A Hybrid Recommender System Based on AHP That Awares Contexts ...
A Hybrid Recommender System Based on AHP That Awares Contexts with Bayesian Networks for Smart TV Ji-Chun Quan and Sung-Bae Cho Department of Computer Science, Yonsei University 50 Yonsei-ro, Seodaemun-gu, Seoul 120-749, Korea [email protected], [email protected]
Abstract. Recently, many researchers are paying close attention to TV program recommendation methods because of the enormous increase of available TV programs for users. As TV programs are often watched by multiple users like a family, this paper proposes a smart TV program recommendation method for multi-users using Bayesian networks and AHP (analytic hierarchy process). The proposed method uses Bayesian networks to infer each user’s genre preference as well as program preference, and uses AHP to predict group genre preference and choose recommended programs. The accuracy of the Bayesian network model is improved through parameter learning from users’ watching history. Experiments verify the inference accuracy of the Bayesian network and the accuracy of programs recommended by the proposed method. Keywords: TV program recommendation, Bayesian network, Analytic hierarchy process, Group recommendation.
1
Introduction
As the technology of smart TV is rapidly developed, users can watch TV program not only from traditional television station but also from Internet [1]. Therefore a system which can recommend TV programs considering users’ preference would help people to save time and energy [2, 3]. For this purpose, a number of researchers have proposed lots of TV program recommendation solutions. These solutions recommend TV programs according to users’ preference which is predicted by users’ profile or watching history. One drawback of these solutions is that they did not consider some information such as period (e.g., Olympic) and event (e.g., the last episode of a drama). Also, the cold start problem is another headache for these methods. In this paper, we propose a TV program recommendation method for multiple users as well as single user using a hybrid recommendation method based on AHP (analytic hierarchy process) and Bayesian networks. A hybrid approach exploits the strengths of the individual method and enhances the performance by their combination [4, 5]. The proposed hybrid method can overcome several problems which cannot be addressed by single method. The contributions of this paper can be summarized as follows. M. Polycarpou et al. (Eds.): HAIS 2014, LNAI 8480, pp. 527–536, 2014. © Springer International Publishing Switzerland 2014
528
J.-C. Quan and S.-B. Cho
• A hybrid method of AHP and Bayesian networks: The proposed method predicts group preference using a hybrid method of AHP and Bayesian networks which can solve the consistency problem of AHP. Also, with this method, users do not need to provide the priority value of candidate decision when comparing each candidate decision with pairwise comparison in AHP. • Solution of cold start problem: There are two kinds of cold start problem. One is happened when there are no data of user’s watching history and the other is when all candidate programs are totally new program or have less watched time. The former problem is addressed by Bayesian networks and the latter is worked out by program audience rating from Internet. The rest of this paper is organized as follows. Section 2 briefly reviews the related works for TV recommendation method and Section 3 presents the proposed recommendation method in detail. In Section 4, we conduct some experiments and analyze the results, and Section 5 concludes the work.
2
Related Works
At most of time, since there is more than one person watching TV, the research for TV program recommendation for multiple users is meaningful [6]. Yu et al. recommended TV programs by using a group profile which is constructed by aggregating all users’ profile in that group [7]. Shin et al. integrated users’ profile, watched TV program profile as well as watching history into a group file and recommend TV programs based on it [8]. Rafael et al. proposed a method which recommended TV programs using collaborative filtering and content-based method with a TV ontology model [9]. Thyagaraju et al. computed group preference using each user’s preference in the group which was extracted from the user’s watching history [10]. Wang et al. proposed a method which could estimate group preference based on external experts’ preference [11]. These studies did not consider some program information which could affect the final decision. For example, during World Cup, user may watch football game instead of drama or other programs. Another drawback of these methods is that recommendation results of them are highly influenced by the number of users in the group. To address these problems, AHP is used to predict group preference and select TV program. AHP is one of the most widely used multi-criteria decision making method [12]. Chen et al. used AHP method to construct a context-aware mobile recommendation system [13]. Park et al. constructed a restaurant recommendation system for multiple users using AHP method [14]. Wu et al. proposed a web services selection method based on AHP and Wiki [15]. Because of using AHP, the proposed system could consider kinds of criteria when predicting group preference and selecting recommended TV program. However, they did not get over the cold start problem. In this paper, the Bayesian network is used to address this problem. The Bayesian network is a directed acyclic graph which is a represantative method giving believable results predicted to users in uncertain environment [16]. It uses cause-effect relationship between parent nodes and children nodes to predict results through
AHP That Awares Contexts with Bayesian Networks for Smart TV
529
probability-based method [17]. Yang et al. constructed a context-aware system in smart TV using Bayesian networks with domain knowledge [18]. Park et al. design a Bayesian netwok with domain knowledge for context-aware robot in home service environment [19]. Park et al. developed a context-sharing system with Bayesian netwok [20]. One advantage of the Bayesian networks is that it could be constructed using domain knowledge, which means that we could construct a Bayesian network for a user without using the user’s specific information. Although the accuracy may be lower, we could update the Bayesian network when we get enough information about the user’s watching history.
