Location-Based Recommendation Systems
A location-based recommendation is an information filtering service, which selectively returns items (e.g., venues, travel routes, friends, or social media) to a user with the consideration of relevant spatial information (e.g., current/historical locations) and the personal preferences. The recommended results typically include k items with the highest predicated scores, which are calculated based on: (1) a recommendation technique/model (such as content-based filtering, link analysis, or collaborative filtering) and (2) the spatial relevance (like Euclidean distance or network distances).
Location-based recommendation system is one of the key social applications in urban space (Zheng et al. 2014), which emerge from two lines of research: (1) location-based services and (2) recommendation services. The traditional location-based services answer spatial queries, such as k nearest neighbor queries (i.e., kNN) and spatial range queries. However, in many cases, users do not satisfy with the results that just considering the geospatial distances and want to incorporate with more context and preferences. To this end, recommendation techniques are introduced to provide “best” results instead of the “closest.”
On the other side, the conventional recommendation services are very successful in providing the suggestions of generic items, like books, movies, and products. With the rapid development in GPS-embedded phones and wireless communications, more users are looking for recommendations related with geospatial information, e.g., suggestions for restaurants, travel routes, and activities. The user’s current location is considered as one of the most important contexts for making recommendations. Moreover, a user’s location history can also be viewed as a resource to extract a user’s preferences. Finally, the aggregated location information generated by all the users in the system can reflect as the social opinions. To this end, the conventional recommendation services also have the urge to incorporate the relevant spatial information to achieve a higher quality of recommendations.
The earliest attempts in location-based recommendations focus on providing stand-alone location suggestions (e.g., restaurants and shops) to a user. For example, Park et al. (2007) uses content-based filtering techniques to make the recommendations by matching the keywords between the venue and user’s profile. After that, e.g., Zheng et al. (2009) includes the crowd wisdom to improve the quality of the recommendation, by mining the trajectory patterns from a large number of users. Then, more personalized recommendations are provided, e.g., Ye et al. (2010).
Most recently, with the popularity of location-based social networking systems, like Foursquare and Yelp, many new types of location-aware recommendation services have been proposed to provide location-based recommendations on: (1) travel routes, (2) users, (3) activities, and (4) social media. More details can be found in a related survey (Bao et al. 2015).
There are mainly three types of methodologies used by location-based recommendations, as being based on: (1) content, (2) link analysis, or (3) collaborative filtering.
Content-based recommendations , such as Park et al. (2007), match user preferences (e.g., income, gender, and race), with features extracted from locations, such as tags, price ranges, and categories, to make recommendations. In this approach, the predication score is calculated based on the similarity measures of the matching contents, as well as the spatial distances.
The advantage of the content-based approach is that it is robust against the cold start (i.e., the users or the locations are newly added and do not have enough history information). However, such systems also face many drawbacks: (1) they do not consider the aggregated opinions from the crowd, which may result in low-quality recommendations, e.g., it may recommend a restaurant with very low review scores from the visitors, and (2) effective content-based recommendation systems require structured information from both users and locations. However, such requirement is very difficult to fulfill, as the contents (i.e., user profiles and location tags) are user-generated.
Link analysis-based recommendations, e.g., PageRank (Page et al. 1999) and Hypertext-Induced Topic Search (HITS) (Kleinberg 1999), extract high-quality nodes and nodes’ closeness by analyzing the network structure. In location-based recommendation scenario, the techniques can be extended to handle three types of networks, i.e., user-user, user-location, and location-location networks. For example, Raymond et al. (2011) extends a random walk-based link analysis algorithm to provide location recommendation. Zheng et al. (2009) extends the HITS algorithm, which is a reinforcement iterations between the user experiences and venue popularities, to recommend experienced users and interesting locations.
The advantages of link analysis-based methodologies are that (1) they take into account the user’s experiences when making recommendations and amplify ratings from experienced users and (2) they are robust against the cold start problem. However, there is one major drawback for such systems: they can only provide generic recommendations for all users, which overlooks users’ personal preferences.
