Place Recommendation from Check-in Spots on Location-Based Online Social Networks

Transcription

1 2012 Third International Conference on Networking and Computing Place Recommendation from Check-in Spots on Location-Based Online Social Networks Chen Hongbo, Chen Zhiming, Mohammad Shamsul Arefin, Yasuhiko Morimoto Graduate School of Engineering, Hiroshima University {m114898, m115395, d105660, Abstract With rapid growth of the GPS enabled mobile device, location-based online social network services become more and more popular, and allow their users to share life experiences with location information. In this paper, we considered a method for recommending places to a user based on spatial databases of location-based online social network services. We used a user-based collaborative filtering method to make a set of recommended places. In the proposed method, we calculate similarity of users check-in activities not only their positions but also their semantics such as shopping, eating, drinking, and so forth. We empirically evaluated our method in a real database and found that it outperforms the naive singular value decomposition collaborative filtering recommendation by comparing the prediction accuracy. I. INTRODUCTION During the past few years, online social network services such as Facebook and Twitter have been developed rapidly. Until today, more than half billion users around the world are using these services. At the same time, with the growing prevalence of GPS enabled mobile devices, such as smart phones like Apple iphones and Google Android phones; Locationbased Online Social Network Services (LOSNS) have become extremely popular, which helps users to create, share, and communicate with other users by attaching their geographical location information. By using check-in function of LOSNS, users can easily publish their feelings and expressions with spatio-temporal information. For instance, Facebook Places is one of these kinds of check-in services. Through the LOSNSs, we can collect a spatio-temporal database containing hundreds of millions of users check-in records. By analyzing the database, we can extract information such as what the users are interested in, where the users like to go, and what the users often do. In this paper, we considered a method for recommending places to a user based on the spatial database of LOSNS. We used a user-based collaborative filtering method to make a set of recommended places for a user, in which we calculate similarity of users and use the similar users records to predict places the user likes. In the proposed method, we calculate similarity of users check-in activities not only their positions but also their semantics such as shopping, eating, drinking, and so forth. We propose a semantic hierarchical category-graph of check-in spots to measure the similarity score between users. The contributions of this paper are as follows: (1) we analyze a real-world check-in spots to get semantic location information for calculating similarity of users interests, and (2) recommend the unvisited places. We, first, make clusters of the check-in records based on the position. Next, for each cluster, we calculate the gravity center of members positions to represent the position of the cluster. We, then, annotate each cluster by using Points Of Interest (POIs) database. If the nearest POI from the center of a cluster is a movie theater, which belong to entertainment category, we annotate words that are related to movie theater and entertainment to the cluster. After we annotated such semantic information to clusters, a semantic hierarchical category-graph framework will be applied to analyze users interests and measure the similarity score between others. We select Top-N similar users and use the users records for user-based collaborative filtering. Then, we recommend some unvisited places. The remaining paper is structured as follows: In section 2, we discuss the related works on location recommendation. In section 3, we describe the check-in records from a real-world LOSNS website. In section 4 we will discuss how to make clusters of the check-in records and how to assign the semantics of each cluster. In section 5, we use the Collaborative Filtering (CF) algorithm to predict ratings of unvisited places for making a set of recommended places. Then, the section 6 will give the experiments and evaluation methods to evaluate our frame have high predictive performance. II. RELATED WORK Location recommendation is a popular service in social network and at present, there are many systems for providing location recommendation. Adomavicius et al. [1] first introduced the idea of collaborative filtering based recommendation. In collaborative filtering based approach, an item is recommended to a user based on past information of the people with similar tastes and preference. Zheng et al. [7] proposed a system by using the location data based on GPS and users comments to discover interesting locations for recommendations. Zheng et al. [8] proposed two types of travel recommendations by mining multiple users GPS traces. The first type tried to recommend a user with top interesting locations in a given geospatial region, and the second type is a personalized one that provides an individual /12 $ IEEE DOI /ICNC

