, identify a group of people’s preferences for analysis; (ii) using a similarity measure, select a sub-group that has the greatest similarity with the target users; (iii) weight the user inside the sub-group in the calculation; and (iv) use the resulting preference function to make a recommendation to the target user that is more in line with the user’s preferences. Commonly used methods are metered similarity, cosine similarity, and Pearson correlation coefficient calculation. Collaborative filtering recommendation systems have many advantages. For example, from the establishment of the model to the preparation of the program, the whole process is very clear; the applicable range is relatively large and includes movie, music, social, commodity and many other aspects; and higher degree. However, shortcomings inevitably remain: in the initial system, there is a lack of basic calculation data to relate to the recommended data, i.e., the cold start problem; when the system is being used to recommend items and for user group secondary filtering, other problems must wait.

interested but also products in which the user is interested but has not yet found. This lack of positive examples adds difficulty to the model. 2. Noise. In contrast to an explicit rating, in implicit feedback data, the user can produce a lot of noise due to various misuses. 3. Large scale. The actual scenario of the recommendation model will involve large-scale data, which requires the model to have sufficiently efficient performance and excellent scalability to handle vast amounts of data. In this paper, the model proposed in the above problem is used to model the implicit feedback data directly by converting the recommended task into the probability of maximizing the probability of user selection behavior. The model will be able to “unselect” the information that is fully utilized while avoiding the introduction of negative cases and noise to improve recommendation quality. Specifically, this article makes the following contributions: (1) it develops an effective approach to fuse social information, rating information and emotional information for recommender systems; (2) assisted by the hybrid features from implicit and explicit data, the performance of the proposed recommender system is significantly improved. 2. Related work 2.1. Recommender systems Traditional recommendation systems are difficult to apply directly to implicit feedback data because the data contain only positive feedback from the user and lack negative feedback. In [9], Pan et al. defines this issue as a single type of collaborative filtering (One Class Collaborative Filtering, OCCF), which is generally summarized as an imbalance problem (Unbalanced Class Problem, UCP) [10]. The main way to deal with this problem is to introduce negative feedback in the following three ways:

2.3. Information fusion In recent years, with the development of data mining, hidden information can be extracted from all types of data. Using effective information can improve the solution of related problems [17]. In recommender systems, the use of converged information is also becoming popular. In [18], a Hybrid Multigroup CoClustering recommendation framework integrated information from user-item rating records, user social networks, and item features extracted from the DBpedia knowledge base. The experimental results demonstrated the superior performance of our approach in top-n recommendation in terms of MAP, NDCG, and F1 compared with other clustering-based CF models. Sun et al. [3] developed a probabilistic factor analysis framework, named RMSQ-MF, which has the ability to exploit multi-source information. Cheng et al. [19] integrated personal information, network structure information and social information and then proposed a fusion recommendation framework. Hence, information fusion is available to improve the performance of recommender systems, especially with the assistance of advanced mobile network to acquits various context data [20].

1. The specificity of the application environment and addition of negative artificial rules [11]. For example, an item that is forwarded in the other two but are not forwarded between microblogs can be considered negative samples as users browse them, but there is no forwarding, which may indicate a lack of interest. This method relies on domain knowledge and cannot be extended. 2. Label an unknown sample as a negative sample in a random sample [12]. This method generally assumes that most samples labeled as unknown are negative; thus, most of the randomly sampled samples are negative. 3. The samples labeled as unknown are negative samples, but a lower weight is set. The weight of the sample is negative, reflecting the confidence level of the sample [13]. The disadvantage of this approach is that it cannot guarantee “true”. In fact, these three methods attempt to add negative samples in the experiment but cannot guarantee the authenticity of negative cases. In addition, the introduction of negative samples will increase the burden of training, and the larger scale will affect the efficiency of the experiment. The model proposed in this paper is based on the idea of a probability generation model. By maximizing the observed probability of user feedback to directly select a user to choose the modeling trend, no negative feedback can be trained. This model is therefore applicable to a variety of implicit feedback recommendation scenarios.

3. System design 3.1. System architecture Fig. 1 shows the architecture of the proposed recommender system, which consists of the following four components: 1. Data Collection: In the proposed scheme, three types of raw data are collected for information fusion, i.e., social network data, rating data and user review data. With the development of crowd-sourced review websites such as Yelp,1 these data are easy to access. 2. Information Fusion: Among the collected data, rating data are

2.2. Collaborative filtering Personalized recommender systems include user-based and contentbased collaborative filtering [14]. Among them, the basic idea of userbased collaborative filtering is to analyze a user’s general rating of an item, including an analysis of the user’s historical situation, preferences, and behavior, and then find the best similarity with those of another

1

142

http://www.yelp.com/datasetchallenge/.