Candidate Selections. The first step of CF-based recommendation systems is to select a subset of candidate nodes to reduce the computational overhead. The traditional CF-based recommendation algorithms use the most similar users (or items) as the candidates, while in location-based recommendations, user friend relations (Ye et al. 2010) and visiting histories (Levandoski et al. 2012) are used to select the candidates.
Similarity Inferences. Similarities between users (or items) are inferred from users’ ratings and location histories. The CF models can be divided into two subgroups: (1) user-based models, which use similarity measures between each pair of users, and (2) item-based models, which use similarity measures between each pair of items (media content, activities, etc.).
Many of the existing location-based recommendation systems, e.g., Ye et al. (2010), provide location recommendations based on the distribution of user’s ratings over their visited locations. Similarity inference between users (and locations) can also be done by analyzing the pattern of location co-visitation. Recently, systems have been proposed to use the number of visitations (e.g., tips and check-ins) at locations as an implicit rating of the location. Location similarity can also be captured using sequential relations or semantic similarities.
Recommendation Score Predication. Finally, CF systems predict a recommendation score for each object (i.e., locations, social media, etc.) in the candidate set. These scores are calculated from ratings given by the set of users and the similarity measures between individual users.
The advantages of the collaborative filtering models are that (1) they do not need to maintain well-structured descriptions of items (locations, activities, etc.) or users and (2) they take advantage of community opinions, which utilize the opinions from similar people and provide high-quality recommendations. However, CF-based model also suffers from several drawbacks: (1) when data is sparse, e.g., the number of user ratings is low, the user-item (location, etc.) rating matrix is very sparse and the CF model fails to make effective recommendations; (2) due to the large number of users and items in the systems, the similarity model construction process is very time consuming, presenting a scalability challenge that is exacerbated by the rapid growth and evolution of user-generated geo-tagged items; and (3) the CF model deals poorly with the cold start problem, providing recommendations for new users or new items in the system.
Stand-alone location recommendation.
Based on a user’s current and historical locations, a stand-alone location recommendation suggests individual venue for the user. The first attempt to build such stand-alone location recommendation systems takes advantage of user’s profile and venue descriptions. For example, Park et al. (2007) matches user’s profile data including age, gender, cuisine preferences, and income against the price and category of a restaurant using a Bayesian network model. In Ramaswamy et al. (2009), Ramaswamy et al., focus on enabling location recommendation on low-end devices capable only of voice and short text messages (SMS). Their approach focuses on using a user’s address and “social affinity,” social connections implied by a user’s address book, to make recommendations. The social affinity computation and spatiotemporal matching techniques in the system are continuously tuned through the user feedback.
With the popularity of location-based social networking systems, like Yelp and Foursquare, user’s historical location information (e.g., their check-ins) is used in the location recommendation systems. For example, motivated by the observation that “people who live in the same neighborhood are likely to visit the same local places”, Horozov et al. (2006) uses the historical ratings from users living close to the user’s query location, which significantly reduces the number of users in the user similarity matrix and thus reduces the computational cost of the recommendation. Ye et al. (2011) use check-ins to study the effects of the CF model, geographical distance, and social structures in making location recommendations. The authors find that geographical distance has the largest impact in their model.
More recently, location semantic information is utilized to further improve the user preference extraction and location recommendation. Shi et al. (2011) proposes a personalized location recommender system based on a category-regularized matrix, which is constructed from the user location histories. The location recommendations consider both the user’s preferences as well as a category-based location similarity. Bao et al. (2012) identifies the “new-city” recommendation issue and proposes a solution with three key components in a location recommender system, (a) the user’s current location, which constrains the location candidates; (b) the user’s location category histories, which reflect the user’s preferences; and (c) the opinions from the local experts. Yin et al. (2013) further extends the problem by proposing an LCA-LDA model, a location-content-aware probabilistic generative model to quantify both the local preference and the item content information in the recommendation process.
Travel route recommendation.
