Group Recommender System Using Hybrid Method

(1)

Group Recommender System Using Hybrid Method

Esa Alfitrassalam, Ade Romadhon, Z K Aabdurahman Baizal^* Fakultas Informatika, Program Studi Informatika, Universitas Telkom, Bandung, Indonesia Email: ¹esaalfitra@students.telkomuniversity.ac.id, ²aderomadhony@telkomuniversity.ac.id,

3,*baizal@telkomuniversity.ac.id

Correspondence Author Email: baizal@telkomuniversity.ac.id

Abstract−In our daily activities, we make a lot of decisions either individually or in groups. The recommender systems is a solution for making decisions. One of the most common recommender systems is the recommendation for tourist destinations, where a number of tourist attractions are given as tourist attractions that are recommended to be visited by someone. There are still few recommended tourist attractions that provide recommendations for a group, while there are several tourist attractions that are more suitable if visited by several people at the same time. In this study, a recommender system for tourist attractions in Bandung-Raya Regency is proposed which is given to user groups. The recommended method used is Hybrid Collaborative Filtering and Knowledge-Based Filtering. In the process of selecting groups that are candidates to be recommended to users, Borda calculations are carried out with votes so that users can determine whether they like or dislike and match or not match the recommender generated by the system. The results of the evaluation of experiments conducted by taking surveys of users showed an average value. the average of the indicators of user satisfaction with the results of group recommender is 4.4 on the scale (1-5).

Keywords: Group Recommender System; Hybrid Method; Collaborative Filtering; Knowledge-Based; Borda

1. INTRODUCTION

At this time, Bandung-Raya Regency has become one of the tourist destinations for domestic and foreign people.

Tourist destinations in Bandung-Raya Regency are very varied, it is recorded that more than 90 tourist destinations have become favorite destinations for tourism. The sources of information obtained by tourists are very diverse, for example, television, newspapers, social media, and many other information media. The number of sources of information has a positive impact on both the Bandung-Raya Regency government and tourists who want to visit.

The development of national tourism areas is very important in advancing the economy or regional opinion [1].

The development of the latest science and technology (IPTEK) to provide a more advanced national tourism development. At this time global tourism is always open to new technologies, even in the development of web application technology, thereby increasing interest in the field of electronic tourism (e-tourism). [2]. In addition, for tourists who want to visit it will be easier to collect information about tourism destinations.

Various information obtained by tourists also has a negative impact, such as tourists finding it difficult to determine which destinations are relevant to visit. Moreover, when tourists want to vacation with groups, they must determine which destinations are suitable for all group members. So, a recommendersystem is needed to determine which destinations might be liked by tourists or groups of tourists. One of the previous studies, in research [3], is a group recommender system made using the Hybrid method that combines Content-Based Filtering (CB), Collaborative Filtering (CF), and Knowledge-Based Filtering (KB). The hybrid calculation method used is Parallel Weighted

From research conducted by Shini Renjith, Anjali C in 2014, regarding the Personalized Cellular Travel Recommendation System using a Hybrid Algorithm, it can produce a personal recommendation system by utilizing information from users, so that the results of the recommendations can be easily accepted by users,the method used is a hybrid system that combines CF, CB, and Demographic Filtering (DF). The use of combining the 3 methods is to mention the respective shortcomings in each CF, CB, and DF. For example, the CF fails to reference because the destination has never been visited by the end-user it can't be certain that it has ever been [4]. But when combined with CB, if the destination matches the user's behavior pattern, it will be recommended at least with a lower rating. New users can be classified into pre-determined classes in certain details where later they will use the K-Nearest Neighbor (KNN) algorithm with cosine similarity for the fitness function [4].

