Movie Recommendation Algorithm Using Social Network Analysis to Alleviate Cold-Start Problem

Abstract

With the rapid increase of information on the World Wide Web, finding useful information on the internet has become a major problem. The recommendation system helps users make decisions in complex data areas where the amount of data available is large. There are many methods that have been proposed in the recommender system. Collaborative filtering is a popular method widely used in the recommendation system. However, collaborative filtering methods still have some problems, namely cold-start problem. In this paper, we propose a movie recommendation system by using social network analysis and collaborative filtering to solve this problem associated with collaborative filtering methods. We applied personal propensity of users such as age, gender, and occupation to make relationship matrix between users, and the relationship matrix is applied to cluster user by using community detection based on edge betweenness centrality. Then the recommended system will suggest movies which were previously interested by users in the group to new users. We show shown that the proposed method is a very efficient method using mean absolute error.

1. Introduction

In recent years, information available on the internet is rapidly growing up; people need more time to select useful information. Social media use continues to grow rapidly, too. There are currently more than 3 billion people around the world using social media each month. The new Global Digital 2018 report has revealed that there are currently more than 4 billion people worldwide using the Internet [1]. Social networks are one of the most popular communications media today and it attracts millions of active users to share and comment on their photos and places with others [2]. To manage the information overload problem, the recommendation system has been developed. A recommendation system is a simple algorithm that allows users to relate and recommend items by filtering user-related data from large data. Recommender systems are software tools that make the recommendation of product or items that are appropriate for a customer’s taste based on the analysis of information of products that many customers are interested in and the customer’s and their past purchasing activity [3]. The recommendations system can help users to make decisions in multiple contexts. The goal of recommendation system is to find what’s likely to be interest to the users. Over the years, many recommender systems were introduced by researchers from different problem areas. We can utilize recommendation systems that use different techniques including collaborative filtering (CF) technique, content-based filtering technique, and hybrid technique. These recommendation systems will recommend popular items that customers liked [4]. CF method is one most successful and widely used in recommendation systems [5]. The basic idea of this algorithm is to use common experiences or similar interests. The important process is to search users that are similar to the target user or find the product that is similar to the predicted product. But, the CF algorithm also has some problems such as a cold-start problem [6]. Cold-start problem is a problem where the system is not able to recommend items to users. For every recommender system, it is required to build a user profile by considering the user’s preferences and likes. The user profile is developed by considering her activities and behaviors they perform with the system. On the basis of user’s previous history and activities, the system makes decisions and consequently recommends items. The problem arises when a new user or new item enters the system, for such users or items that the system does not have enough information to make decision. For example, if a new user has not rated some items and not yet visited/viewed some items, then it would be difficult for the system to build a model on that basis. This can lead to inaccuracies in estimating similarities between users. Many approaches have been studied to solve the existing problem. Embarak [7] suggested two types of recommendation such as node recommendation and batch recommendation, and then compared the suggested method with three other alternative methods including Naïve Filterbots Method, Media Scout Stereotype Method, and Triadic Aspect Method to solve the cold-start problem. Basiri et al. [8] suggested a new hybrid approach, which focuses on improving performance under cold start problem. This method can give a reasonable and appropriate recommendation.

With the development of technology, the user’s behavior and personal information can be tracked and recorded on social networking sites or online shopping sites. This type of technology makes it easier and it is very useful for analyzing user preferences. We analysis social network analysis (SNA) methods [9,10] and introduced the betweenness centrality in SNA into a CF approach. This paper presents a movie recommendation algorithm using SNA to alleviate the cold-start problem. In the proposed method, the user’s personal information such as age, gender, and occupation are used to establish a relationship between users. Then the relationship matrix between users will be applied for clustering the user into several communities or groups. In this process, the centrality of SNA is used to detection communities; after that, the system will recommend movies in the group that is similar to the target users by considering CF. The main objective of this article is to develop techniques that can recommend the most suitable movies for target users based on personal characteristics.

The proposed work is briefly described as follows. In Section 2, we will explain about the related work which is the methods used in this paper. Our proposed algorithm for movie recommendation algorithm using SNA to alleviate cold start problem will describe in Section 3. In Section 4, will present experimental analysis and experimental results and finally, the conclusion will be concluded in Section 5.

2. Related Works

In this section, we have a purpose to briefly explain about the relevant research, including the recommendation system, CF, and SNA that are required for the movie recommendation system in this study.