The second type of travel route recommendation uses the trajectory information. GPS trajectories contain a rich set of information, including the duration a user spent at a location and the order of location visits. Chang et al. (2011) proposes a route recommender system that takes into account a user’s own historically preferred road segments, mined from the user’s historical trajectories. The intuition for this approach is that users may feel more comfortable traveling on familiar roads. The itinerary recommender system (Yoon et al. 2012) further extends the previous works by incorporating additional constraints, such as (1) a total time constraint on the trip, e.g., a user only has 8 h for traveling; (2) a destination constraint, which indicates that the user wants to end the trip with a selected location, e.g., a user may need to return to a hotel or the airport; and (3) a constraint on specific ratio metrics, including (a) the elapsed time ratio (ETR) between the duration of the recommended trip to the total time constraint, which captures a user’s desire to utilize as much available time as possible; (b) the stay time ratio (STR) between the amount of time a user stays at location to the amount of time spent traveling between locations, which captures a user’s desire to maximize the time in the interesting locations; and (c) the interest density ratio (IDR), which is the summation of interest scores for all the locations in the trip over the maximum total interest.
Location-based social media recommendation.
Based on the user’s current location and preferences, a location-based social media recommendation system can provide suggestions for popular user-generated and geo-tagged media, such as photos, news, and messages. For example, Levandoski et al. (2012) proposes a novel location-aware recommendation framework, LARS, to exploit users’ ratings of locations using a technique that uses the distance of querying users to influences recommendations.
On the other side, some recommender systems also provide tag suggestions, when the users post photo on a location-based social networking system. For example, Silva et al. (2011) improves the quality of the image tags using a recommender system to automatically infer and suggest candidate location tags.
Location-based user recommendation.
Several studies (e.g., DeScioli et al. 2011) reveal that geographical information actually plays a vital role in determining user relationships within social networks. Location-based user recommendation provides three main functions: (1) popular user identification, (2) friend recommendation, and (3) community detection.
Popular user identification. In the context of location-based social networking/recommender system, we consider “popular users” to be the users with more knowledge about the locations. Finding experienced users is very important for the recommender systems, as these users can provide high-quality location recommendations. Zheng et al. (2009) finds that a user’s traveling experiences are regional, and a user’s experience is best determined by considering the qualities of the locations in addition to the number of locations visited. The authors propose a system to identify experienced travelers by applying a HITS inference model over a tree-based hierarchical graph of users’ historical trajectories. Similarly, Ying et al. (2011) extends the previous work and proposes four metrics that are used for analysis on EveryTrail (a website for sharing trips). They found that users who share more trajectories get more attention from other users, and users who are popular are more likely to connect to other popular users.
Friend recommendation. Recommending friends using the information extracted from users’ location history is based on the intuition that user location histories reveal preferences, and thus users with similar location histories have similar preferences and are more likely to become friends. Several publications investigate the impact of users’ geographical locations on their social relations. Yu et al. (2011) builds a pattern-based heterogeneous information network to predict connection probabilities using an unsupervised link analysis model. The connections inside the information network reflect users’ geographical histories as well as their social relationships. The connection probability and the friend recommendation score are calculated by a random walk process over the user-location network. Other works, such as Cho et al. (2011), study the relationship between user movement and friendships through an analysis of mobile phone communications and check-ins. The authors discover that users’ short-term periodical movement is irrelevant to social structure, but their long-distance movement significantly affects their social structure.
Each value in the corresponding matrix indicates the relevance measure between location, feature, and activity. The system uses a filtering approach to train the location-activity recommender system using collective matrix factorization (Singh et al. 2008). An objective function is also defined to infer the missing values. This function is iteratively minimized using gradient descent. Finally, based on the result in location-activity matrix, the top-k values are suggested as activities for the location.
There are still many open questions and challenges to be addressed in location-based recommendation systems. In particular, we summarize potential research directions as to improve the effectiveness and efficiency of recommender systems.
Effectiveness of Recommendation
Diverse Data Sources. Most location-based recommender systems currently use only one type of the data source to make recommendations. However, there are many different types data available, e.g., users’ friendships, online interactions, and user location histories. By considering more diversified data sources, more effective recommendations can be provided. For instance, the user online interactions, social structures, and location histories are all very relevant to friend recommendation. If two users have more online interactions, are close in the social structure, and have overlapped location histories, these users are likely to be compatible. A friend recommender system that can consider all these factors will make higher quality friend recommendations. In addition, other data sources, such as POIs, road networks, and traffic conditions, can also be considered in the recommendation. Fusing the knowledge from multiple heterogeneous data sources into a recommendation system is also a challenge.