3
A TV Program Recommendation Method for Multiple Users
As shown in Fig. 1, the program recommendation method consists of group genre preference inference part, candidate program integration part and program recommendation part.
Fig. 1. Overview of the program recommendation method
• Group Genre Preference Inference: To predict group genre preference, first, each user’s genre preference(GPR) is inferred by the GPR Bayesian network model. After that, the group‘s GPR is computed using GPR AHP model and each user’s GPR which is inferred in the previous step. Finally, the top two genres are selected as the group’s prefer genre.
530
J.-C. Quan and S.-B. Cho
• Candidate Program Integration: In this part, the method selects candidate programs based on the inferred group’s GPR as well as each user’s watching history in that group and a program selection algrorithm. • Program Recommendation: To recommend the most attractable programs to users, firstly each user’s preference to each candidate program is calculated by the program preference (PPR) Bayesian network model. Then, the group PPR AHP model is used to rank the cadidate programs and after that, the top N programs are recommended to the users. In this paper, we select 3 programs to users. 3.1
Inference with Bayesian Networks
In Bayesian networks, the belief of node A with evidence B is calculated by equation (1). |
|
The joint probabilities of the Bayesian network is calculated by chain rule and we can get equation (2). ,
,…,
∏
| …, | ,…,
(1) ,
,… ,
| ,…, (2)
With Markov propery, the equation (2) could be changed to equation (3). ,
,…,
∏
|
(3)
To recommend appropriate programs to users, we construct two Bayesian network models for each user as shown in Fig. 2 and Fig. 3. These models use equation (1) and (3) to calculate each user‘s preference.
Fig. 2. The Bayesian network model for program preference inference
Fig. 2 is the Bayesian network model for program preference inference. It is calculated using the program’s information, watching history of the program as well as the user’s genre preference information. Fig. 3 is the genre preference inference Bayesian network model which represents the relationship between a user’s genre preference and time.
AHP That Awares Contexts with Bayesian Networks for Smart TV
531
Fig. 3. The Bayesian network model for genre preference inference
3.2
Byesian Network Learning
Each user’s genre preference Bayesian network model should be updated regularly to represent the user’s patterns of watching TV. In this paper, the Maximum Likelihood (ML) estimation algorithm is applied to learn and update the Bayesian network model using one user’s watching history and profile. ML algorithm is a parameter learning algorithm when the structure of a Bayesian network model is fixed [21]. The conditional probability table (CPT) of each Bayesian network model’s node is generated by equation (4). ,
| In the equation (4), the and Here, 3.3
(4)
,
is the number of occurrences that come out at the same time in the data set. is the appearance time of in the data set.
Candidate Program Selection
To select candidate programs based on inferred group GPR, the equation (5) is used to rank the programs in the user’s watching history. 1
where
(5)
In the proposed method, the top N programs are put into candidate program pool. In the equation (5), R is the score of the program, E is the audience rate of the program, and P is the watching ratio of that program. In addition, T is the watching time
532
J.-C. Quan and S.-B. Cho
of the program and S is the total show time of the program on TV. The value is used to decide which criteria would be given more weight. Because of the value , these newly opened TV programs, like a new TV series, could be competitive with other programs. In this paper, if the value P is smaller than 0.4, the value of is set to 0.6, while it is set to 0.4 if the P is larger than 0.4. 3.4
Group Decision by a Combination of AHP and Baeysian Network
To predict group GPR and select appropriate programs to group users, we constrcut group GPR AHP model (Fig. 4) and group PPR AHP model (Fig. 5), respectively.
Fig. 4. AHP model for group genre preference
An AHP model consists of three hierarchies which are goal, criteria and alternatives. As shown in Fig. 4, the criteria of group GPR AHP model are each user’s genre preference and the alternatives are the program genre. Similarly, as shown in Fig. 5, the criteria of group PPR AHP model are each user’s program preference and the alternatives are candidate programs.