Information Fusion 46 (2019) 141–146

Y. Qian et al.

Fig. 1. System architecture of EARS.

the users, we can take some privacy preserving measures, such as anonymity, which protects the user’s privacy by hiding the user’s sensitive information, such as ID, name and so on. In fact, such a processing does not affect the performance of the recommendation system proposed in this paper, because the two processes are vertical. We will not discuss the data privacy preserving here. This work will be carried out in the future.

explicit information that can be processed through various approved approaches for recommender systems, such as CF, SVD, and matrix factorization [21]. However, social network data and user review data are implicit information whose features should be extracted by available approaches. In particular, social information can be extracted by social computing from social network data [22], and emotional information can be extracted from user review data by sentiment analysis [23]. 3. Algorithms and Models: After information fusion, the raw data are preprocessed, and valuable information is extracted for recommendations. Available algorithms and models are used to provide personalized services based on the hybrid information. 4. Recommendation Services: In different scenarios, various recommendation services are appropriate: (i) statistics-based applications that list the most popular items [24]; (ii) knowledge-based applications that discover users’ interests according to historical data [25]; and (iii) prediction-based applications that recognize users’ demands through advanced machine learning and artificial intelligence [26].

3.2. Explicit feedback analysis On the basis of the algorithm mentioned above, we can provide the following model: Assume that the user’s choice behavior is determined by the user’s “selectivity” of the product. Definition 1 assumes that user i and product j in potential product Kdimensional feature space are represented as a set of vectors Ui = (Ui1, Ui2, ⋯, Uik ), Uj = (Uj1, Uj2, ⋯, Ujk ) . User characteristics and product characteristics together determine the behavior of the user’s choice, and thus Aij represents the absolute extent of the tendency of the product user ij: K

It is worth noting that in the process of data collection and information fusion, the data we used may involve user’s privacy information, such as emotional information. In order to protect privacy of

Aij = Ui·Vj =

∑ Uik Vjk k=1

(1)

However, the absolute tendency itself does not fully reflect the user’s 143

Information Fusion 46 (2019) 141–146

Y. Qian et al.

between the start and stop of the collection of time intervals.

choice of behavior, so a comparison is necessary. Assuming that user i’s preference for product j is 20, user i is likely not to select the product if user i’s preference for most products is higher than this value (the average selection is 30) If user i is less likely to choose the majority of products with this value (the average choice is 10), user i is likely to choose this product. Suppose M is the number of products, i δij represents the relative degree of tendency of the user to select product j. When δij is larger, user i is more likely to select product j; by contrast, a δij closer to 0 indicates that user i is less likely to select product j. The relative resistance can be defined to obtain

δij =

Aij Ai

According to the above definitions, if a user’s original microblogging forwarding number is N, the decayed forwarding number can be calculated through Eq. (7). −k (Timerate − Timebegin)

Ndec = N × e

M

1 M

∑h = 1 Aih

(2)

where Ai indicates the user tends to average the M selected products. The user selection probability Prij that i j occurs is determined by the product δij. To facilitate data processing, the sigmoid function δij is normalized to (0,1) to give formula (3.3).

y = (x − Min)/(Max − Min)

δij

Prij = φ (δij ) =

1 + δij

∏

Prij =

(i, j ) ∈ O

∏ (i, j ) ∈ O

δij 1 + δij−1

(4)

The objective of the model is given by a set of O that maximizes the potential features matrix U, V of the posterior probability and the Bayes formula. Assuming U, V with mean 0 and variance with a Gaussian prior distribution, the following probability is obtained:

∏

P (U , V O) ∝

(i, j ) ∈ O

δij 1 + δij−1

∏ N (Uik 0, σ 2) ∏ N (Vik 0, σ 2) i, k

3.4. Sentiment analysis for user reviews (5)

j, k

Considering the emotional offset in user reviews of items, sentiment analysis is necessary to calculate the offset for revising the original rating. Specifically, we calculate the emotional offset through a nonsupervisory learning method based on the sentiment lexicon in the following steps: Assume that there are i words in subsentences sub; we can calculate the sentiment value according to Eqs. (9) and (10).

