GroRec:集成社交,移动和大数据技术的以群体为中心的智能系统

所需积分/C币:1 2018-03-30 10:49:34 997KB PDF

GroRec:集成社交,移动和大数据技术的以群体为中心的智能系统,对用户情绪的定量分析,从而改善数据的客观性,基于评分,兴趣和社会关系的模型
IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL 9. NO 5. SEPTEMBER/OCTOBER 2016 GroRec Multidimensional Group Preference Modeling Cyber 、 ysical Social Place User Group Preference Modeling Multi-source Context Data SUB: l0:10 Cafe Shop Adam Cofee (spatiotemporal) Rating revision based on Emotional Offset Reviews Sentiment Analysis Rating revision d the Enchilada Eggs. add green Chile, chicken and you have a feast suitable for the Gods Th Time is definitely a spur spot, but it'smy 号导 avIor faverite. But tbe Front desk competency nt At busy times the staff is visibly overwhelmed. Ive had to go Fig. 3. SUB model hack twice beeause of their mistakes source, heterogeneous and high-dimensional context Group Discovery According to Behavior Similarity data is a highly chalenging task, especially given the sparsity caused by the uneven distribution of user data. Hence, the multidimensional preference 碱,是 modeling module in groRec is designed to compre hensivcly analyze user preferences to providc morc 减 intelligent recommendation services 3.2 System Model and Problem Formulation The proposed GroRec system consists of three modules, i.e group discovery, rating revision and multidimensiona Fig 2 System architecture preference modeling to separately address three underly- ing problems: 1)user classification based on behavioral sim- 3 PROPOSED SYSTEM ARCHITECTURE AND ilarity, 2) quantification of emotional offset in user reviews, SYSTEM MODEL and 3)comprehensive group preference modeling 3.1 Proposed system Architecture The proposed GroRec system is expected to be able to pro- 3.2.1 User Classification Based on Behavioral vide effective objective and accurate recommendations in Similarity Ss. As illustrated in Fig. 2, GroRec consists of the follow- The purpose of group discovery is to classify users based on ing modules behavioral similarity. In a CPSs the available user behavior Group Discovery Based on Behavioral Similarity data consist of four elements, i.e., time, place, user, and behav- To address the issue of the explosion of context data ior. Generally, the time and place information can be merged for dimensional reduction Therefore, we define user behavior plied to decrease the computational complexity s data using a Spacetime-User-Behavior (SUB)model. How- the recommender system. However, conventional ever, the sUB model is merely a data structure; it must be rep group discovery methods are based on the clustering resented using a proper mathematical approach of low-dimensional data and thus are not applicable A tensor is able to represent multidimensional arrays [171 for CPSSs In GroRec, the group discovery module is and thus is a suitable means of representing the sub model designed to rapidly and accurately discover groups which describes three-dimensional data. Hence, the SUB rom high-dimensional data based on the similarit model can be represented by a 3rd-order tensor, as shown of user behavior in Equation 1, where Is, lu and IB represent the vector Rating Revision Based on Emotional Offset: As spaces of spacetime, user and behavior, respectively. Specif- shown in Fig. 1 L, therc can bc a significant offsct ically aii=1, where aijk E AsUB, represents that user 3 between user reviews and ratings, which can lower exhibits behavior k at the spacetime point i, Whereas the accuracy of recommendation results. In GroRec. aijk=0 otherwise rating data are revised in accordance with their emo- tional features, which are extracted from user AsmB∈ RSXIUXIB (1) reviews through sentiment analysis. The rating revi- {0,1ak∈AsB} sion module is the foundation for ensuring the objec tivity of the recommender, thereby improving the Furthermore, we apply Tucker decomposition, which is a accuracy of the recommendation results method of high-order principal component analysis, for fur- Multidimensional Group Preference Modeling: In ther dimensional reduction. As shown in Equation(2), the a CPSS, extracting user preferences from multi- initial tensor AE RiX. XIN, is decomposed into a core ZHANG: GROREC: A GROUP-CENTRIC INTELLIGENT RECOMMENDER SYSTEM INTEGRATING SOCIAL, MOBILE AND BIG DATA tensor,i.e,BE RRX.XRN, and a series of matrix multiplica- 3. 