Devloping a Recommander Location Algorithm Utilizing Temporal Influence, Geographical Influence and Social Influence

Social network-based services are among the services that have been welcomed by smart phone users. Location recommendation is a popular service in social network. This service suggested the unvisited places to the users based on the is based on the users’ visiting histories and location related information such as location categories. The existing methods which utilize check-in data and category information, only consider temporal and spatial information. Since the influence of user’s social relations can play an important role in location recommendation since it can improve the algorithm performance.
In this paper, a PCLRSGT method is developed that consider temporal, spatial and social components. The spatial component models a user’s probability of checking in to a location. The spatial model obtains user’s home location by using his check-in dataset. It calculates the distance from the location to the user’s home. The spatial PDF filters out those locations that are far away from a user’s home and are not in the users’ interest. These locations should not be recommended to the user.
The temporal component employs similar users’ check-in probabilities to model a user’s probability of checking in to a location. It constructs users’ temporal curves to represent a user’s periodic check-in behavior. A User Temporal Curve U for category j is defined as a sequence of probability values. The probability value is denoted as that means probability of checking into category j in hour m (1≤m≤24).The probability sequence for user u into category is denoted as . Since the distances between users temporal curves are used to find users’ similarity, in this paper the distances are measured by curve coupling method. Temporal similarity is used to predict user’s probability of checking in to a location. The periodic behavior of a certain user is predicted by a weighted summation of the periodic behaviors of his similar users. If two users are more similar in terms of temporal similarity, they influence each other's periodic check-in behavior more. The social component models a user’s probability of checking in to a location by considering similarity between user and his friends in terms of social connection, periodic check-in behavior and check-in activities into locations. The social influence weight between two friends is concluded based on all three similarities between a user and his friends in terms of social connection, periodic check-in behavior and check-in activities into locations. Therefore, the social influence weight between two friends is calculated by combining the three above factors. The social influence weight between two friends is used to predict user’s probability of checking in to a location.
The dataset employed in this paper was collected from Gowalla. Gowalla was one of the popular location based social network launched in 2007 and closed in 2012. .The data set contains 1000 users and 15905 check-in records. A check-in record indicates a user has visited a location at a given time. It contains the user ID, location ID, and time stamp of the check-in. To evaluate the performance of the recommendation algorithm, the dataset was divided into training and testing datasets. So, one of the check-in records of each user was randomly moved to the testing dataset. The rest of the dataset formed the training dataset. As the result, the testing dataset contained 1000 check-in records, and the training dataset contained 14905 check-in records. In this paper, Precision and Recall were used to evaluate the performance of the location recommendation algorithm, which are widely accepted as the performance measurement for recommender systems.
The performance of the proposed recommendation algorithm is compared with two existing location recommendation methods, Probabilistic Category-based Location Recommendation (PCLR) and Probabilistic Category-based Location Recommendation Utilizing Temporal Influence and Geographical Influence (sPCLR). The performance of proposed algorithm is reported by recommending top-N recommendation list in the testing set (N=1, 2, 5, 10, 15 and 20). Experimental result prove that PCLRTGS performed better than all other algorithms, in terms of both precision and recall about 10 to 15 percent. This proved that using the social influence helps to improve the location recommendation. It can also be observed that the precision value decreases when the number of recommendations increases. There are two reasons for the precision decreasing, the number of correct recommendations decreases when the number of recommendation increases, or the number of correct recommendation increases with a lower rate compared to the number of recommendations. We should check the recall values to see which one is true. It also is seen that the recall values increase as the number of recommendations increases. Since the number of correct answers is a constant value, so it can concluded that the number of correct recommendations is increasing when the number of recommendations are increasing.