Taking formula (3.5) and logarithmic inversion, the optimization target equivalent is obtained:

argminU , V L: =

∑

ln(1 + δij−1) + λ ( U

(i, j ) ∈ O

2 F

+ V

2⎞ F⎟

⎠

(6)

1 2δ2

in the parameter is adjusted according to the complexity where λ = of the control parameters to prevent over-fitting. Thus, based on the recommended model, implicit feedback data are fed into the model for an optimization problem, where the matrix U is obtained according to the potential given set of O, V, for the prediction of user i and a given candidate product set {j1 , j2 , …, jk } . According to formula (3.1) obtained by Aij, a recommendation list sorted in descending order is generated.

− 1 if the number of negative words in sub is odd Pol (sub) = ⎧ 1 if the number of negative words in sub is even ⎨ ⎩ (9)

Sen (sub) =

⎛ ∑ SentiWordNet (wi) × Pol (sub) ⎞ ⎝ wi∈sub ⎠

(10)

where Pol(sub) represents the polarity of sub; SentiWordNet(wi) represents the sentiment value of word i calculated according to SentiWordNet 3.0. However, user review lengths differ significantly, and the emotional offset of a longer review is most likely higher than that of a shorter one. To balance the influence that different lengths of review exert on the overall variance, normalization is essential to calculate the overall emotional offset of a user review. Assume that there are j sub sentences in review re that include n sentiment words; we can calculate the emotional offset according to Eq. (11).

3.3. Implicit feedback analysis This part is the highlight of the whole algorithm. The main research concept of this paper is to solve the problem of user interest drift by adding a time forgetting function to the original traditional recommendation algorithm. To more clearly describe the whole process, the following symbols are defined:

• Time : Reference time. The initial time in the experiment with the time span of the collected data is the reference time. • Time : Absolute time. The last time that the microblogging is begin

rate

•

(8)

where x, y respectively represent the raw value and normalized value and Max and Min respectively represent the maximum and minimum values of the experimental data. After the time decay of the score is input into the Pearson correlation coefficient formula, for the data after the time decay, the similarity weight of the time closer to the data should be higher. As in the original Pearson correlation coefficient and the calculation of the user as a user of the project score, in the actual problem of this experiment, the mapping of a microblogging user microblogging is the number of forwarding. The process from the original initiative becomes passive and a problem of the mean. In this paper, this process is simplified to select the sample time after the normalization of the mean.

(3)

The user selects a set of behaviors referred to as the assumed behavior. Each behavior is selected independently, and O gives the set of likelihood probability:

P (O U , V ) =

(7)

As in Sina microblogging, the microblogging forwarding number is affected by a variety of factors. There will be a large gap between the sample situations. To make the calculation of the results converge faster, this paper normalizes different sources of data unified to a reference system. This article uses MATLAB. The time after microblogging and the microblogging forwarded number results are normalized. Linear function transformation is chosen as the normalization method, as presented in Eq. (8).

Aij

=

Length

forwarded is the absolute time. This time may be the last time to be forwarded with a larger time interval, including the possibility that the special data are large, which will easily affect the final score. The user therefore publishes the microblogging time as the actual scoring time. Length: Time span. The time span is the length of the experiment

Offset (re ) =

∑subj ∈ re Sen (subj ) n

(11)

Obviously, the emotional offset of each user review falls in the range of [−1, 1], which is used to revise the original rating. Generally, user reviews contain detailed personalized evaluations 144

Information Fusion 46 (2019) 141–146

Y. Qian et al.

4.3. Experimental results and analysis

Table 1 Comparison Table 1 recommending effectiveness of EARS and alignment algorithms. Data set

Contrast algorithm

MAP

EARS upgrade(%)

F-Measure (%)

EARS upgrade (%)

Watercress

User-based CF Item-based CF EARS

0.131 0.135 0.145

10.7 7.4 –

55.80 56.08 70.35

14.55 14.27 –

4.3.1. Performance evaluation In the experiment, the effectiveness of EARS was validated by comparison with two methods. As shown in Table 1, the advantage of EARS in the MAP evaluation index is relatively large, indicating that EARS improves the recommendation accuracy rate in top-K with outstanding performance. In the F-Measure evaluation index, EARS is also superior to the other two algorithms, proving that EARS indeed increases the accuracy of the recommended list.

Table 2 Effect of potential features and dimensions on MAP and F-Measure (λ = 10−5 ).

MAP F-Measure (%)

K=5

K = 10

K = 20

K = 50

0.131 70.15

0.143 70.25

0.144 70.35

0.143 70.20

Table 3 Effect of n parameters on MAP and F-Measure for (K = 10).