2.3 Comprehensive Group Preference Modeling tion patterns, i. e, U(m)=un, .., gl]E RInxRn, through Prcfcrancc modeling is the most important issuc for a rcc- Tucker decomposition [18] ommender system In GroRec, we propose a comprehensive approach to extracting group preferences from ratings, interests and social relationships. Considering the high A rN=l (2) dimensionality of the context data available in a CpSS matrix factorization (MF) is a suitable mcans of mapping =B×1U(1)… C/( (N) the complex relationships between users and items into a low-dimensional space of latent factors. Through Mf, users In particular, if r< RN, then the approximate tensor of and items are mapped to a low-dimensional latent factor the sub tensor, which is a high-order compression of the ini- space that explains the user ratings of the items, and the tial tensor can be calculated using equations (1) and(2)as uscr-itcm rating matrix is regarded as the product of the =B×1U(5×2U10x3Um users and items, as presented in Equation( 8). Here, k repre- ()sents the number of selected latent factors and P and Q rep Through Tucker decomposition, the volume and resent the weights of each user and item, respectively, with dimensionality of the user behavior data are reduced but espect to each characteristic in the latent factor space, more importantly, the sparsity is alleviated to some extent which are the results of factorizing the rating matrix R and because the approximate tensor is a densified approxima are used for rating prediction. Notably, some latent factors tion of the initial tenso can typically be ignored; therefore, k:<. ×7 Pm×k米Q 3.2.2 Quantification of emotiona/ Offset in User Reviews Generally, Stochastic Gradient Descent (SGD)[19] is Considering that user reviews can include a significant often applied to minimize the loss function, which is pre- emotional offset, sentiment analysis is an important means sented in Equation(9), for the calculation of P and Q. In this of calculating this offset to revise the original ratings function, A(PlF+Qf) is a bias unit to avoid over-fitting, Suppose that there are i words in subsentences sub; then, A is the regularization Parameter, and p? and QF are we can calculate the sentiment value using as Frobenius norms. Specifically, the minimization of Equa Pol(sby 1 odd-numbered negative words tion 9) is equivalent to the parameter training of the partial (4)derivatives presented in Equation(10) 1 even-numbered negative words ∑(Rn-P2Q)2+P+|Q) Scn(sb)=Po)(∑ Sentiwordneti(a)(⑤) R ∈tra2m U;∈sub dp=-2 2Ru ctrain ( Rui- PuQ1)Q+ 2AP Here, Pol(sub) represents the polarity of sa ar a/ →Rn;∈ train (Rui-PuQi)P+2xQ (10) SentiWordNet(w) represents the sentiment value of word as calculated according to SentiWordNet 3.0 Because of the good performance of MF, many research However, user reviews can significantly differ in length, ers have attempted to extend this model by incorporating and the emotional offset of a longer review is likely to be information othcr than rating data, such as social nctwork higher than that of a shorter one. To balance the influence of data [20] and even multidimensional context data [21l,to different review lengths on the overall variance normaliza- provide more accurate recommendations. In groRec, we tion is essential when calculating the overall emotional off propose a comprehensive preference model based on multi set of user reviews dimensional mf Suppose that there are j subsentences