2.1 Recommendation System

In recent years, the recommender system has become more popular. It has been used in many areas including book, news, movie, music, and products. There are also recommender systems for experts [11], restaurants, collaborators [12], garments, jokes, financial services [13], romantic partners, life insurance, and Twitter pages [14]. These recommendation systems use one type of filtering to predict ratings and user satisfaction, which allows users to purchase products based on their interests or needs. Having information about user’s life can give hints about how the user will react when faced with different situations [15]. The recommender system is a useful alternative of search algorithms because it can help users discover what they may not find. It is usually performed using a non-traditional indexing search engine. A recommender system is a technology that makes automatics predictions about the relationship between customers or between items and searches for items that users may need. There are many approaches to make a recommender system, including the following.

● Content-based filtering: Content-based filtering methods depend on the item’s features and user preferences [16]. These methods work with data that the user provides. Based on this data, a user profile will be created which will be used to give advice to users. As the user provides more input and accepts recommendations, the engine becomes more and more accurate. These algorithms try to recommend the products that are similar to those that the user liked in the past. In particular, many nominated the products are compared to the products that were previously ranked by the user. This method provides the foundation for data retrieval and data filtering research.

● Collaborative filtering: CF method is used to automatically predict or filter information about user interests by collecting settings or tasting data from multiple users (collaborating). The basis of the CF method is assuming that if user U₁ has the same taste with user U₂, user U₁ tends to have the taste of U₂ on issues that differ from those of random users [17]. Collaborative recommender systems receive a list of recommended items by analyzing similarities between users and predicting user ratings of an item based on similar user ratings on the same list [18].

● Hybrid recommender system: Hybrid methods can be used in many ways. This approach can be making content-based filtering and collaborative-based filtering predictions separately and it can combine them together. It can be used to enhance content-based capabilities through a collaborative-based approach. It can combine all of them into one format [19].

The purpose of many studies about the referral system is to focus on the ability to recommend products that satisfy customers primarily. CF is the most commonly used method for identifying similarities between items.

2.2 Collaborative Filtering Algorithm

The CF method is derived from collecting and analyzing large amounts of data about user behavior, activities or user preferences, and predicts which users will be similar to other users. To understand what CF is, one can think of a simple question; for instance, if a person wants to read a book, but that person does not know which book is good, what will that person do? Usually, most people like to ask friends to see which books are good. We like to receive a suggestion from friends or people who have the same taste as us. This is the main idea of the CF method [20]. It can be separated into two types: user-based and item-based CF. Item-based CF depends on the similarity between the items calculated using the ratings of people for those items. For example, when users who like item I₁ also like item I₂, the similarity between these two items is are considered similar. User-based CF takes advantage of the similarities between users in the forecast. For example, if Mr. James and Mr. Paul have seen the same movie and they are also giving the same ratings, the similarities between them is 1. On the other side, assuming that they give the different rating to the item, the similarities will decrease as differences.

In this paper, we adopted a user-based CF algorithm. The algorithm predicts that if users’ personal characteristics are similar, then their interests in products or items are also similar. The algorithm searches for the most similar users according to the target user information. Based on the most similar interests or preference, a user’s interest can be predicted. The recommender system will carry out information for suggest for relevant users. Cosine similarity is used to measure the similarity of users. Its formula [21] is shown in Eq. (1).

\(\operatorname{sim}(A, B)=\frac{\sum_{i=1}^{n} A_{i} B_{i}}{\sqrt{\sum_{i=1}^{n} A_{i}^{2}} \sqrt{\sum_{i=1}^{n} B_{i}^{2}}}\)       (1)

In Eq. (1), A and B are two different vectors. A_i is the component of vector A and Bi is the component of vector B, respectively. The result of similarity ranges from −1 to 1. The negative value is the opposite; it refers to two different vectors. A positive value represents two similar vectors. With 0 indicating orthogonally or de-correlation while in-between values indicate intermediate similarity or dissimilarity.

2.3 Social Network Analysis

The SNA is a method used to analyze social network properties. It characterizes networked structures in terms of nodes, such as individual actors, people, or things in the network. Vertices indicate objects or entities while edges indicate links to show relationships or interactions. Both objects and the links may have attributes. Networks are constructed from general, real-world data. They propose several unexpected challenges pending to data domains themselves, e.g., information distillation, pre-processing, and data structures used for displaying knowledge and storage [22]. Social networks represent relationships that exist within the community. Social networks provide tools for collaborative education, especially through the theory developed in analyzing social networks [23]. Even in the same community, there may be various types of social networks depending on social relationships as friends, mutual support and cooperation. The similarity is a common standard used to construct social relationship components of a community. Actors or nodes in social networks can be individuals, groups, objects, organizations or events, as long as certain relationships remain together. Centrality is an agent indicator used in SNA. There are many types of centrality including degrees’ centrality, closeness centrality, and betweenness centrality.