In this study, we apply Weighted parallel Hybridization in building a recommender system that combines Collaborative Filtering (CF) and Knowledge-Based Filtering (KB). The use of the Hybrid method is expected to provide users with accurate rating prediction results. After getting the rating prediction results from the Hybrid method, then displaying the 5 highest rating predictions will become a group recommender. In CF, the SVD (Singular Value Decomposition) algorithm is used to find the predicted rating value. The KB calculates Constrains Cost and Category to determine the user's desire, then calculates the rating prediction for KB. After getting the rating prediction results from CF and KB, then look for the rating prediction value using the Hybridization method.

The hybrid method uses the Weighted Technique to get more accurate results. In the group process, selecting destinations with a predicted rating of 4 will be made into groups based on members with the same tourist destination and users who give a rating = 5. In the final stage, ranking for each destination is carried out by members using the Borda method.

(2)

2. RESEARCH METHODOLOGY

2.1 Related Work

At this time many people have made decisions in a group way, both in daily life and in many other fields. In research [4], a group recommender system has been developed that uses the tourism domain, using 3 methods including Content-Based Filtering (CB), Collaborative Filtering (CF), and Knowledge-Based (KB). In CB apply similarity value search to the user with Cosine Similarity method and in CF use Pearson Correlation to get similarity value between user and neighborhood to get predicted rating CF and KB apply to Constrains Cost, Category, and GeoLocation methods to determine user needs and get precalculation results rating prediction. The results of CB, CF, and KB are to determine rating predictions and are combined in the Hybrid method using the Weighting technique.

In research [5] the recommender system method used is CF and CB where the process is carried out as follows, if the CB method does not find recommender that match user preferences, then CF will be carried out by looking at the assessment of a tourist attraction that has been carried out by the user. So that it can proceed to the calculation process involving all attributes and also destination categories where when the user determines the desired parameter, the system proceeds to the stage of finding the highest value which will use the weight that will be calculated with K-Nearest Neighbor (KNN) to get the recommender results from the case. old to a new case with the Euclidean distance equation.

Many previous studies have developed a recommendation system in the tourism sector, using a knowledge- based filtering approach, by utilizing the Multi Attribute Utility Theory (MAUT) to determine tourism recommendations that are in accordance with the user criteria and in this study using the limitations of popularity, cost and amount. objects that have been visited [6][7].

In this study, we built a tourism group recommender system based in Bandung-Raya Regency. The recommender system is made using the weighting hybrid method. The result of the hybridization between CF and KB is that it becomes the top 5 group recommender that will be recommended. CF is a process to model and train data by using Singular Value Decomposition (SVD) to get predictive ratings. In the family planning process, we ask the user to enter the travel budget and category of the tourist. After getting the top 5 results from Hybrid where the next process is to eliminate the assessment of group members who meet the criteria to become a group. In the final stage, we evaluate using Root Means Square Error (RMSE) and user satisfaction survey.

2.2 Recommender System

A recommender system is a system that provides recommender to potential buyers. Two techniques that are widely used to build recommender systems to date are collaborative screening and knowledge-based approaches [8]The purpose of the recommender system itself is to suggest items that may be of interest to users, to identify items that have an effect on users, the system must provide a value for recommending these items [9][10].

Content-Based Filtering provides recommender to users based on prior knowledge and what the user likes [5]. Collaborative Filtering is a recommended method that calculates the similarity value between users [8][11][3].

Knowledge-Based Filtering recommends users according to user needs and aligns user needs with available items.

[12].

2.2 Collaborative Filtering

Collaborative Filtering (CF) is a method that is commonly used in recommender systems. This recommender is based on the similarity between users and the predicted value of the rating [12]. CF is a user and item-based filtering, where the matrix factorization with least-squares in the feed is implicit in the matrix factorization weight [13].CF is divided into two classes, namely, Item-Based and User-Based. On Item-Based recommender that refer to the user on the value assessment. While in User-Based, using statistical techniques to get user groups [3].

In this study, we apply CF with the User-Based method and the SVD (Singular Value Decomposition).

Algorithm to obtain the best model and then predict the rating. In this study [3], we apply the Library Surprise SVD.

𝑅 ≈ (Ս) ∗ (𝑉𝑡) (1) R is the original matrix, U and V are the results of the original matrix divided into two.