Fig. 5.AHP model for group program preference
In the original AHP method, users have to allocate the relative importance to each alternative decision with respect to each criterion. However, for TV program recommendation, the relative importance must be computed automatically as it is
AHP That Awares Contexts with Bayesian Networks for Smart TV
533
impossible to let users give it to every candidate program and genre by themselves. To address this problem, we use a hybrid method of AHP and Bayesian networks in which the importance value of each AHP alternative decision is calculated by a Bayesian network model and rank of these candidate decisions is computed by AHP method. Another advantage of this method is that there is no need to check the consistency of AHP as the comparison with numerical importance value will be always transitive. The process of the method consists of the following steps and the group GPR computation process will be used as an example to explain these steps. At the first step, the comparison matrix is constructed for criteria and the alternatives with respect to each criterion through pairwise comparison. To construct the matrix for criteria of group GPR AHP model, like equation (6) , the users’ relative importance value ( in the equation) collected from user’s profile will be used. ,
(6)
To construct the comparison matrix, , , ,…, in which is representative of the alternative matrix with respect to criterion n, for candidate genres with respect to each criterion, like equation (7), each user’s genre preference predicted by the Bayesian network model will be used. In the equation, means user n’s preference regarding to genre i. ,
(7)
At the second step, the priority of criterion and the alternatives with respect to each criterion is computed by equation (8). In the equation, i represents a criterion or an alternative and n and m represent the number of colums and rows in the matrix, respectively. For group GPR decision, a priority set, , ,…, , is constructed for criteria and a priority set, , ,…, , is constructed for candidate genres. ∑
∑
(8)
At the last step, the final priority of each alternative is computed by using equation (9). (9) After that, all alternatives are sorted by the final priority and top N alternatives are recommended to the users.
534
4
J.-C. Quan and S.-B. Cho
Experimental Results
In this section, we conduct two experiments to evaluate the accuracy of the genre Bayesian network model and the users’ satisfaction about the recommended TV programs. The TV program information was downloaded from NAVER TV guide and 20 people attended out at the experiments. In addition, we collected the users’ watching history for a month through survey. 4.1
Accuracy Evaluation of the Genre Bayesian Networks
In this experiment, we compared the result predicted by a Bayesian network model after learning from a user’s watching history, a Bayesian network constructed by the domain knowledge, audience rate data from the internet, and rule-based inference method. We selected 5 times randomly to infer user’s genre preference and the rulebased inference method always select the most watched TV program’s genre as the predicted result. The accuracy is estimated by asking the participants in the experiment on which result is the most satisfied. 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Rule based method
1 0.5
2 0.5
3 0.55
4 0.65
5 0.6
Domain knowledge network
0.55
0.6
0.65
0.75
0.7
Learned network
0.75
0.9
0.85
0.85
0.8
Fig. 6. Accuracy of the genre preference Bayesian network model
As shown in Fig. 6, the learned network get the best accuracy. Through the learning process about one user’s watching history, the Baeysian network could represent the user’s watching pattern. For example, users select a drama which they have watched before instead of a drama which have the most audience rate. 4.2
Accuracy Evaluation of the Proposed Group Recommendation Method
In this experiment, participants were divided into ten 2-person groups, nine 3-person groups, seven 4-person groups as well as four 5-person groups. And we compared the proposed group recommendation method with a rule-based method and a neural network-based method. Rule-based method recommends the program which has the most audience rate while the neural network-based method is trained by all users’ watching history.
AHP That Awares Contexts with Bayesian Networks for Smart TV
535
As shown in Fig. 7, the proposed method has the best performance as it recommends TV programs considering group’s genre preference, users’ program preference and the relative importance of each user. After that is the neural network-based method and the rule-based method has the worst performance. It means that not all people prefer the program with higher audience rate, and some information like period and event of the program really affects the accuracy of the recommendation. 1
0.8
0.6
0.4 Rule based method
1
2
3
4
5
0.5
0.57
0.47
0.6
0.62
Neural network
0.72
0.7
0.75
0.68
0.71
Poposed method
0.77
0.83
0.8
0.87
0.83
Fig. 7. Accuracy of the proposed group recommendation method
5
Concluding Remarks
In this paper, we have proposed a TV program recommendation method for group users. The proposed method uses a hybrid method of Bayesian networks and AHP to predict group’s genre preference and recommend TV program to users from candidate program pool. Also, we propose a TV program selection algorithm to generate the candidate program pool. Furthermore, the cold start problem is addressed by applying Bayesian networks. Finally, we verify the advantage of the proposed method by two accuracy evaluation experiments. As future work, we will develop a smart TV-based system which has user-friendly interface and can collect users’ real watching history data. Also, more experiments will be done with these real data and more participants.