MAP F-Measure (%)

λ = 10−1

λ = 10−3

λ = 10−5

λ=0

0.132 70.15

0.135 70.20

0.143 70.35

0.142 70.30

4.3.2. Effects of experimental parameters To assess the potential features of EARS required to set the dimension K, λ parameters of two parameters, a further experimental study of the influence of different parameters was performed. As shown in Table 2, the potential feature dimension did not affect MAP and F-Measure, and a value of feature dimensions between 10 and 20 is recommended.

about doctors, such as their feelings towards the medical environment and the doctor’s ability and attitude. Therefore, the LDA model is adopted by iDoctor to extract the topics of user latent preference; doctor features can be extracted from user review comments on doctors, which involves matrix factorization to provide more accurate and personalized recommendations. 4. Experiment 4.1. Data set In the experiment, the experimental datasets were obtained from watercress, a leading community site that offers audio book recommendation, city activities, and group exchanges under topic lines featuring a variety of services. This experiment considers a song book or product (Item). All information is collected to obtain the implicit feedback stream sorted by time. Item 383 grabs the user’s information from watercress online, including the historical behavior of their selection of books and music, the two products of interest. The behavior of several books is selected as 373,648, and 346,242 is selected as musical acts. From the tab feature, the following product information is extracted as the product features: song information (including artist information, song type) or book information (including the type of book, author information). A comprehensive evaluation experiment is performed using the F-Measure and average mean accuracy (MAP) as an evaluation index.

As shown in Table 3, the regularization parameter has little effect on MAP and F-Measure but is still effective. When λ is 0, the data potential features in the data matrix remain small. This illustrates that compared with traditional matrix decomposition, EARS provides an advantage by preventing over-fitting because the model optimizes the relative proportions between the selected product and the general product, rather than fitting a value. A regularization parameter value of between 10−3 and 10−5 is recommended.

4.2. Contrast algorithm In this paper, we choose the following two algorithms as a contrast algorithm: 1. User-based collaborative filtering (user-based CF): The algorithm first finds for the target user a set of users with similar interests, finds a user-friendly collection that the target user never heard of, and recommends items to the target user. 2. Item-based collaborative filtering (item-based CF): The algorithm calculates the similarity between items and generates a recommendation list for the user based on the similarity and user behavior history items. 145

Information Fusion 46 (2019) 141–146

Y. Qian et al.

5. Conclusion In this paper, an empirical model based on implicit feedback data is studied experimentally. First, the implicit feedback data stream is explained, and the potential feature implicit feedback recommendation model based on matrix decomposition is then given. The probability of the proposed problem is modeled as an optimization problem by maximizing the probability of the user’s selection behavior. The model is verified by a relevant data set, and the superiority of the model is verified by comparison with other recommended models. A limitation of this paper is that it does not implement online recommendations. In addition, this model only applies to implicit feedback data. However, implicit feedback data and display feedback data may exist at the same time in reality, and how to integrate these two resources to increase the accuracy of recommendations requires further study. The contrast algorithm selection remains inadequate, and for greater contrast, the new algorithm can reflect the performance of EARS. In the future, I will apply the characteristics of online recommendation to this model to create a model for online recommendation. I will also attempt to add a network into a realistic recommendation scene. Many factors that influence the choice of users must be elucidated to achieve more refined modeling.

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16] [17]

Acknowledgments This work was supported by the China National Natural Science Foundation under Grant 61702553 and the Project of Humanities and Social Sciences (17YJCZH252) funded by the China Ministry of Education (MOE).

[18]

[19]

[20]

References

[21]

[1] M. Chen, J. Yang, J. Zhou, Y. Hao, J. Zhang, C.-H. Youn, 5G-smart diabetes: toward personalized diabetes diagnosis with healthcare big data clouds, IEEE Commun. Mag. (2018) 3. [2] M. Chen, Y. Qian, Y. Hao, Y. Li, J. Song, Data-driven computing and caching in 5G networks: architecture and delay analysis, IEEE Wirel. Commun. 25 (1) (2018) 2–8. [3] J. Sun, Y. Xiong, Y. Zhu, J. Liu, C. Guan, H. Xiong, Multi-source information fusion for personalized restaurant recommendation, International ACM SIGIR Conference on Research and Development in Information Retrieval, (2015), pp. 983–986. [4] M. Chen, Y. Hao, L. Hu, K. Huang, V.K. Lau, Green and mobility-aware caching in 5G networks, IEEE Trans. Wireless Commun. 16 (12) (2017) 8347–8361. [5] I. Palomares, F. Browne, P. Davis, Multi-view fuzzy information fusion in collaborative filtering recommender systems: application to the urban resilience domain, Data Knowl. Eng. (2017). [6] H. Zhao, Q. Yao, J. Li, Y. Song, D.L. Lee, Meta-graph based recommendation fusion over heterogeneous information networks, Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM, 2017, pp. 635–644. [7] X. He, L. Liao, H. Zhang, L. Nie, X. Hu, T.-S. Chua, Neural collaborative filtering,