2 with locations matching the user s travel preferences. Zheng et al. [9] introduced a social networking service, called GeoLife, which can understand trajectories, locations and users, and mine the correlation between users and locations in terms of user-generated GPS trajectories. Zheng et al. [10] moved their direction towards on a personalized friend and location recommender for the geographical information systems on the web. In their paper they built a hierarchical graph to model each user s location history and measure the similarity among users. All of [7, 8, 9, 10], were mining users similarity from their individual GPS trajectory. The process of these papers is similar to ours, as both of us try to find the similarity between users by their location history firstly. They identified the stay points from the GPS trajectory and clustered the stay points. Then they matched clustered sequences to get similar users. However, this kind of framework paid more attention to users physically location, the higher user similarity means two users are physically close such as neighborhoods. In general, people do not think neighborhoods are similar to each other. Our proposed method do not only focuses on mining users physically location, but also including their interests. Berjani et al. [4] proposed a personalized recommender for location-based online social networks, based on the check-ins of the users. They crawled check-in information to generate a user/spot rating matrix. By predicting the interest of users in certain spots, their technique recommends places they have not been visited previously. This approach is useful to find the interest of users in certain spots. In [6], Takeuchi et al. [6] proposed a novel real-world recommendation system based on users individual preferences and needs. This system automatically figures out each user s frequently visited spots and makes recommendations by using the spots as input to the item-based collaborative filtering algorithm. McDonald et al. [12] proposed an expert locating system that recommends people for possible collaboration within a work place. Chow et al. [5] presented GeoSocialDB, a holistic system, providing location-based news feed, location-based news ranking, and location-based recommendation. Within their framework, they discussed the research challenges and directions towards the realization of location based social networking services. All of the above works did not consider the issue of the semantics of GPS location. On the other hand, in our proposed method at first the names of the noted places are attached to GPS location data. Then, we have built hierarchical categorygraph framework and transactional database. Moreover, we have shown that our method can provide recommendation by applying a typical CF approach that was not considered in earlier works. III. DATA SAMPLE In this paper, we used SINA microblog as data source to collect users check-in spots. SINA microblog is a Chinese microblogging website. It is one of the most popular LOSNS in China, which is used by over 30% of Internet users, and has more than 300 million registered users as of February Additionally, users are also allowed to insert graphical Fig. 1. Examples of the check-in spots dataset. emoticons or attach own image, music, video files and checkin location information in every post. We extracted the user s unique identification as attribute userid, blogs identification as blogid, and geographical position as <latitude, longitude> to constitute our checkin spots dataset. The schema of our check-in spots dataset is shown in Table I. Examples of the check-in spots dataset are shown in Figure 1. TABLE I THE SCHEMA OF DATESET parameter type userid varchar(10) blogid varchar(15) latitude float longitude float When we perform the check-in function, we can add some short sentences to express our feelings. The blogid of the check-in spots identifies the documents that was posted with the check-in. Though we do not use the blog information, we add it for future use. IV. INTERESTS DISCOVERY AND SIMILARITY CALCULATION In this section, we will propose a framework to discover interests and measure the similarity score between users in terms of their check-in spots. This framework refers to semantic hierarchical category-graph framework, which models people s check-in spots and estimates the similarity between users effectively. A. Interests Discovery 1) The Shared Framework: To learn users interests, we need to analyze the check-in spots dataset. We first use a density-based clustering method to cluster all of users checkin spots into several regions. Examples of the clusters are shown in the layer 1 of Figure 2 (a). In the figure, C 1, C 2,..., C 15 are the clusters. The reasons for clustering the check-in spots into regions are as follows: (1) Users may visit one place several times and leave a lot of GPS records in the place. For example, users may frequently visit a famous landmark and may check-in each time they visit. (2) GPS data that are very close to each other are stand for the same place. For example, users may check-in different positions of a park, although their GPS data may be different. Hence, before semantic location discovering, we have to do this preprocessing. 144