References 1. Kim, M.K., Park, J.H.: Demand forecasting and strategies for the successfully deployment of the smart TV in Korea. In: 13th Int. Conf. on Advanced Communication Technology (ICACT), pp. 1475–1478 (2011) 2. Kim, E.-H., Pyo, S.-J., Park, E.-Y., Kim, M.-C.: An automatic recommendation scheme of TV program contents for (IP)TV personalization. IEEE Transactions on Broadcasting 58(3), 674–684 (2011) 3. Kwon, H.-J., Hong, K.-S.: Personalized smart TV program recommender based on collaborative filtering and a novel similarity method. IEEE Transaction on Consumer Electronics 57(3), 1416–1423 (2011)
536
J.-C. Quan and S.-B. Cho
4. Wozniak, M., Graña, M., Corchado, E.: A survey of multiple classifier systems as hybrid systems. Information Fusion 16, 3–17 (2014) 5. Borrajo, M.L., Baruque, B., Corchado, E., Bajo, J., Corchado, J.: Hybrid neural intelligent system to predict business failure in small-to-medium-size enterprises. International Journal of Neural Systems 21(04), 277–296 (2011) 6. Konstantinos, C.: Personalized and mobile digital TV applications. Multimedia Tools and Applications 36(1-2), 1–10 (2008) 7. Yu, Z., Zhou, X., Hao, Y., Gu, J.: TV program recommendation for multiple viewers based on user profile merging. User Modeling and User-Adapted Interaction 16(1), 63–82 (2006) 8. Shin, C., Woo, W.: Socially aware TV program recommender for multiple viewers. IEEE Transactions on Consumer Electronics 55(2), 927–932 (2009) 9. Rafael, S., Yolanda, B.-F., Martin, L.-N., Alberto, G.-S.: TV program recommendation for groups based on multi-dimensional TV-anytime classifications. IEEE Transactions on Consumer Electronics 55(1), 248–256 (2009) 10. Thyagaraju, G., Kulkarni, U.P.: Interactive democratic group preference algorithm for interactive context aware TV. In: IEEE Int. Conf. on Computational Intelligence and Computing Research, pp. 1–5 (2010) 11. Wang, X., Sun, L., Wang, Z., Meng, D.: Group recommendation using external followee for social TV. In: Int. Conf. on Multimedia and Expo, pp. 37–42 (2012) 12. Saaty, T.L.: How to make a decision: The analytic hierarchy process. European Journal of Operational Research 48, 9–26 (1990) 13. Chen, D.-N., Hu, P.J.-H., Kuo, Y.-R., Liang, T.-P.: A web-based personalized recommendation system for mobile phone selection: Design, implementation, and evaluation. Expert Systems with Applications 37(12), 8201–8210 (2010) 14. Park, M.-H., Park, H.-S., Cho, S.-B.: Restaurant recommendation for group of people in mobile environments using probabilistic multi-criteria decision making. In: Lee, S., Choo, H., Ha, S., Shin, I.C. (eds.) APCHI 2008. LNCS, vol. 5068, pp. 114–122. Springer, Heidelberg (2008) 15. Wu, C., Chang, E.: Intelligent web services selection based on AHP and Wiki. In: Int. Conf. on Web Intelligence, pp. 767–770 (2007) 16. Charniak, E.: Bayesian networks without tears. AI Magazine 12, 50–63 (1991) 17. Kleiter, G.D.: Propagating imprecise probabilities in Bayesian networks. Artificial Intelligence 88(1-2), 143–161 (1996) 18. Yang, K.-M., Cho, S.-B.: Probabilistic modeling for context-aware service in smart TV. In: Intelligent Systems Modeling and Simulation, pp. 78–83 (2013) 19. Park, H.-S., Cho, S.-B.: A modular design of Bayesian networks using expert knowledge: Context-aware home service robot. Expert Systems with Applications 39(3), 2629–2642 (2012) 20. Park, H.-S., Oh, K.-H., Cho, S.-B.: Bayesian network-based high-level context recognition for mobile context sharing in cyber-physical system. Distributed Sensor Networks, 1–10 (2011) 21. David, H.: A tutorial on learning with Bayesian networks. In: Holmes, D.E., Jain, L.C. (eds.) Innovations in Bayesian Networks. SCI, vol. 156, pp. 33–82. Springer, Heidelberg (2008)
Abstract. Recently, many researchers are paying close attention to TV program recommendation methods because of the enormous increase of available TV programs for users. As TV programs are often watched by multiple users like a family, this paper proposes a smart TV program recommendation method for multi-users using Bayesian networks and AHP (analytic hierarchy process). The proposed method uses Bayesian networks to infer each user’s genre preference as well as program preference, and uses AHP to predict group genre preference and choose recommended programs. The accuracy of the Bayesian network model is improved through parameter learning from users’ watching history. Experiments verify the inference accuracy of the Bayesian network and the accuracy of programs recommended by the proposed method. Keywords: TV program recommendation, Bayesian network, Analytic hierarchy process, Group recommendation.