2.3 Knowledge-Based Filtering

Knowledge-Based Filtering (KB) is a recommended method based on user knowledge and considering what items are most suitable for users. An example of family planning is the Personal Logic system that helps make decisions about various products, ranging from tourism, cars, and many other things [8]. KB recommends domain knowledge work on the influence of features on user needs and preferences. [11] .The design of a RS with knowledge explicit modeling, which represents all the needed knowledge to provide accurate recommendation [7]. KB filters out tourist destinations that are similar to what users want [3].

(3)

In this study, we use family planning with the Constrains Based method to determine user needs which are divided into 2 dimensions, namely tourism and cost categories. The formula applied is sourced from research [4].

As follows :

𝑟̂𝑢, 𝑖 = ∑ 𝑘∈𝐷 𝑤𝑘∗𝑠𝑐(𝑖,𝑘)

∑ 𝑘∈𝐷^𝑤𝑘 (2)

𝑠𝑐(𝑖, 𝑘) = 1 −^{(𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑}^{𝑐𝑜𝑠𝑡} −max 𝑏𝑢𝑑𝑔𝑒𝑡,0)

max _𝑏𝑢𝑑𝑔𝑒𝑡 (3)

𝑤𝑘 = 1 (4)

Where, is the prediction of user us a rating for item i while sc(i,k) is determining the user is a score for item i. wk is the weight of the feasible method KB = 1.

2.4 Hybrid

Hybrid is a method that combines several methods in a recommender system [3]. In this study, we use a combination of Collaborative Filtering (CF) and Knowledge-Based Filtering (KB) methods. The hybrid formula is used to combine the values of the rating predictions. In this study, we propose the Weighted Hybrid method by combining 2 methods, namely CF and KB. The w or weighted value is in both CF and KB methods. Because the weight or w if added is equal to 1 [4][14]. The formula used is as follows:

𝑟̂ℎ𝑦𝑏𝑟𝑖𝑑 = 𝑤𝑐𝑓 ∗ 𝑟̂𝑐𝑓 + 𝑤𝑘𝑏 ∗ 𝑟̂𝑘𝑏 (5)

𝑤𝑐𝑓 + 𝑤𝑘𝑏 = 1 (6)

Where 𝑤𝑐𝑓 is the weight of Collaborative Filtering (CF) 𝑟̂𝑐𝑓 is the result of the predicted CF rating, wkb is the weight of Knowledge-Based Filtering (KB) and 𝑟̂𝑘𝑏 is the predicted rating KB.

3. RESULTS AND DISCUSSION

3.1 Dataset

The dataset used in this research is sourced from research [14]. The data originated from the results of user registration websites and tourist destinations in the Bandung-Raya Regency. On the website, there is user rating data and details of tourism destinations such as tourism categories, entry fees, and more detailed information about these tourist destinations. The results of the data are csv and xlsx format files. Figure 1 shows the dataset processing process.

Figure 1. Dataset

Figure 1 describes the dataset processing process used in this study as follows:

1. User data: A collection of user registration data.

2. Tourism destination data: Data collection of Bandung-Raya Regency tourism destinations.

3. Preprocessing: Removing punctuation elements and Rp at cost, deleting data that does not provide a rating, normalizing kb rating data from (0-2) to (1-5).

4. Final Dataset: Total user and destination rating data.

3.2 System Design

In this study, we aim to develop a group recommender system with candidates for the same tourist destination.

This recommender system uses the Hybrid method between Collaborative Filtering (CF) and (KB), then the winner will be given a ranking using the Borda method. Figure 2 shows the basic system design regarding the system to be designed.