in review re, including n sentiment words; then, we can calculate the 4 PROTOTYPE IMPLEMENTATION AND DETAILED emotional offset using as METHODOLOGY b;∈Te Sen(subi 4.1 Group Discovery Based on Behavioral Similarity Offset(re) In a cpss, the available user behavior data are four-dimen Obviously, the emotional offset of each user review falls sional, including time, place, user and behavior informa- tion. In GroRec, we propose to represent user behavior data in the range of [-1, 1; this fact is used to revise the original in tensor form, analyze the behavioral similarity of users via rating by means of Equation (7), where R(re) is the rating Tucker decomposition, and discover groups via clustering corresponding to review re and p is a value that is deter- Fig. 4 illustrates the flowchart of the group discovery pro- mined through training and is used to adjust the relative cess in groRec. Specifically, group discovery in grorec con weights of the original rating and the emotional offset sists of the following steps Rrev(re)=pR(re)+(1-pOffset(re) (7) 1) Tensor-based Representation of User Behavior: As introduced in Section 3.2.1, the SUb model is used to 1.http://sentiwordnet.isticcnr.it/ represent the user behavior data in the form of the IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL 9. NO 5. SEPTEMBER/OCTOBER 20 Behaviors Tensor Tucker Approximate Clustering Decomposition Te ensor B AsuB B SUB U Fig. 4. Group discovery flowchart tensor AsuB, the structure of which is schematically 1) Text Preprocessing: The reviews, which predomi- illustrated in Fig. 5 nantly consist of text, must be preprocessed, including 2) Behavioral Similarity Analysis via Tucker Decom- word segmentation, morphological normalization, position: As shown in Fig. 6, the approximate tensor and the removal of stopwords and punctuation AsuB can be calculated via Tucker decomposition e) Emotional Offset Calculation: Using Eq ua According to the definition of the approximate ten tions(),(5)and(6), the emotional offset of each sor, it is known that Asug is an approximate repre- review can be calculated scntation of user behavior, whcrcas aC AsUB 3) Rating revision: Based on the corresponding calcu describes the probability of a specific user behavior ated emotional offset, each rating is revised using which can be used to analyze behavioral similarit Equation (7) Group Discovery via Clustering: In Fig. 7, the struc- ture of the approximate tensor AsuB is illustrated in 4.3 Multidimensional Group Preference Modeling ree-dimensional coordinates. This approximate In the proposcd GroRcc system, a comprchcnsivc group tensor can be used to identify groups via clustering preference model is built that integrates ratings, interests algorithms. In the proposed scheme, the approxi- and social relationships. Fig 8 presents the flowchart of the mate tensor AsuB is clustered using the k-Nearest preference modeling process in GroRec, which consists of Neighbors(KNN)algorithm [22] ithm [22] after normalization. group-centric data fusion and group preference modeling 4.2 Rating Revision Based on Emotional Offset 4.3. 1 Group-Centric Data Fusion In GroRec, the emotional offset is calculated through senti- Through data fusion, the context of each member is merged ment analysis and used to revise the corresponding rating into a group-centric context. Considering the different data to improve the objectivity of user evaluations. Specifi- weights of each member in a group, linear superposition is cally, we calculate the emotional offset using a non-supervi- not suitable for group-centric data fusion. As shown in sory learning method based on a sentiment lexicon as follows: Figs. 5 and 6, it is obvious that the elements in the approxi- mate tensor a' are different from the elements in the initial Fig 5. Tensor-based representation of user behavior data Fig 6 Behavioral similarity analysis via