1
Introduction
As the technology of smart TV is rapidly developed, users can watch TV program not only from traditional television station but also from Internet [1]. Therefore a system which can recommend TV programs considering users’ preference would help people to save time and energy [2, 3]. For this purpose, a number of researchers have proposed lots of TV program recommendation solutions. These solutions recommend TV programs according to users’ preference which is predicted by users’ profile or watching history. One drawback of these solutions is that they did not consider some information such as period (e.g., Olympic) and event (e.g., the last episode of a drama). Also, the cold start problem is another headache for these methods. In this paper, we propose a TV program recommendation method for multiple users as well as single user using a hybrid recommendation method based on AHP (analytic hierarchy process) and Bayesian networks. A hybrid approach exploits the strengths of the individual method and enhances the performance by their combination [4, 5]. The proposed hybrid method can overcome several problems which cannot be addressed by single method. The contributions of this paper can be summarized as follows. M. Polycarpou et al. (Eds.): HAIS 2014, LNAI 8480, pp. 527–536, 2014. © Springer International Publishing Switzerland 2014
528
J.-C. Quan and S.-B. Cho
• A hybrid method of AHP and Bayesian networks: The proposed method predicts group preference using a hybrid method of AHP and Bayesian networks which can solve the consistency problem of AHP. Also, with this method, users do not need to provide the priority value of candidate decision when comparing each candidate decision with pairwise comparison in AHP. • Solution of cold start problem: There are two kinds of cold start problem. One is happened when there are no data of user’s watching history and the other is when all candidate programs are totally new program or have less watched time. The former problem is addressed by Bayesian networks and the latter is worked out by program audience rating from Internet. The rest of this paper is organized as follows. Section 2 briefly reviews the related works for TV recommendation method and Section 3 presents the proposed recommendation method in detail. In Section 4, we conduct some experiments and analyze the results, and Section 5 concludes the work.
2
Related Works
At most of time, since there is more than one person watching TV, the research for TV program recommendation for multiple users is meaningful [6]. Yu et al. recommended TV programs by using a group profile which is constructed by aggregating all users’ profile in that group [7]. Shin et al. integrated users’ profile, watched TV program profile as well as watching history into a group file and recommend TV programs based on it [8]. Rafael et al. proposed a method which recommended TV programs using collaborative filtering and content-based method with a TV ontology model [9]. Thyagaraju et al. computed group preference using each user’s preference in the group which was extracted from the user’s watching history [10]. Wang et al. proposed a method which could estimate group preference based on external experts’ preference [11]. These studies did not consider some program information which could affect the final decision. For example, during World Cup, user may watch football game instead of drama or other programs. Another drawback of these methods is that recommendation results of them are highly influenced by the number of users in the group. To address these problems, AHP is used to predict group preference and select TV program. AHP is one of the most widely used multi-criteria decision making method [12]. Chen et al. used AHP method to construct a context-aware mobile recommendation system [13]. Park et al. constructed a restaurant recommendation system for multiple users using AHP method [14]. Wu et al. proposed a web services selection method based on AHP and Wiki [15]. Because of using AHP, the proposed system could consider kinds of criteria when predicting group preference and selecting recommended TV program. However, they did not get over the cold start problem. In this paper, the Bayesian network is used to address this problem. The Bayesian network is a directed acyclic graph which is a represantative method giving believable results predicted to users in uncertain environment [16]. It uses cause-effect relationship between parent nodes and children nodes to predict results through
AHP That Awares Contexts with Bayesian Networks for Smart TV
529
probability-based method [17]. Yang et al. constructed a context-aware system in smart TV using Bayesian networks with domain knowledge [18]. Park et al. design a Bayesian netwok with domain knowledge for context-aware robot in home service environment [19]. Park et al. developed a context-sharing system with Bayesian netwok [20]. One advantage of the Bayesian networks is that it could be constructed using domain knowledge, which means that we could construct a Bayesian network for a user without using the user’s specific information. Although the accuracy may be lower, we could update the Bayesian network when we get enough information about the user’s watching history.