(4)

Figure 2. System Design

The first step we do is preprocessing so that the data can be processed. Furthermore, the CF data process is modeled and used using Singular Value Decomposition (SVD) to obtain a predictive rating. Furthermore, at the implementation stage of KB Cost and Category Constraints. The user inputs costs and categories. After getting information on tourist destinations that are in sync with tourism categories and costs. Next, we calculate predictions for family planning. We combine the results of the CF and KB rating predictions with the Hybrid calculations to get the rating prediction results. The results of the top 5 rating predictions by Hybrid are the group candidates for certainty. However, we use a Threshold user who gives a 5 rating on the destination, which will be grouped for a vacation together. Furthermore, at the Borda stage, voting for each user who gives a rating of 5 is carried out in order to get a weight for ranking by Borda. The highest Borda score is the winner for the given group recommender.

3.3 The Most Influential Group Recommender Test Results

The results of this test are obtained from the final data of the top 5 prediction results of the Hybrid user rating with a rating of 5 which are included in one group with the same tourist destination. Next, the Borda method is used to give weights. So, each user votes first to the recommended destination from (1-5). The result of this test is the attribute of the largest group to get the weight. Table 1 shows the 5 most influential groups.

Table 1. 5 Most Influential Group Recommender Tour Group

Museum Pos Indonesia

Vihara Vipassana

Graha

Museum Nike Ardila

Museum Asia

Afrika Geology Museum

Weight 137 104 185 91 143

Based on the results of ranking using the Borda method, the “Nike Ardilla Museum” group received the greatest weight, so the winner of this group's recommender system was “Nike Ardila Museum”.

3.4 Evaluation

In the evaluation stage, we use the evaluation metric Root Mean Square Absolute Error (RSME). The purpose of this evaluation is to measure the accuracy of the methods used in this study. The RMSE value can range from 0 - and the smaller or closer to the 0 value, the better the value and vice versa. RMSE evaluation is widely used to develop a recommender system model [15]. Here's the RMSE formula:

RMSE = √^{∑ i,j∈T(r}^ij^{− r̂}^ij^{) 2}

𝑇 (7)

Where 𝑟𝑖𝑗 is the actual rating of user i to item j, 𝑟̂𝑖𝑗 is the predicted rating of user i to item j while T is the number of predicted values.

At the evaluation stage of the Borda method, Borda method is a voting method that can determine group decisions and get ratings and points for each alternative [16], we compare the results of group recommender with users and the results of the difference in group recommender with an additional 25 users to prove whether the group's recommender results remain the same. In the second evaluation, we used a user satisfaction survey for the system application. group recommender in this study.

(5)

3.5 Result of Generating and Preprocessing data

The data processing in this study consisted of 97 tourist destinations, 755 users, and 72,506 ratings. The dataset is processed into two parts, the first is at the Collaborative Filtering stage, processing and modeling rating data so that rating predictions are made. The Knowledge-Based Filtering stage of the destination data is carried out first, such as removing punctuation elements, Rp, and changing the data type from string to integer so that arithmetic operations can be performed to determine Cost and Category Constraints. The results of the knowledge-based filtering prediction rating are 0-2, so normalization is needed so that the knowledge-based filtering prediction results can be combined with the collaborative filtering prediction results that are 1-5. Normalization of data is a grouping of various values so that they are on the same scale to be compared [6].

3.6 Analysis of Test Results

In this study, we apply RMSE (Root Means Square Error) to test the performance of each method, testing the group's recommender system compares the weights of the ranking results of the first Borda with the user dataset, then performs comparisons by adding 25 new users. In the final evaluation, we used a user satisfaction survey on the group recommender system application in this study.

Figure 2. RMSE of each personal method Figure 2. explain the results of personal evaluation using RMSE Evaluation

Figure 3. Results of the Group Recommender System Satisfaction Survey

Figure 3. explain the results of the percentage of responses from answers to questions that I believe in, I can easily give an assessment of what I like or don't like from the survey given

Figure 4. Results of the Group Recommender System Satisfaction Survey 0.608

1.1642 1.3672

0 0.5 1 1.5

Collaborative Filtering

Knowledge-Based Hybrid

Performance Test Results with RMSE

4% 8%

20%

32%

36%

I believe, I can easily give a rating what I like or dislike

Very Dissatisfied (1) Not satisfied (2) Quite Satisfied (3) Satisfied (4) Very Satisfied (5)