Betweenness centrality is the measure of the center of the graph based on the shortest path. For all vertices in the connected graph, there is at least one shortest path between the vertices, which is each number of edges that pass through or the sum of the weight of the minimized edge. The betweenness centrality for each vertex is the number of shortest paths that pass through the vertex. Girvan and Newman [24] have presented a community detection algorithm in social networks and biological networks based on edge betweenness to avoid the flaws of hierarchical clustering methods. This algorithm continuously detects communities by removing the edges from the original network, and the connected components of the network that remain are communities. Instead of trying to create a measure that tells us which edge is most central to the community, but this algorithm focuses on the edge that seems to be “between” communities. The algorithm steps for community detection are summarized as Fig. 1.

Fig. 1. Community detect based on edge betweenness algorithm.

An example in Fig. 2 shows the effective algorithm Girvan-Newman for edge community detection based on betweenness.

Fig. 2. Community structure: (a) input graph, (b) computed betweenness values for the edges in graph, (c) at the end of iteration 1, remaining edges in the graph after removed edge 4–6 with highest betweenness score of 25.0, (d) at the end of iteration 6, remaining edges in the graph after removed edges 6–7 with betweenness score is 7.0, (e) at the end of iteration 7, remaining edges in the graph after removed edges 3 and 4 with betweenness score is 7.0, (f) at the end of iteration 10, remaining edges in the graph after removed edge 0–2 with betweenness score of 5.0, (g) at the end of iteration 11, remaining edges in the graph after removed edges 5–6 with betweenness score is 5.0, (h) at the end of iteration 12, remaining edges in the graph after removed edge 8–10 with betweenness score is 5.0, and (i) at the end of iteration 19, final partitioning of the network graph into communities.

3. Movies Recommender Algorithm Using Social Network Analysis to Alleviate Cold-Start Problem

The process of data collection and workflow processes that are sufficient for the recommended system are shown in Fig. 3. ① The system needs to collect user data and movie listings into the database used as a test dataset. ② The system requests that new users log in if they want to join. ③ The system needs to collect new user data and movie listings into the database used as a test dataset. ④ In this process we make a relationship table for a user based on their personal propensity such as age, gender, and occupation; we call this table an adjacency matrix. ⑤ The relationship between the user table or adjacency matrix is used to evaluate a community out of several communities (groups) by analyzing the relationship table of the user and comparing it using community detection based on edge betweenness. ⑥ In this process, after a community (group) of the user has been selected for evaluation, then we modify the group for the new user by computationally comparing the similarity of the new user to other users in each group from their personal information. For computing, the similarity of user cosine similarity is used. ⑦ After that we match a group with the new user, the group with the highest similarity will be selected as a group of new users. ⑧ The movies that were watched and rated by users in the matching group of the new user will be arranged in the order of popularity. ⑨ In this process, the recommended system will select the top 5, 10, 20, 30 and 40 most popular movies and finally ⑩ the system will choose the most popular movies that were watched by the members in the group recommend these to the new user.

Fig. 3. Movie recommender system configuration diagram.

3.1 Detail of Proposed Processes

3.1.1 New user provides personal information

To recommend the best movies for users, we need to collect some necessary personal information about these users. Users who have the same behavior or same personal propensity may like the same item. Personal propensity that we use for performance in this research includes age, gender, and occupation.

3.1.2 The relationship between users table

In our proposed algorithm, the centrality of SNA was applied. Therefore, we have to make a relationship table between users according to their personal propensity including age, gender, and occupation. We assume that if the users have the same personal information it means they have a relationship to each order. This table can be formalized as a classical mathematical relationship that can be seen as an unspecified graph.

3.1.3 Community detection based on edge betweenness.

After getting the relationship between users table, in this process, we want to cluster users into several groups by applying the relationship between users table. Community detection based on edge betweenness is used for clustering users, a community representation as a group. The concept of detection community based on edge betweenness is that the possibility those edge connections separate modules have the highest betweenness value because the shortest path from one module to another module must cross through those modules. Therefore, if we gradually remove the edge with the highest edge betweenness value, we will get a hierarchical map or a rooted tree, called a dendrogram of the graph. The leaves of the tree are the person’s vertices and the root of the tree means the whole graph in the network.