3
A TV Program Recommendation Method for Multiple Users
As shown in Fig. 1, the program recommendation method consists of group genre preference inference part, candidate program integration part and program recommendation part.
Fig. 1. Overview of the program recommendation method
• Group Genre Preference Inference: To predict group genre preference, first, each user’s genre preference(GPR) is inferred by the GPR Bayesian network model. After that, the group‘s GPR is computed using GPR AHP model and each user’s GPR which is inferred in the previous step. Finally, the top two genres are selected as the group’s prefer genre.
530
J.-C. Quan and S.-B. Cho
• Candidate Program Integration: In this part, the method selects candidate programs based on the inferred group’s GPR as well as each user’s watching history in that group and a program selection algrorithm. • Program Recommendation: To recommend the most attractable programs to users, firstly each user’s preference to each candidate program is calculated by the program preference (PPR) Bayesian network model. Then, the group PPR AHP model is used to rank the cadidate programs and after that, the top N programs are recommended to the users. In this paper, we select 3 programs to users. 3.1
Inference with Bayesian Networks
In Bayesian networks, the belief of node A with evidence B is calculated by equation (1). |
|
The joint probabilities of the Bayesian network is calculated by chain rule and we can get equation (2). ,
,…,
∏
| …, | ,…,
(1) ,
,… ,
| ,…, (2)
With Markov propery, the equation (2) could be changed to equation (3). ,
,…,
∏
|
(3)
To recommend appropriate programs to users, we construct two Bayesian network models for each user as shown in Fig. 2 and Fig. 3. These models use equation (1) and (3) to calculate each user‘s preference.
Fig. 2. The Bayesian network model for program preference inference
Fig. 2 is the Bayesian network model for program preference inference. It is calculated using the program’s information, watching history of the program as well as the user’s genre preference information. Fig. 3 is the genre preference inference Bayesian network model which represents the relationship between a user’s genre preference and time.
AHP That Awares Contexts with Bayesian Networks for Smart TV
531
Fig. 3. The Bayesian network model for genre preference inference
3.2
Byesian Network Learning
Each user’s genre preference Bayesian network model should be updated regularly to represent the user’s patterns of watching TV. In this paper, the Maximum Likelihood (ML) estimation algorithm is applied to learn and update the Bayesian network model using one user’s watching history and profile. ML algorithm is a parameter learning algorithm when the structure of a Bayesian network model is fixed [21]. The conditional probability table (CPT) of each Bayesian network model’s node is generated by equation (4). ,
| In the equation (4), the and Here, 3.3
(4)
,
is the number of occurrences that come out at the same time in the data set. is the appearance time of in the data set.
Candidate Program Selection
To select candidate programs based on inferred group GPR, the equation (5) is used to rank the programs in the user’s watching history. 1
where
(5)
In the proposed method, the top N programs are put into candidate program pool. In the equation (5), R is the score of the program, E is the audience rate of the program, and P is the watching ratio of that program. In addition, T is the watching time
532
J.-C. Quan and S.-B. Cho
of the program and S is the total show time of the program on TV. The value is used to decide which criteria would be given more weight. Because of the value , these newly opened TV programs, like a new TV series, could be competitive with other programs. In this paper, if the value P is smaller than 0.4, the value of is set to 0.6, while it is set to 0.4 if the P is larger than 0.4. 3.4
Group Decision by a Combination of AHP and Baeysian Network
To predict group GPR and select appropriate programs to group users, we constrcut group GPR AHP model (Fig. 4) and group PPR AHP model (Fig. 5), respectively.
Fig. 4. AHP model for group genre preference
An AHP model consists of three hierarchies which are goal, criteria and alternatives. As shown in Fig. 4, the criteria of group GPR AHP model are each user’s genre preference and the alternatives are the program genre. Similarly, as shown in Fig. 5, the criteria of group PPR AHP model are each user’s program preference and the alternatives are candidate programs.