3 Fig. 2. The Semantic Hierarchical Category-graph Framework (a) The shared framework. (b) A personal framework In the next step, for each cluster, we calculate the gravity center to represent the position of the cluster. We, then, annotate each cluster by using POIs database. If the nearest POI from the center of a cluster is a movie theater, which belong to entertainment category, we annotate words that are related to movie theater and entertainment to the cluster. Then, we construct a Semantic Hierarchical Category-graph Framework (SHCF) as in Figure 2 (a). This hierarchical framework contains three layers: In the bottom layer, there are clusters which relate to real-world check-in spots are distributed. In the middle layer, we assign categories reflecting clusters semantics. And in the top layer, we assign more general categories to the categories in the middle layer. The SHCF aggregates users check-in records. So, we can consider it as a shared framework. Based on this shared framework, each user can build his/her own personal framework. Figure 2 (a) is an example of the SHCF. Users check-in spots are aggregated and clustered into 15 different geographical clusters as C 1 to C 15 in layer 1. All precise categories reflecting clusters semantic classification lay in layer 2 and all general categories are in layer 3. From Figure 2 (a), we can see that cluster C 13 at layer 2 belongs to category Baseball, while its general category is Sports in layer 3. 2) A Personal Framework: Based on the shared framework, each user can build a personal framework, Personal Semantic Hierarchical Category-graph (PSHC). Figure 2 (b) gives an example of a PSHC. In the figure, clusters and categories are came from Figure 2 (a). Oval pins stand for a user a s checkin spots. As we can see from the user s personal framework, there are 4 cluster nodes that belong to KTV category and there are 3 cluster nodes that belong to Bar category. Therefore, we can imagine that the user is interested on KTV and Bar, and he/she might like entertainment very much. Those hierarchical features are used to differentiate users in different degrees of similarity. For example, the users who share the check-in spots in same clusters on the bottom layer might be more correlated than those who share interests in same categories on a higher layer. B. Similarity Calculation Similarity calculation is an important step for finding users potential interests. We find top-n similar users and use the users records for user-based collaborative filtering. In this step, we use the SHCF as mentioned before, and calculate the similarity weight between users layer by layer. At the end, after adding each layer s similarity weight, we can get the similarity score. 1) Inverse Document Frequency: In order to normalize the popularity of different locations, we use the the idea of inverse document frequency (IDF) that is used for document analysis. If the location A is not popular and visited by a few people and the location B is popular and many people visited, then the co-occurrence of A is stronger than that of B for calculating similarity in the idea of inverse document frequency. A cluster can be regard as a document, while the users who have checked in this cluster can be as terms. If a large number of users (n i ) that checked in a cluster (c j ), the IDF j = log U n i of this cluster would become very small. Therefore, this cluster will not offer many contributions to the similarity score of these two users on the cluster layer. 2) Similarity Weight Measure Of Bottom Layer: To measure the similarity between two users in layer 1, we use a popular Pearson Correlation approach. Suppose N = C j clusters are generated on the bottom layer (cluster layer) in our SHCF. Assume that two users u a and u b visited m j and m j times in a cluster C j. Let w j and w j be the ratings of u a and u b to C j. We can compute ratings w j and w j with the help of the equations (1) and (2); w j = m j IDF j, (1) w j = m j IDF j, (2) 145

4 So, two vectors can be respectively constructed for U a and U b as follows, U a =<w 1,w 2,...,w j,...,w N >, (3) and U b =<w 1,w 2,...,w j,...,w N >. (4) The Pearson similarity, which have been widely used in recommendation systems, is computed as: SimW eightcluster a,b = sim pearson (u a,u b ) j = (w j Ūa)(w j Ūb), (5) j (w j Ūa) 2 (w j Ūb) 2 3) Significance Score of Category Propagation: In general, if we find that a user frequently visits different coffee shops and seldom visits KTV houses, we can consider that the person likes drinking coffee more than singing in KTV. The similarity weight of category layers are computed as: SimW eightcategory a,b = 1 h 2 sim pearson(u a,u b ) h = i (SS a(i) Ūa)(SS b (i) Ūb). (9) h h 2 i (SS a(i) Ūa) 2 (SS b (i) Ūb) 2 5) Similarity Score Measure: SimScore a,b denotes the similarity score between user a and b on all of the shared SHCF layers, it is calculated as follows: SimScore a,b = SimW eightcluster a,b +SimW eightcategory a,b. (10) V. PLACES RECOMMENDATION After we estimated the similarity between users, a userbased CF methods are employed to infer users interests in unvisited places. We used the profiles of a cluster by exploring the knowledge from POIs. We could recommend different types of places (clusters) matching the users preferences. In addition, by using the profiles