[22]

[23]

[24] [25]

[26]

146

Proceedings of the 26th International Conference on World Wide Web, International World Wide Web Conferences Steering Committee, 2017, pp. 173–182. T. Joachims, A. Swaminathan, T. Schnabel, Unbiased learning-to-rank with biased feedback, Proceedings of the Tenth ACM International Conference on Web Search and Data Mining, ACM, 2017, pp. 781–789. R. Pan, Y. Zhou, B. Cao, N.N. Liu, R. Lukose, M. Scholz, Q. Yang, One-class collaborative filtering, Data Mining, 2008. ICDM’08. Eighth IEEE International Conference on, IEEE, 2008, pp. 502–511. L. Duan, M. Xie, T. Bai, J. Wang, A new support vector data description method for machinery fault diagnosis with unbalanced datasets, Expert Syst. Appl. 64 (2016) 239–246. B. Ning, Q.-L. Han, Z. Zuo, J. Jin, J. Zheng, Collective behaviors of mobile robots beyond the nearest neighbor rules with switching topology, IEEE Trans. Cybern. (2017). J. Lian, F. Zhang, M. Hou, H. Wang, X. Xie, G. Sun, Practical lessons for job recommendations in the cold-start scenario, Proceedings of the Recommender Systems Challenge 2017, ACM, 2017, p. 4. L.M. de Campos, J.M. Fernández-Luna, J.F. Huete, L. Redondo-Expósito, Positive unlabeled learning for building recommender systems in a parliamentary setting, Inf. Sci. 433–434 (2018) 221–232. Y. Zhang, D. Zhang, M.M. Hassan, A. Alamri, L. Peng, Cadre: cloud-assisted drug recommendation service for online pharmacies, Mob. Netw. Appl. 20 (3) (2015) 348–355. M. Chen, Y. Hao, K. Hwang, L. Wang, L. Wang, Disease prediction by machine learning over big data from healthcare communities, IEEE Access 5 (2017) 8869–8879. B. Yang, Y. Lei, J. Liu, W. Li, Social collaborative filtering by trust, IEEE Trans. Pattern Anal. Mach. Intell. 39 (8) (2017) 1633–1647. F. Zhu, Y. Zhang, N. Yorke-Smith, G. Guo, X. Chen, IFUP: Workshop on multi-dimensional information fusion for user modeling and personalization, Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, ACM, 2018, pp. 804–805. S. Huang, J. Ma, P. Cheng, S. Wang, A hybrid multigroup coclustering recommendation framework based on information fusion, ACM Trans. Intell. Syst.Technol. 6 (2) (2015) 1–22. S. Cheng, B. Zhang, G. Zou, M. Huang, Z. Zhang, Friend recommendation in social networks based on multi-source information fusion, Int. J. Mach. Learn. Cybern. (5) (2018) 1–22. M. Chen, Y. Hao, M. Qiu, J. Song, D. Wu, I. Humar, Mobility-aware caching and computation offloading in 5G ultra-dense cellular networks, Sensors 16 (7) (2016) 974. X. Guan, C.-T. Li, Y. Guan, Matrix factorization with rating completion: an enhanced SVD model for collaborative filtering recommender systems, IEEE Access 5 (2017) 27668–27678. Y.-D. Seo, Y.-G. Kim, E. Lee, D.-K. Baik, Personalized recommender system based on friendship strength in social network services, Expert Syst. Appl. 69 (2017) 135–148. C. Musto, M. de Gemmis, G. Semeraro, P. Lops, A multi-criteria recommender system exploiting aspect-based sentiment analysis of users’ reviews, Proceedings of the Eleventh ACM Conference on Recommender Systems, ACM, 2017, pp. 321–325. M. Song, X. Zhao, E. Haihong, Z. Ou, Statistics-based CRM approach via time series segmenting RFM on large scale data, Knowl. Based Syst. 132 (2017) 21–29. J.K. Tarus, Z. Niu, A. Yousif, A hybrid knowledge-based recommender system for elearning based on ontology and sequential pattern mining, Future Gener. Comput. Syst. 72 (2017) 37–48. J. Chen, H. Wang, Z. Yan, et al., Evolutionary heterogeneous clustering for rating prediction based on user collaborative filtering, Swarm Evol. Comput. 38 (2018) 35–41.