0%

0% 20%

40%

I believe these recommendations useful

(6)

Figure 4. explain the results of the percentage of responses from answers to questions that I believe these recommendations useful from the survey given

Figure 5. explain the results of the percentage of responses from answers to questions that If this application is available to the public, I would consider using this app from the survey given

Figure 6. explain the results of the percentage of responses from answers to questions that overall, I am statisfied with this recommendation from the survey given

Table 2. Group recommender system evaluation error value Tour Group

Museum Pos Indonesia

Vihara Vipassana

Graha

Museum Nike Ardila

Museum

Asia Afrika Geology Museum

Weigted 730 User 137 104 185 91 143

Weigted 755 User 172 145 205 125 165

Difference 38 41 20 34 22

Average 33

Based on Figure 2, the results of the RMSE test show that Collaborative Filtering (CF) is superior to Knowledge-Based Filtering (KB) and Hybrid. The performance test results obtained are as follows, CF 0.6080, KB 1.1642, and Hybrid 1.3672. The results of the performance test above prove that the recommender system using the CF method is more attractive to users, that the results of the varied recommender get many choices of tourist destinations, the same as a research [9]. While in the evaluation of the group recommender system, the error value obtained from the evaluation of the original users and adding 25 new users was 33, and the value of the winner of the group recommender was still won by "Nike Ardila Museum". In the next evaluation, we use a user survey proven in Figure 3, users feel satisfaction using this application to determine the rating with an average value of 3.88, in Figure 4, users also feel confident with the recommender of this application with an average value of 4, 2, in Figure 5, users consider that if they use this application with a satisfied average result of 4.24 and finally in Figure 6, users are satisfied with the performance of the given group recommender system with an average result of 4.4.

0%

0% 24%

28%

48%

If this application is available to the public, I would consider using this app

0%

0% 12%

52% 36%

Overall, I am satisfied with this recommendation

(7)

4. CONCLUSION

The conclusion of the implementation and group recommender system using the Weighted Hybrid method by combining the Collaborative Filtering (CF) and Knowledge-Based Filtering (KB) methods for the Bandung-Raya Regency domain. Based on the results of the evaluation with the RMSE performance test.. While in the evaluation of the group recommender system, the error value of the original user evaluation and adding 25 new users was 33 points, and the winner of the group recommender was still won by "Nike Ardila Museum". So, the 10th user will group with the “Nike Ardila Museum” because that group gets the highest weight in the calculation using the Borda method. Suggestions for further research to create more user input for group calculations and add different evaluation scenarios.

REFERENCES

[1] Kemenpar, “Renstra Kementerian Pariwisata 2015 - 2019,” Kementeri. Pariwisata, no. September 2015, pp. 1–100, 2015.

[2] P. J. Sheldon, “eTourism: Information Technology for Strategic Tourism Management,” Ann. Tour. Res., vol. 31, no. 3, pp. 740–741, 2004, doi: 10.1016/j.annals.2004.03.002.

[3] T. De Pessemier, J. Dhondt, K. Vanhecke, and L. Martens, “TravelWithFriends: a Hybrid Group Recommender System for Travel Destinations,” 9th ACM Conf. Recomm. Syst., 2019.

[4] S. Renjith and C. Anjali, “A personalized mobile travel recommender system using hybrid algorithm,” 2014 1st Int. Conf.

Comput. Syst. Commun. ICCSC 2014, no. December, pp. 12–17, 2014, doi: 10.1109/COMPSC.2014.7032612.

[5] J. D. C. L. Soares, Suyoto, and A. J. Santoso, “M-guide: Hybrid recommender system tourism in east-timor,” Proc. - 2017 Int. Conf. Soft Comput. Intell. Syst. Inf. Technol. Build. Intell. Through IOT Big Data, ICSIIT 2017, vol. 2018- January, pp. 303–309, 2017, doi: 10.1109/ICSIIT.2017.16.