tucker decomposition ZHANG: GROREC: A GROUP-CENTRIC INTELLIGENT RECOMMENDER SYSTEM INTEGRATING SOCIAL, MOBILE AND BIG DATA 400 T 300 Group-centered Context Fusion 200 来米 Hike 钟钟钟 100 Social Network 30 Emotional ofset Social Featire Fig. 7. Group discovery via clustering M U tensor A. In accordance with the significance of the approxi mate tensor, an element in the approximate tensor represents the prevalence of the corresponding behavior in a group Multidimensional Group Preference Model Moreover, as shown in Fig. 7, group discovery is performed ased on Matrix Factorization based on the approximate tensor. Therefore, the weight of each behavior can be calculated using Equation (11), which is the foundation of group-centric data fusion WO-S Q2∈ ∈gr0up Recommendation Here, ab is the element corresponding to behavior b in the approximate tensor A and 2iEgroum ai is the sum of all clc Fig. 8. Flowchart of multidimensional group preference modeling ments included in a group. Hence, through group-centric data fusion, the behavior data of all members are integrated (Rai-paq Ro∈ train the form of individual data. In other words the group as a whole is treated as an individual user ∑(G-P242)2 4.3.2 Group Preference Modeling The purpose of group preference modeling is to extract +B∑(I-Q42) Bo∈ Train preferences from ratings, interests and social relationships, each of which must be modeled separately +A(|P|F+‖QHF+‖A‖ Rating-based Matrix Factorization (RMF): As described in Section 3.2.2, the rating data are revised 0.J P ∑(Rn-P)Q to obtain more objective user evaluations of items The rmf process is a basic mf procedure, which can be applied using Equations(9)and (10) ∑a(G-PA2)A+2AP Interest-based matrix Factorization (IMF): To inte Rm;∈tain grate user interests into the preference model, the J enerally accepted approach for discovering inter ∑(R-PQ)P ests is used, namely, latent dirichlet Allocation R∈trai (LDA)[23]. Specifically, the distributions of interests and items associated with a group are extracted 2B∑a(I-Q4)A+2)Q B;∈ train through LDa, i. e. the interest matrix G and the item matrix I are calculated. Based on the extracted distri- butions of interests and items, IMF is performed 4=-2a ∑(G-P4)P R using the loss function presented in Equation(12) and the partial derivatives presented in Equa- 2B∑a(I-Q4)Q+2入A tion(13). Here, a and B are the weights of the user interests and item features, respectively and a is the latent-topic-mapping matrix of the latent-factor-rat-. Social-based Matrix Factorization(SMF): Various ing matrix 1g recommender systems exploit social network 792 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL 9. NO 5. SEPTEMBER/OCTOBER 2016 Data Channel 1U Control Channel Network switch ∑(R-PQ Roai∈tran 4U Management Node (G-P94 4U Management Node-Admin cluster Roi∈ train 1U KVM Management console B∑(=QA Roi∈ train 2U DataNode 2U DataNode +y∑sm(,)P-Pl 2U DataNode Worker cluster 2U DataNode 2U DataNode +X(|P|F+‖QHF++‖4F) 2U DataNode 2U DataNode Expansion slot 0.J Expansion slot ap ∑(Bn-P?)Q Expansion slot Rn∈ trcin Expansion slot Expansion slot 2a∑a(-P4)A xpansion slot RG∈tran Expansion slot +y2smn(g.9)P-9-+2入P Fig 9. Inspur in-cloud smartData appliance 0.J (18) records to alleviate sparsity and improve accu- Rn-PQ1)P+2入Q Fag∈trit racy. In the proposcd GroRec system, a friend a/ group d is defined based on the friendships of -2∑(G-PA)P each member in group g. Then, the similarity Ba∈ train between g and g can be calculated in terms of the Pearson product-Moment Correlation Coefficients 26∑a(-:4)Q+2入A Ra:∈lrut (PPMCCS) for ratings and interests as presented in Equation (14), where Prat is the rating-based PPMCC of g and g, Pint is the interest-based 5 EXPERIMENTS AND EVALUATIONS PPMCC of g and g, and a is an adjustment factor for the weights of Prat and pint 5.1 Experimental Data and Evaluation Standards The experimental data used in this paper were obtained Sim(g, g)=pRat(9, g)+(1-a)Pint (, ).