3.1.4 Cosine similarity and Collaborative filtering

Traditional methods to measure the similarity of users just consider the similarities of user ratings. In reality, the similarity of users is not only linked to ratings for items, but is also linked to the preference for certain item categories, that is, user interest for the item category feature. In addition, if two users have similar personal information, these two users are considered highly similar. Therefore, our research modified the group for the new user by using the cosine similarity measure. After users are divided into several groups by community detection based on edge betweenness, then the similarity between the new user and the other users in each group is computed by filtering personal information including age, gender, and occupation. For computing the similarity between users Eq. (1) is used.

3.1.5 The matching group for the new user

After completing the calculation of similarities between new users and other users in each group, we compared the average similarities of new users and each group. The group with the highest similarity was selected as the most similar group for related users.

3.1.6 Ranking the popularity movies

When finding a group that was most similar to a new user, movies watched by group members will be ranked according to their popularity, which was counted from the rating that each member of the group had given to each movie.

3.1.7 Recommended movies to the new user

After ranking popular movies from users that were similar to new users, the most popular movies were selected to be recommended for new users. However, the final decision about which movies the user will watch will depend on the new user.

In this article, we try to combine existing CF techniques with SNA. We use between centrality identification method and the introduction of a movie recommendation system that may give the best predictions about the movie program that the target user might be interested in. The algorithm used in our recommendation system is shown in Algorithm 1.

Algorithm 1. movies recommender system using social network analysis and collaborative filtering

4. Experimental Analysis

4.1 Experimental Dataset

In the experiment in this research, we used the MovieLens dataset provided by the GroupLens research group at the University of Minnesota, USA [25]. It comprises data from three orders of scale. Each data set has related user’s information, user’s ratings, and the movie’s information. This dataset consisted of 100,000 ratings; the rating is in range of 1 to 5. This dataset includes 943 users and 1,682 movies. Each user gives a rating to movies at least 20 movies. This dataset also has simple information for each user, such as gender, age and occupation.

To evaluate the quality of our proposed method, the dataset is divided into two parts: the training and testing set. It is very important to evaluate performance using data not involved in formulating the model. To improve the accuracy of the recommendation system, both sets of dataset included 10 random datasets from users who gave ratings to at least 20 movies, users who gave ratings to at least 50 movies, users who gave ratings to at least 100 movies, and users who gave ratings to at least 200 movies. This means we have to implement the method a total of 40 times.

Training dataset is used after a model has been processed. In the setting of this recommender system, partitioning is performed by randomly selecting some users and some ratings from all users. There are 800 users in this dataset. Training set is implemented to build up a model. Testing dataset is used to test the model by making predictions. There are 143 users in this dataset. Some users and some ratings are then randomly selected from all users.

Fig. 4 shows part of the user’s personal information provided by MovieLens. Fig. 5 displays a section of rating and matrix data generated from rating data by users who watch certain movies. In Fig. 5, the columns are user numbers or user IDs, while the rows are movie numbers or movie IDs. Table 1 shows the result of user relationships after comparing user’s personalization information.

Fig. 4. MovieLens’ user information part

Fig. 5. Users–movies matrix part.

Table 1. The relationship table between users​​​​​​​

4.2 Experimental Environment

Hardware and software used to evaluate the methods proposed in the paper are shown in Table 2.

Table 2. Resources​​​​​​​

4.3 Experimental Results

Several metrics have been proposed to evaluate the accuracy of the CF method algorithm. Mean absolute error (MAE) is one of the most commonly used tools for measuring the accuracy of the recommender system [26]. MAE evaluates the accuracy of a prediction algorithm by comparing numerical deviation of the predicted rating from the respective actual user rating. Formally, if n is the number of an actual item that is purchase by a target user and MAE is assigned as the mean absolute difference between the pair. And assuming that the predicted rating set of the target user u is {p_u1, p_u2, ..., p_uN} and actual rating set is {r_u1, r_u2, ..., r_uN}, then MAE is defined as follows:

\(M A E=\frac{\sum_{i=1}^{N}\left|P_{u i}-r_{u i}\right|}{N}\)       (2)

After performing the proposed method, we want to predict the accuracy of our proposed method. We want to know when our system recommended movies, how many movies were actually watched from the user. MAE can help measure the level of satisfaction and evaluate the accuracy of the recommender system. Normally, the lower the value of the MAE, the higher the accuracy of the recommendation. To the computation of MAE, the Eq. (2) is used.