[6] Uwaisy, M. Anranur, Z. K. A. Baizal, and M. Yusza Reditya. "Recommendation of Scheduling Tourism Routes using Tabu Search Method (Case Study Bandung)." Procedia Computer Science 157 (2019): 150-159.

[7] Baizal, ZK Abdurahman, et al. "Generating travel itinerary using ant collony optimization." Telkomnika 16.3 (2018):

1208-1216.

[8] T. Tran, “Combining Collaborative Filtering and Knowledge-Based Approaches for Better Recommendation Systems,”

J. Bus. Technol., 2007, [Online]. Available:

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.89.9245&rep=rep1&type=pdf.

[9] L. Boratto and S. Carta, “The rating prediction task in a group recommender system that automatically detects groups:

architectures, algorithms, and performance evaluation,” J. Intell. Inf. Syst., vol. 45, no. 2, pp. 221–245, 2015, doi:

10.1007/s10844-014-0346-z.

[10] G. Geetha, M. Safa, C. Fancy, and D. Saranya, “A Hybrid Approach using Collaborative filtering and Content based Filtering for Recommender System,” J. Phys. Conf. Ser., vol. 1000, no. 1, 2018, doi: 10.1088/1742-6596/1000/1/012101.

[11] K. Yoo and U. Gretzel, “Recommender Systems Handbook,” Recomm. Syst. Handb., pp. 455–477, 2011, doi:

10.1007/978-0-387-85820-3.

[12] G. Geetha, M. Safa, C. Fancy, and D. Saranya, “A Hybrid Approach using Collaborative filtering and Content based Filtering for Recommender System,” J. Phys. Conf. Ser., vol. 1000, no. 1, 2018, doi: 10.1088/1742-6596/1000/1/012101.

[13] Aprilianti, Mustika, Rahmad Mahendra, and Indra Budi. "Implementation of weighted parallel hybrid recommender systems for e-commerce in indonesia." 2016 International Conference on Advanced Computer Science and Information Systems (ICACSIS). IEEE, 2016.

[14] S. Robertson, “Understanding inverse document frequency: On theoretical arguments for IDF,” J. Doc., vol. 60, no. 5, pp. 503–520, 2004, doi: 10.1108/00220410410560582.

[15] L. Ambarwati and Z. K. A. Baizal, “Group Recommenders System Using Hybrid Filtering for Tourism Domain,” vol.

4, no. August, pp. 21–30, 2019, doi: 10.21108/indojc.2019.4.2.258.

[16] N. Nassar, A. Jafar, and Y. Rahhal, “A novel deep multi-criteria collaborative filtering model for recommendation system,” Knowledge-Based Syst., vol. 187, p. 104811, 2020, doi: 10.1016/j.knosys.2019.06.019.

[17] Mauko, Arfan, B. Muslimin, and Putu Sugiartawan. "Sistem Pendukung Keputusan Kelompok Pemilihan Saham LQ45 dengan menggunakan metode AHP, Promethee dan BORDA." Jurnal Sistem Informasi dan Komputer Terapan Indonesia (JSIKTI) 1.1 (2018): 25-34.

[18] Uwaisy, M. Anranur, Z. K. A. Baizal, and M. Yusza Reditya. "Recommendation of Scheduling Tourism Routes using Tabu Search Method (Case Study Bandung)." Procedia Computer Science 157 (2019): 150-159.

[19] El Bouhissi, Houda, et al. "Towards an Efficient Knowledge-based Recommendation System." IntelITSIS. 2021.

[20] B. Prasetyo, H. Haryanto, S. Astuti, E. Z. Astuti, and Y. Rahayu, “Implementasi Metode Item-Based Collaborative Filtering dalam Pemberian Rekomendasi Calon Pembeli Aksesoris Smartphone,” Eksplora Inform., vol. 9, no. 1, pp. 17–

27, 2019, doi: 10.30864/eksplora.v9i1.244.