(14) from an open data set provided by the crowd-sourced review website Yelp. In these experiments, approximately Based on the similarity of the friend group, 80 percent of these data were randomly selected for train social factors can be integrated into the group ing, and the remaining data were used to verify the perfor preference model. Specifically, SMF is performed mance of the proposed system. The experimental data set using the loss function presented in Equation (15) included information about local businesses in 10 cities and the partial derivatives presented in Equa- across four countrie Specifically, the data set included 2.2 tion(16) M reviews and 591 K tips provided by 552 K users for 77K businesses and involved a social network of 552 K users, for total of 3.5 led Rni∈ train The hardware environment used in our experiments was +y∑sim(g,9川P。-Pp1 (15) an Inspur In. Cloud SmartData Appliance(SDA)provided by the embedded and pervasive computing lab at huaz- hong University of Science and Technology. As illustrated +A(P|F+‖Q|F) in Fig. 9, this SDa consists of two main clusters: 1)an admin iding 64 CPU 256 GB of RAM and 3.6 TB of storage, and 2)a worker cluster with 0.J seven nodes, providing 84 CPU cores, 336 GB of RAM and P 2∑(BRm-P2Q2)Q Foa∈ train 252 TB of storage. enerally, the root mean Square Error(rmse) and the +y>sim(9,d)P-g-P,+ 2AP; (16) Mean Absolute Error (MAE)are the most important indica- tors for the evaluation of recommender systems. Moreover a 2∑(Bn-PQ)P+2A2 the Recall, precision and Fl Score are the typical indicators Roi∈ train used to evaluate clustering-based recommendations [24] Therefore, we evaluated the performance of the group dis Based on the RME, IMF and SMF procedures, the loss covery module in terms of the Recall, Precision and F1 Score function and partial derivatives for group preference modeling based on mf can be derived as 2.http://www.yelp.com/dataset_challenge/ ZHANG: GROREC: A GROUP-CENTRIC INTELLIGENT RECOMMENDER SYSTEM INTEGRATING SOCIAL, MOBILE AND BIG DATA When K-5, the performance TABLE 1 of clustering is considerable Parameters to be determined Parameter Significance Optimal value learning rat 0.001 0.8 regularization parameter 0.1 number of iterations 100 number of latent factors 15 0.6 1g 07 number of latent topics 30 0.4 weight of user interests 0.1 weight of item fcatures 0.1 ■ Recall 0. -d-Precision 5.3 Evaluation of the recommendations provided ●=F1 by GroRec Using equation(22), the rmse between the predicted rat 3 6 8 ings and the actual ratings can be calculated. A smaller K RMSE indicates a better recommendation performance. In Fig 10. Evaluation of group discovery performance this equation, n is the number of records in the test data set p: represents a predicted rating calculated by groRec, and ri and evaluated the recommendation performance of GroRec represents the corresponding actual rating from the data set using the rmse 5.2 Evaluation of Group Discovery Performance RASF=7∑(m2-n (22) As stated in Section 4.1, in GroRec, group discovery based on behavioral similarity is performed using the knn clus The experiment reported here was designed to compare tering algorithm. Because of its simplicity and suitability, our proposed GroRec system with typical systems for item is often used to cluster low-dimensional data. In the based CF [251, user-based CF [26] and MF-based recommen- proposed group discovery scheme, the volume and dations [19]. As described in Section 4.3, the parameters dimensionality of the context data are effectively com- summarized in table 1 should be determined before the pressed and reduced, allowing these data to be processed comparison. using the knn algorithm. To simplify the evaluation of In Table 1, n, A and N are the three basic SGD parameters the proposed scheme, we selected approximately 6,000 