Fig. 5.AHP model for group program preference
In the original AHP method, users have to allocate the relative importance to each alternative decision with respect to each criterion. However, for TV program recommendation, the relative importance must be computed automatically as it is
AHP That Awares Contexts with Bayesian Networks for Smart TV
533
impossible to let users give it to every candidate program and genre by themselves. To address this problem, we use a hybrid method of AHP and Bayesian networks in which the importance value of each AHP alternative decision is calculated by a Bayesian network model and rank of these candidate decisions is computed by AHP method. Another advantage of this method is that there is no need to check the consistency of AHP as the comparison with numerical importance value will be always transitive. The process of the method consists of the following steps and the group GPR computation process will be used as an example to explain these steps. At the first step, the comparison matrix is constructed for criteria and the alternatives with respect to each criterion through pairwise comparison. To construct the matrix for criteria of group GPR AHP model, like equation (6) , the users’ relative importance value ( in the equation) collected from user’s profile will be used. ,
(6)
To construct the comparison matrix, , , ,…, in which is representative of the alternative matrix with respect to criterion n, for candidate genres with respect to each criterion, like equation (7), each user’s genre preference predicted by the Bayesian network model will be used. In the equation, means user n’s preference regarding to genre i. ,
(7)
At the second step, the priority of criterion and the alternatives with respect to each criterion is computed by equation (8). In the equation, i represents a criterion or an alternative and n and m represent the number of colums and rows in the matrix, respectively. For group GPR decision, a priority set, , ,…, , is constructed for criteria and a priority set, , ,…, , is constructed for candidate genres. ∑
∑
(8)
At the last step, the final priority of each alternative is computed by using equation (9). (9) After that, all alternatives are sorted by the final priority and top N alternatives are recommended to the users.
534
4
J.-C. Quan and S.-B. Cho
Experimental Results
In this section, we conduct two experiments to evaluate the accuracy of the genre Bayesian network model and the users’ satisfaction about the recommended TV programs. The TV program information was downloaded from NAVER TV guide and 20 people attended out at the experiments. In addition, we collected the users’ watching history for a month through survey. 4.1
Accuracy Evaluation of the Genre Bayesian Networks
In this experiment, we compared the result predicted by a Bayesian network model after learning from a user’s watching history, a Bayesian network constructed by the domain knowledge, audience rate data from the internet, and rule-based inference method. We selected 5 times randomly to infer user’s genre preference and the rulebased inference method always select the most watched TV program’s genre as the predicted result. The accuracy is estimated by asking the participants in the experiment on which result is the most satisfied. 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Rule based method
1 0.5
2 0.5
3 0.55
4 0.65
5 0.6
Domain knowledge network
0.55
0.6
0.65
0.75
0.7
Learned network
0.75
0.9
0.85
0.85
0.8
Fig. 6. Accuracy of the genre preference Bayesian network model
As shown in Fig. 6, the learned network get the best accuracy. Through the learning process about one user’s watching history, the Baeysian network could represent the user’s watching pattern. For example, users select a drama which they have watched before instead of a drama which have the most audience rate. 4.2
Accuracy Evaluation of the Proposed Group Recommendation Method
In this experiment, participants were divided into ten 2-person groups, nine 3-person groups, seven 4-person groups as well as four 5-person groups. And we compared the proposed group recommendation method with a rule-based method and a neural network-based method. Rule-based method recommends the program which has the most audience rate while the neural network-based method is trained by all users’ watching history.
AHP That Awares Contexts with Bayesian Networks for Smart TV
535
As shown in Fig. 7, the proposed method has the best performance as it recommends TV programs considering group’s genre preference, users’ program preference and the relative importance of each user. After that is the neural network-based method and the rule-based method has the worst performance. It means that not all people prefer the program with higher audience rate, and some information like period and event of the program really affects the accuracy of the recommendation. 1
0.8
0.6
0.4 Rule based method
1
2
3
4
5
0.5
0.57
0.47
0.6
0.62
Neural network
0.72
0.7
0.75
0.68
0.71
Poposed method
0.77
0.83
0.8
0.87
0.83
Fig. 7. Accuracy of the proposed group recommendation method
5
Concluding Remarks
In this paper, we have proposed a TV program recommendation method for group users. The proposed method uses a hybrid method of Bayesian networks and AHP to predict group’s genre preference and recommend TV program to users from candidate program pool. Also, we propose a TV program selection algorithm to generate the candidate program pool. Furthermore, the cold start problem is addressed by applying Bayesian networks. Finally, we verify the advantage of the proposed method by two accuracy evaluation experiments. As future work, we will develop a smart TV-based system which has user-friendly interface and can collect users’ real watching history data. Also, more experiments will be done with these real data and more participants.