of users and the semantic meaning of places, it is also able to combine the contentbased method with collaborative filtering to reduce the cold start problem of recommender systems. Fig. 3. The significance score of different categories. As in Figure 3, if many clusters, which are checked in by a user, belong to category X and few clusters belong to category Y, then the significance score of category X is bigger than category Y. To formulize this, we let SS a (x) be the significance score of category node x of a user a. It is calculated as follows: SS a (x) = Category a (x) totalnumberofcheckinspots a, (6) where Category a (x) is the number of check-in spots belong to category node x s sibling nodes, totalnumberofcheckinspots a is the total number of check-in spots of user a. 4) Similarity Weight Measure Of Upper Layers: Considering a user u a as a vector U a R i, where Ua i = SS a (i) is user u a s category score of category i. Supposing k is the total number of category on layer h of the shared SHCF. So, two vectors U a and U b can be respectively constructed for user u a and u b as follows: U a =<SS a (1),SS a (2),...,SS a (i),...,ss a (k) >, (7) and U b =<SS a (1),SS a (2),...,SS a (i),...,ss a (k) >. (8) A. Users Similarity Matrix After we have measured the similarity score between two users, we then formulate a similarity matrix (SM) for all the users in a community, where SM = SimScore j,k, 1 j U, 1 k U,j k, j, k are two different users; U is the total number of all users. From Table II, given a user (u j ) as a query, we can rank other people in this community (u k U, j k) according to their similarity score (SimScore j,k )tou j. Then, a group of people (U ) with relatively high similarity scores can be retrieved as similar users for u j. B. Similar Users and Unvisited Places Discovering When a query u j has been given, we retrieve from the SM the vector v j containing the similarity scores between u j and others, where V j = SimScore j,k,k U,j k. Then, SimScore j,k will be normalized by Equation: SimScore j,k = SimScore j,k min(v j ). (11) Max(v j ) min(v j ) Thus, the top N users with high SimScore j,k will be retrieved as u j s similar users (U j ). After we retrieve u j s similar users U j, we are more likely to discover some regions (places) that might interest u j but have not been visited by u j. And we define this set of places discovered from U j as candidate places C j of u j. 146

5 TABLE II SIMILARITY MATRIX. SimScore u a u b... u k... u u u a 1 SimScore a,b... SimScore a,k... SimScore a,u u b SimScore b,a 1... SimScore b,k... SimScore b,u u j SimScore j,a SimScore j,b... SimScore j,k... SimScore j,u u u SimScore u,a SimScore u,b... SimScore u,k... 1 C. CF-based Recommender Before we recommend places to users, let s see a significant problem of our hierarchical category framework. For instance, in Figure 4, we assume user u b is one of the similar users of u a. We can imagine that, u a may be interested in regions (places) C 6, C 11, C 14 which are predictd from u b s data. However, u a has never been there. Because we have already know the semantic information of these places, C 6 is a KTV house and C 14 is a Bar. User u a really like these kinds of places. On the other hand, C 11 is a bookstore and u a has never checked in this category. Although, it is correct to recommend C 11 to u a, he/she may not like go to there indeed. where SS q j is the significance score of category propagation of region q, as formula (6). ˆr q j comes from the user-based prediction, it likes: ˆr q j = r k U j + SimScore j,k(r q k r k) k U SimScore, (13) j,k r j = 1 r c C j,rj c 0. (14) c C As we can see from equation (13), SimScore j,k is the similarity score between users u j and u k, is calculated by using our SHCF and normalized on formula (11). Here it is used as a weight. It means the more similar u j and u k are, the more weight r q k will carry in the prediction of ˆrq j. In fact, the equation (13) and (14) shown above are well known by many papers and have been used widely in many recommender systems. Hence, we do not explain them in more detail. VI. EXPERIMENT We use the real-world check-in spots dataset to make an offline experiment to evaluate the quality of the recommendations of our framework. Fig. 4. Personal interests distribution hierarchical graph of user a and b. To improve the recommendation accuracy, we plus the significance score of category propagation for predicting the rating of each candidate places. After candidate places C j discovered for a particular user u j, CF-based methods are employed to infer the right places to the user. In our SHCF, the equation describe the process predicting u j s rating r q j onaregionq( C j ) like this: r q j = SSq j +ˆrq j, (12) A. Experimental Setup We used SINA microblog, which is one of the most famous location-base online social network website in China, in this experiment. We collected 30,515 check-in spots over 10,827 users in a city. Amond them, we select 268 users who checked in more than 25 times. The number of check-in records of those 268 users is 15,713. We used a density-based clustering algorithm called DB- SCAN (Density-Based Spatial Clustering of Applications with Noise) to cluster the check-in spots. In our clustering analysis, we set the neighborhood radius ε equals actual distance 50 meters, and each cluster contains at least a minimum number, MinPts equals 1. Using these parameters, DBSCAN algorithm, made 10,476 clusters on the bottom layer. After the clustering, we use Foursquare (one of most spotlighted location based social network services) database as our POIs data to annotate clusters. Examples of clusters semantic information shown are shown in Table III. B. Evaluation To evaluate prediction accuracy of the proposed recommendation we examined the Precision and Recall of recommended places and compare the accuracy with a naive 147