for ME. Through simple substitution, the results can be behavior data corresponding to 57 users at similar times found to be satisfactory when n=0.001,A=0.1,and and places from the experimental data. The Recall, Preci- N-100. The other optimal parameters were determined as sion and f1 Score,which were calculated using Equa- follows.p, K, a and B are used only in HMF, and we deter tions(19),(20) and (21), respectively, were selected as the mined the optimal value for each of these parameters by fix evaluation indicators for this experiment. In these equa- ing the other three tions,TP represents the number of true positives, TN rep- resents the number of true negatives, and Fp represents 1) M: The determination of M is essential for all mf the number of false positives based recommendations based on the basic mf loss function presented in [191, we evaluated the rmse TP for M=5, 10,., 30, 35]. As shown in Fig. 11a, Recall= TP+TM when M=15, the rmse reaches its minimum. Therefore, the number of latent factors in the evalua tion was determined to be 15 TP ecsd (20) 2) p: In the simulation, we set K-20,a=0.1 and TP+ FP B=0. 1 and attempted to minimize the rmse with p=0,0.1.., 0.9, 1. Here, p=0 indicates that only 2· Precision· Recall the emotional offset is used for rating prediction, F1 Precision+ Recall whereas p= 1 means that only the original rating used. Fig 11b shows that when p=0.7, the RMse of Fig. 10 illustrates the evaluation of the proposed kNN the mf calculated based on the emotional-oftset based group discovery procedure for several values of K.It revised ratings (Revised-MF, green line in Fig. 11b is seen that when K is small, FP is small but FN is relatively reaches a minimum. The rmse of the basic MF(MF large; therefore, the Recall is low and the Precision is high red line in Fig. 11b) is also shown as a constant value, By contrast, when K is large, FPis large but FN is relatively corresponding to a special case of the revised MF small, resulting in a high recall and a low precision. When with p= l. The findings indicate that the weight of K=5, the Recall, Precision and F1 Score are all high. There- the emotional offset should be smaller than that of fore, the experiment confirms that the proposed group dis the original rating, because the original rating is usu- covery scheme is suitable for discovering groups based on ally submitted after some deliberation and is reason- the behavioral similarity among users ably representative of the users intent IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL 9. NO 5. SEPTEMBER/OCTOBER 20 127 I-MF 28 4-Revised-MFl 1.27 When M=15. the RMSE is minimum 1.26 1.25 1.24 When p=0.7, the 1.24 RMSE is minimum 123 101520253035 020.40.60.8 M (a)Determination of M b)Determination of p 127 4-|MF When a=0.1 and B=0.1. the 1.24 RMSE is minimum When K=30. the RMSE is minimum 1 0203040506070 (c)Determination of K (a)Determination of a and B Fig. 11. Determination of the optimal parameters 3) K: In the simulation, we fixed p=0.7 and set 4) a and B: Based on the parameter determination a=0. 1 and B=0. 1. Fig. 11c presents the experi rosults presented above, the final paramctcrs to bc mental results for the rmse of the imf blue line in determined were a and B. We attempted to opti- Fig. 11c) as calculated for K=[10, 20,, 60, 70, mize a, BE[0, 1 through a grid search; as shown revealing that the rMse reaches its minimum when in Fig. 11d, the RMse reaches a minimum when K=30. If K is too small, the topic distributions of a=0.1 and B=0.1. These results indicate that the the user interests and item features cannot be suffi weights of the user interests and item features ciently represented in the MF. Conversely, when K hich are used to revise the basic mf to achieve is too large, the degree of fusion of the user inter more accurate recommendations should be rather ests, item features and latent factors is too low