Context Awareness. Current location-based recommender systems use mainly a user’s history to extract preferences. However, other important context of user’s is currently ignored. A context-aware recommender system would need to consider (1) user context, including static attributes like income, profession, and age, as well as dynamic attributes including current user location, mood, and status (e.g., at home or in meeting), and (2) environmental context, including information about the surrounding environment, e.g., the current time, weather, traffic conditions, events, etc.
Hybrid Methodologies. The recommendation methodologies used in the existing location-based recommender systems each have their own drawbacks. For example, in collaborative filtering-based recommender systems, data sparsity and cold starts are challenging problems. Link analysis-based recommender systems avoid these problems, but only provide generic recommendations that ignore users’ personal preferences. By integrating CF and link analysis-based techniques, a hybrid recommender system could overcome the weaknesses of both.
Diverse Recommendation Results. Most of the existing location-based recommendation systems focus on improving the accuracy of the recommended result, while little attention has been paid to improve the diversity of the recommended result. In the restaurant recommendation scenario, a user may not be satisfied with the results, if most of them came from the same category (e.g., fast-food). A more diverse result, which includes some other location categories like Italian restaurants and steak houses, may be preferred by the user. Moreover, in location-based recommendations, spatial diversity is also another important factor. Providing a spatially diverse recommendation results gives a user more freedom to choose her destination.
Efficiency of Recommendations
User Mobility. Users in a location-based recommendation system interact with the service using mobile devices and want up-to-date recommendations based on their current location. However, processing continuous recommendation requests as multiple individual requests is inefficient as many redundant computations are undertaken between the consecutive recommendation queries. To address this, more advanced recommendation algorithms are required that leverage prior computations to reduce the cost of continuous recommendation requests.
Frequent User Updates. Users in a location-based recommendation system can be very active, e.g., writing tips, giving ratings, and check-ins. They visit many locations over short-time spans, which adds information related to their preferences at a high rate. It is very inefficient to recompute the user preferences and user similarities every time a user undertakes a new activity. As a result, new recommendation techniques are required to efficiently address the rapid update frequency.
- Bao J, Zheng Y, Mokbel MF (2012) Location-based and preference-aware recommendation using sparse geo-social networking data. In: Proceedings of the 20th international conference on advances in geographic information systems, Redondo Beach, pp 199–208Google Scholar
- Chang K-P, Wei L-Y, Peng W-C, Yeh M-Y (2011) Discovering personalized routes from trajectories. In: Proceedings of the 3rd ACM SIGSPATIAL international workshop on location-based social networks, Chicago, pp 33–40. ACMGoogle Scholar
- Cho E, Myers SA, Leskovec J (2011) Friendship and mobility: user movement in location-based social networks. In: Proceedings of the 17th ACM SIGKDD international conference on knowledge discovery and data mining, San Diego, pp 1082–1090Google Scholar
- Horozov T, Narasimhan N, Vasudevan V (2006) Using location for personalized POI recommendations in mobile environments. In: International symposium on applications and the Internet (SAINT), Phoenix, pp 6–ppGoogle Scholar
- Huang L, Li Q, Yue Y (2010) Activity identification from GPS trajectories using spatial temporal POIs’ attractiveness. In: Proceedings of the 2nd ACM SIGSPATIAL international workshop on location based social networks, San Jose, pp 27–30Google Scholar