References 1. Kim, M.K., Park, J.H.: Demand forecasting and strategies for the successfully deployment of the smart TV in Korea. In: 13th Int. Conf. on Advanced Communication Technology (ICACT), pp. 1475–1478 (2011) 2. Kim, E.-H., Pyo, S.-J., Park, E.-Y., Kim, M.-C.: An automatic recommendation scheme of TV program contents for (IP)TV personalization. IEEE Transactions on Broadcasting 58(3), 674–684 (2011) 3. Kwon, H.-J., Hong, K.-S.: Personalized smart TV program recommender based on collaborative filtering and a novel similarity method. IEEE Transaction on Consumer Electronics 57(3), 1416–1423 (2011)
536
J.-C. Quan and S.-B. Cho
4. Wozniak, M., Graña, M., Corchado, E.: A survey of multiple classifier systems as hybrid systems. Information Fusion 16, 3–17 (2014) 5. Borrajo, M.L., Baruque, B., Corchado, E., Bajo, J., Corchado, J.: Hybrid neural intelligent system to predict business failure in small-to-medium-size enterprises. International Journal of Neural Systems 21(04), 277–296 (2011) 6. Konstantinos, C.: Personalized and mobile digital TV applications. Multimedia Tools and Applications 36(1-2), 1–10 (2008) 7. Yu, Z., Zhou, X., Hao, Y., Gu, J.: TV program recommendation for multiple viewers based on user profile merging. User Modeling and User-Adapted Interaction 16(1), 63–82 (2006) 8. Shin, C., Woo, W.: Socially aware TV program recommender for multiple viewers. IEEE Transactions on Consumer Electronics 55(2), 927–932 (2009) 9. Rafael, S., Yolanda, B.-F., Martin, L.-N., Alberto, G.-S.: TV program recommendation for groups based on multi-dimensional TV-anytime classifications. IEEE Transactions on Consumer Electronics 55(1), 248–256 (2009) 10. Thyagaraju, G., Kulkarni, U.P.: Interactive democratic group preference algorithm for interactive context aware TV. In: IEEE Int. Conf. on Computational Intelligence and Computing Research, pp. 1–5 (2010) 11. Wang, X., Sun, L., Wang, Z., Meng, D.: Group recommendation using external followee for social TV. In: Int. Conf. on Multimedia and Expo, pp. 37–42 (2012) 12. Saaty, T.L.: How to make a decision: The analytic hierarchy process. European Journal of Operational Research 48, 9–26 (1990) 13. Chen, D.-N., Hu, P.J.-H., Kuo, Y.-R., Liang, T.-P.: A web-based personalized recommendation system for mobile phone selection: Design, implementation, and evaluation. Expert Systems with Applications 37(12), 8201–8210 (2010) 14. Park, M.-H., Park, H.-S., Cho, S.-B.: Restaurant recommendation for group of people in mobile environments using probabilistic multi-criteria decision making. In: Lee, S., Choo, H., Ha, S., Shin, I.C. (eds.) APCHI 2008. LNCS, vol. 5068, pp. 114–122. Springer, Heidelberg (2008) 15. Wu, C., Chang, E.: Intelligent web services selection based on AHP and Wiki. In: Int. Conf. on Web Intelligence, pp. 767–770 (2007) 16. Charniak, E.: Bayesian networks without tears. AI Magazine 12, 50–63 (1991) 17. Kleiter, G.D.: Propagating imprecise probabilities in Bayesian networks. Artificial Intelligence 88(1-2), 143–161 (1996) 18. Yang, K.-M., Cho, S.-B.: Probabilistic modeling for context-aware service in smart TV. In: Intelligent Systems Modeling and Simulation, pp. 78–83 (2013) 19. Park, H.-S., Cho, S.-B.: A modular design of Bayesian networks using expert knowledge: Context-aware home service robot. Expert Systems with Applications 39(3), 2629–2642 (2012) 20. Park, H.-S., Oh, K.-H., Cho, S.-B.: Bayesian network-based high-level context recognition for mobile context sharing in cyber-physical system. Distributed Sensor Networks, 1–10 (2011) 21. David, H.: A tutorial on learning with Bayesian networks. In: Holmes, D.E., Jain, L.C. (eds.) Innovations in Bayesian Networks. SCI, vol. 156, pp. 33–82. Springer, Heidelberg (2008)