to sma all achieve a small rMse Using the determined optimal parameters, we evaluated the performance of the proposed GroRec system In Fig. 12, it is obvious that the rmse of grorec is lower than that of 1.6 MF, which indicates that groRec is able to provide more accurate recommendations. Furthermore, considering the similarity between the number of latent factors in MF (i.e M) and the number of neighborhoods in CF, which is one of the most common techniques used in recommender sys tems, it is reasonable to include the results for item -based CF and user-based CF in the comparison, with the finding that the rmse of conventional CF-based recommendation is higher than that of MF-based recommendation 0.6 6 CONCLUSIO Item CE At present, CPSSs are evolving rapidly and becoming MF widely accepted. However, the theoretical foundations of GroRec recommender systems are immature, presenting great challenges in improving the sufficiency, objectivity and accuracy of such systems in CPSSs. To address this need, we propose a group-centric intelligent recom- Fig 12. Comparison between GroRec and conventional recommender mender system named GroRec, which integrates social systems based on CF and ME. ld big data technologies to provide effective, ZHANG: GROREC: A GROUP-CENTRIC INTELLIGENT RECOMMENDER SYSTEM INTEGRATING SOCIAL, MOBILE AND BIG DATA objective and accurate recommendation services in [171 J. Liu, P. Musialski, P Wonka, and j Ye, Tensor completion for CPSSs. Specifically, a group-centric approach based on estimating missing values in visual data, IEEE Trans. Pattern Anal. Mach. Intell., vol. 35, no. 1, PP. 208-220, Jan 2013 the behavioral similarity among users is proposed to [18] I. V. Oseledets, "Tensor-train decomposition, "SIAM J. Scientific decrease the complexity of conventional individual-cen- Comput.,vol.33,no.5Pp.2295-2317,2011 tric recommender systems, a method of quantifying emo [19] R. Gemulla, E Nijkamp, P.J. Haas, and Y. Sismanis, " Large-scale tional offset based on sentiment analysis is proposed to matrix factorization with distributed stochastic gradient descent, in Proc. 17th ACM SIGKDD Int. Conf Knowl. Discovery Data Min revise user ratings for improved objectivity of rating data;and a comprehensive approach based on three [20] M Jamali and M Ester,"A matrix factorization technique with aspects, i.e. ratings, interests and social relationships, is trust propagation for recommendation in social networks, in 135142 proposed to comprehensively extract group preferences. [211 L. Baltrunas, B Ludwig, and F Ricci, " Matrix factorization techni- Furthermore, experiments verify that the performance of ues for context aware recommendation, in Proc 5th ACM Conf the proposed GroRec system is superior to that of the Recommender Syst, 2011, pp 301-304 conventional CF-bascd and MF-bascd approaches How- [22] S Yang, M. A Cheema, X Lin, and W. Wang, "Reverse k nearest neighbors query processing: Experiments and analysis, "Proc. ever, the proposed scheme can still be further improved, VLDB Endowment, vol 8, no 5, pp. 605-616, 2015 considering that the KNN-based group discovery mod- [23] Z Xu, R Lu, L Xiang, and Q Yang, Discovering user interest on ule has limited processing power. Therefore, in our future work, we will attempt to develop a deep-learnin ACMInt. Conf. Web Intell. Intell. Agent Technol., vol. 1, Pp 422-429, 2011. based approach to improve the efficiency and accuracy [24] J. Das, S Majumder, D. Dutta, and P Gupta, "Iterative use of of group discovery weighted voronoi diagrams to improve scalability in recom- mender systems,"in Advances in Knowledge Discovery and Data Mining. Berlin, Germany: Springer, 2015, pp. 605-617 REFERENCES 125] D. Li et al. ,"An algorithm for efficient privacy-preserving item- [1] F.