6 TABLE III EXAMPLES OF CLUSTERS SEMANTIC INFORMATION Cluster Spots LocationName Category FatherCategory 1 6 Return97 Bar Nightclub Entertainment 2 15 KFC Fast Food Food 3 22 Starbucks Cafe Food 4 3 Wangfujing Bookstore Bookstore Education 5 13 Mcdonald s Fast Food Education Disneyland Park Entertainment Singular Value Decomposition Collaborative Filtering (SVD- CF) recommender which is a popular and advanced method in recommendation system. Figure 5 shows the results. From the results, our SHCF works effectively in recommendation, and it outperforms the SVD-CF for every various recommender places from different top-n similar users. Our framework is 21.5% higher in the precision and 56.7% in the recall than the SVD-CF on the average. Fig. 5. Experimental results with various recommender places from different top-n similar users. VII. CONCLUSION In this paper we have proposed semantic hierarchical category-graph framework, an effective framework to deal with the real-world location-based online social network users check-in information and measure similarity score by analyzing each individual interests. Through analyzing users interests and extracting the check-in spots from similar users, we recommend the unvisited places. The result of the experimental evaluation shows that our framework can work effectively with collaborative filtering algorithm for achieving our recommendation. As a future work, we would aggregate the user s current location by calculating the real physical distances between candidate places, to get more accurate and meaningful results. Also, by understanding the profiles of users and the semantic meaning of places, it is able to combine the content-based method to reduce the cold start problem for our recommendation. ACKNOWLEDGMENT This work is supported by KAKENHI ( ) Japan. Mohammad Shamsul Arefin is supported by the scholarship of MEXT Japan. REFERENCES [1] G. Adomavicius and A. Tuzhilin, Toward the next generation of recommender systems: Asurvey of the state-of-the-art and possible extensions, IEEE Transactions on Knowledge and Data Engineering, 2005, pp [2] A. Gunawardana and G. Shani, A Survey of Accuracy Evaluation Metrics of Recommendation Tasks, J. Mach. Learn. Res., 2009, pp [3] J. L. Herlocker, J. A. Konstan, L. G. Terveen, and John T. Riedl, Evaluating collaborative filtering recommender systems. ACM Trans. Inf. Syst., 22, 2004, pp [4] B. Berjani and T. Strufe, A Recommendation System for Spots in Location- Based Online Social Networks. SNS 11, ACM 2011, pp [5] C. Y. Chow, J. Bao and Mohamed F. Mokbel, Towards Location-based Social Networking Services. Proceedings of the 2nd ACM SIGSPATIAL International Workshop on Location Based Social Networks, 2010, pp [6] Y. Takeuchi and M. Sugimoto, CityVoyager: An Outdoor Recommendation System Based on User Location History. In Proc. of UIC, 2006, pp [7] V. W. Zheng, Y. Zheng, X. Xie, and Q. Yang, Collaborative Location and Activity Recommendations with GPS History Data. In Proc. of WWW, 2010, pp [8] Y. Zheng and X. Xie, Learning Travel Recommendation from User- Generated GPS Trajectories. ACM TIST, 2011, pp [9] Y. Zheng, X. Xie and W. Y. Ma, GeoLife: A collaborative social networking service among user location and trajectory. IEEE Data Engin. Bull. 33, 2010, pp [10] Y. Zheng, L. Z. Zhuang, X. Xie and W. Y. Ma, Recommending Friends and Locations Based on Individual Location History. ACM Trans. 2011, pp [11] S. Deerwester, S. Dumais, G. W. Furnas, T. K. Landauer and R. Harshman Indexing by Latent Semantic Analysis. Journal of the Society for Information Science 41, 1990, pp [12] D. W. McDonald, Recommending collaboration with social networks: a comparative evaluation. CHFCS 2003, pp [13] A. Mislove, M. Marcon, Krishna P. Gummadi, P. Druschel, and B. Bhattcharjee Measurement and analysis of online social networks. In Proceedings of the Seventh ACM Internet Measurement Conference, 2007, pp [14] M. Ester, H. P. Kriegel, J. Sander, X. W. Xu, A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. Proc. 2nd Int. Conf. on Knowledge Discovery and Data Mining, Portland, OR, AAAI Press, 1996, pp [15] M. J. Lee and C. W. Chung, A User similarity Calculation Based on the Location for Social Network Services. DASFAA 2011, pp [16] N. Li and G.L. Chen, Sharing Location in Online Social Networks. IEEE Network, 2010, pp [17] D. Ashbrook, and T. Starner, Using GPS to learn significant locations and predict movement across multiple users. UBIQ. 2003, pp [18] T. Horozov, N. Narasimhan and V. Vasudevan, Using location for personalized POI recommendations in mobile environments. In Proceedings of the International Symposium on Applications on Internet, 2006, pp [19] SINA Micro Bolg, [20] Foursquare API, 148