- Hung C-C, Chang C-W, Peng W-C (2009) Mining trajectory profiles for discovering user communities. In: Proceedings of the international workshop on location based social networks, Seattle, pp 1–8Google Scholar
- Jia-Ching Ying J, Lee W-C, Ye M, Ching-Yu Chen T, Tseng VS ( 2011) User association analysis of locales on location based social networks. In: Proceedings of the 3rd ACM SIGSPATIAL international workshop on location-based, social networks, Chicago, pp 69–76Google Scholar
- Levandoski J, Sarwat M, Eldawy A, Mokbel M (2012) LARS: a location-aware recommender system. In: 2012 IEEE 28th international conference on data engineering (ICDE), Washington, DC, pp 450–461Google Scholar
- Page L, Brin S, Motwani R, Winograd T (1999) The PageRank citation ranking: bringing order to the web. Technical reportGoogle Scholar
- Ramaswamy L, Deepak P, Polavarapu R, Gunasekera K, Garg D, Visweswariah K, Kalyanaraman S (2009) Caesar: a context-aware, social recommender system for low-end mobile devices. In: Tenth international conference on mobile data management: systems, services and Middleware (MDM), Taipei, pp 338–347Google Scholar
- Raymond R, Sugiura T, Tsubouchi K (2011) Location recommendation based on location history and spatio-temporal correlations for an on-demand bus system. In: Proceedings of the 19th ACM SIGSPATIAL international conference on advances in geographic information systems, Chicago, pp 377–380Google Scholar
- Shi Y, Serdyukov P, Hanjalic A, Larson M (2011) Personalized landmark recommendation based on geotags from photo sharing sites. In: Fourth International AAAI Conference on Web and Social Media, Menlo Park, vol 11, pp 622–625Google Scholar
- Silva A, Martins B (2011) Tag recommendation for georeferenced photos. In: Proceedings of the 3nd ACM SIGSPATIAL international workshop on location based social networks, Chicago. ACMGoogle Scholar
- Singh AP, Gordon GJ (2008) Relational learning via collective matrix factorization. In: Proceedings of the 14th ACM SIGKDD international conference on knowledge discovery and data mining, Las Vegas, pp 650–658Google Scholar
- Tai CH, Yang D-N, Lin LT, Chen MS (2008) Recommending personalized scenic itinerary with geo-tagged photos. In: IEEE international conference on multimedia and expo, Hannover, pp 1209–1212Google Scholar
- Wei L-Y, Zheng Y, Peng W-C (2012) Constructing popular routes from uncertain trajectories. In: Proceedings of the 18th ACM SIGKDD international conference on knowledge discovery and data mining, Beijing, pp 195–203Google Scholar
- Ye M, Yin P, Lee W-C (2010) Location recommendation for location-based social networks. In: Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems, San Jose, pp 458–461Google Scholar
- Ye M, Yin P, Lee W-C, Lee D-L (2011) Exploiting geographical influence for collaborative point-of-interest recommendation. In: Proceedings of the 34th international ACM SIGIR conference on research and development in information retrieval, Beijing, pp 325–334Google Scholar
- Yin Z, Cao L, Han J, Zhai C, Huang T (2011) Geographical topic discovery and comparison. In: Proceedings of the 20th international conference on World Wide Web, Hyderabad, pp 247–256Google Scholar
- Yin H, Sun Y, Cui B, Hu Z, Chen L (2013) LCARS: a location-content-aware recommender system. In: Proceedings of the 19th ACM SIGKDD international conference on knowledge discovery and data mining, Chicago, pp 221–229Google Scholar
- Yu X, Pan A, Tang L-A, Li Z, Han J (2011) Geo-friends recommendation in GPS-based cyber-physical social network. In: International conference on advances in social networks analysis and mining (ASONAM), Kaohsiung, pp 361–368Google Scholar
- Zheng Y, Zhang L, Xie X, Ma W-Y (2009) Mining interesting locations and travel sequences from GPS trajectories. In: Proceedings of the 18th international conference on World Wide Web, Madrid, pp 791–800Google Scholar
- Zheng VW, Zheng Y, Xie X, Yang Q (2010) Collaborative location and activity recommendations with GPS history data. In: Proceedings of the 19th international conference on World Wide Web, Raleigh, pp 1029–1038Google Scholar
- Zheng Y, Capra L, Wolfson O, Yang H (2014) Urban computing: concepts, methodologies, and applications. ACM Trans Intell Syst Technol (ACM TIST) 5(3):38Google Scholar