-Y. Wang, The emergence of intelligent enterprises: From CPs based collaborative filtering, Future Generation Comput. Syst vol.55,pp.311-320,2016 to CPSS,"IEEE Intell. Syst. vol. 25, no. 4, Pp 85-88, JuL /Aug. [26] P. Pirasteh, D. Hwang, and J. E. Jung, "Weighted similarity 2010 [2] G. Xiong, et al. ,"Cyber-physical-social system in intelligent trans schemes for high scalability in user-based collaborative filtering, portation, "IEEE/CAA/. Automatica Sinica, vol. 2, no. 3, pp. 320 Mobile Netw. Appl. vol 20, no 4, pp. 497-507, 2015 33,jul.2015 [3 Y Hu, F.-Y. Wang, and X. Liu, "A CPSS approach for emergency Yin Zhang is a faculty member in the School of evacuation in building fires, IEEE Intell. Syst., vol. 29, no. 3 Information and safety Engineering, Zhongnan PP.48-52,May/Jun.2014 University of Economics and Law. He worked as [ 4] H. Ning, H. Liu, J. Ma, L. T. Yang, and R Huang, " Cybcrmatics a poster-doctoral fellow in the School of Com Cyber-physical-social-thinking hyperspace based science and puter Science and Technology, Huazhong Univer technology, " Future Generat. Comput. Syst, vol. 56, PP. 504-522 sity of Science and Technology, China. He is 2015 vice-chair of IEEE Computer Society Big Data [5] M. Chen, S. Mao, and Y. Liu, " Big data: A survey Mobile netw STC. His research interests include data analy Appl,vol.19,no.2,pp.171-209,2014 sis,data mining, healthcare big data, and social [61 A Sheth, P. Anantharam, and C. Henson, " Physical-cyber-social network. He has published more than 30 presti gious conference and journal papers. He serves computing: An early 21st century approach, "IEEE Intell. Syst., as topic editor, associate editor or guest editor for the IEEE Sensors vol28,no.1,pp.78-82,Jan./Feb.2013 Journal, the /EEE Access, and the Journal of Medical Systems and New [7] Y.L. Chen, L.-C. Cheng, and C.-N. Chuang,A group recommen- Review of Hypermedia and Multimedia. He also served as TPC co-chair dation system with consideration of interactions among group of CloudComp 2015, and local chair of TRIDENTCOM 2014. Since A Salehi-abari andc. boutilier,"preference-oriented social net- 2016, he is a senior member of the IEEE works: Group recommendation and inference, "in Proc. 9th ACM Conf Recommender Syst, 2015, pp 35-42 9] L. Boratto and S Carta, The rating prediction task in a group rec ommender system that automatically detects groups: Architec- tures, algorithms, and performance evaluation, Intell Inform Syst.,vol.45,no.2,pp.22l-245,2015 [10] Y. Zheng, B. Mohasher, and R D. Burke, " The role of emotions in context-aware recommendation, Decisions@ Rec Sys, vol. 2013 FP.21-28,2013 [11] Y. H. Chin, S. H. Lin, C. H. Lin, E. Siahaan, A. Frisky, and J. C Wang,"Emotion profile-based music recommendation, in Proc IEEE7th Int Conf. Ulbi-Media Comput Workshops, 2014, pp 111-114 [12] C. Orellana-Rodriguez, E. Diaz-Aviles, and g affective context in short films for emotion -aware recommen dation, in Proc. 26th ACM Conf. Hypertext Soc. Media, 2015, 185-194 [13] S Fazeli, B Loni, A. Bellogin, H. Drachsler, and P Sloep, Implicit versus explicit trust in social matrix factorization, in Proc. 8th ACM Conf Recommender Syst, 2014, pp 317-320 [14] F Wu, Z Li, W.C. Lee, H. Wang, and z Huang,"Semantic ann taion of mobility data using social media, " in Proc. 24th Int. Conf World Wide Web, 2015, Pp 1253-1263 [15] R Xia, Y Pan, L Du, and J. Yin, Robust multi-view spectral clus tering via low-rank and sparse decomposition, in Proc. 28th AAAl Conf. Artificial Intell., 2014, pp. 2149-2155 [16] Q. Song, J. Cheng, and H. Lu, " Incremental matrix factorization via feature space re-learning for recommender system in Pr 9th ACM Conf Recommender Syst, 2015, pp 277-280

...展开详情
试读 10P GroRec:集成社交,移动和大数据技术的以群体为中心的智能系统
img
  • GitHub

    绑定GitHub第三方账户获取

关注 私信 TA的资源

上传资源赚积分,得勋章
    最新推荐
    GroRec:集成社交,移动和大数据技术的以群体为中心的智能系统 1积分/C币 立即下载
    1/10
    GroRec:集成社交,移动和大数据技术的以群体为中心的智能系统第1页
    GroRec:集成社交,移动和大数据技术的以群体为中心的智能系统第2页
    GroRec:集成社交,移动和大数据技术的以群体为中心的智能系统第3页

    试读已结束,剩余7页未读...

    1积分/C币 立即下载 >