As described in Section 4.1 we performed experiments in four cases: random testing datasets from users who gave ratings to at least 20, 50, 100, 200 movies. To reduce inaccuracy of the MAE values, we performed 10 experiments using each dataset. In total, of an experiment are 40 times. The average results of the movie recommendation system using SNA and CF for 10 experiments are shown in Fig. 6.

Fig. 6. The result of social network analysis (SNA) and collaborative filtering (CF) method.

Fig. 5 shows the average of all experiments from 10 random times; the best result for the number of movie recommendations was 5. Therefore, we averaged the result from four cases; the average result from four cases was 3.55 when 5 movies were recommended. The average result from four cases was 3.79 when 10 movies were recommended. The average result from four cases was 4.34 when 20 movies were recommended. The average result from four cases was 4.59 when 30 movies were recommended. The average result from four cases was 4.78 when 40 movies were recommended. The maximum MAE value is 10. Therefore, we can say that our proposed method is very effective and can solve the cold-start problem. From these results, we can also interpret that the user was interested in more than 3 out of 5 movies recommended by the system.

In order to confirm the effectiveness of the methods that are presented in this paper, we have compared with other methods including density-based on clustering [27] method, the CF with k-NN, and CF method.

The results of the movie recommendation system using density-based on clustering by used the same dataset as a proposed method are shown in Fig. 7.

Fig. 7. The result of density-based on clustering method.

In Fig. 6, the average result from four cases was 4.19 when 5 movies were recommended. The average result from four cases was 4.76 when 10 movies were recommended. The average result from four cases was 5.28 when 20 movies were recommended. The average result from four cases was 5.44 when 30 movies were recommended. The average result from four cases was 5.63 when 40 movies were recommended. The maximum MAE value is 10. The best number of movies recommended in this method was also 5.

Fig. 8 showed the results of the movie recommender system using k-NN and CF, the average result from four cases was 3.60 when 5 movies were recommended. The average result from four cases was 3.96 when 10 movies were recommended. The average result from four cases was 4.35 when 20 movies were recommended. The average result from four cases was 4.61 when 30 movies were recommended. The average result from four cases was 4.85 when 40 movies were recommended. The maximum MAE value is 10. The best number of movies recommended in this method was also 5, as in previous methods.

Fig. 8. The result k-NN and collaborative filtering method.

Traditional datasets that are not tuned have been tested using traditional CF algorithm. For the movie recommendation system based on original CF, the neighbor for the user who has the same taste as the new user was found. The result of the movie recommender system using CF is shown in Fig. 9. When 5 movies were recommended the result was 5.75. When 10 movies were recommended the result was 5.84. When 20 movies were recommended the result was 6.06. When 30 movies were recommended the result was 6.29. When 40 movies were recommended the result was 6.48. The best number of movies recommended using this method was also 5; same as other methods.

Fig. 9. The result of collaborative filtering method.

After that, we will show the efficiency of each method by comparing the total average of the MAE results of the four methods. The total average result of the movie recommendation system that is proposed to use k-NN and CF shows better results than the total average result of the movie recommendation system that is proposed to use density-based clustering. The k-NN and CF method also shows better results than the results of the movie recommendation that is proposed to use original CF. The total average result of the method we propose, which is a movie recommendation system that is proposed to use SNA and CF, is more accurate than the other three methods as shown in Fig. 10. We also have to argue that the efficiency of our method is better than the other three methods. We also have to argue that the efficiency of the three methods is the best of our method.

Fig. 10. Comparative the result of the experimental.

5. Conclusions and Future Work

This paper aims to solve the problem of CF by using SNA in the recommender system. We design and implement in R programming which is an open source scripting language for predictive analytics and data visualization. The recommendation system is a way to help the users to find the information that they want easily. CF is one of the most popular uses one and a successful method used in the recommender system. However, it has weakness such as cold start problem. To overcome this problem, in this paper we proposed an alternative approach for the recommender system using both SNA and CF. We found the community or group for the user based on edge betweenness centrality. The method that we proposed here is very effective for making movie recommendations. Analyzing results showed that the total number of 20 movies recommended was better than 40 and the best number of movies recommended was 5. In addition, the method presented in this paper showed the best performance, followed by k-NN and CF, Density-based clustering, and CF.

However, the implementation processed in this paper by using R programming took a very long time. Therefore, we propose to reduce the experiment time and to improve the accuracy and effectiveness of the recommender system by applying various types of datasets in future research.