Article Preview
Top1. Introduction
Location-Based Social Networks (LBSNs) like Foursquare, are fast becoming an important part of our daily lives, wherein one can connect with friends, share images and at the same time, discover new geographical areas and locations (Zheng et al., 2011) LBSNs help us to comprehend, investigate, explore, and geographically record the spots we live in. Foursquare permits us to spatially check-in to famous places of a city, updating rich databases that hold computerized engravings of our associations at the same time (Cao, Cong, & Jensen, 2010). Social-users update and share what they do, where they are and how they feel at a particular time or place. These can notify a social-user about the check-in status of their friends to nearby geographical locations with the help of the network itself or by a third-party service provider. User preferences can be analyzed to enable personalized location-based systems by analyzing the data entered. In addition, users uncover when and where they are experiencing a passionate emergency, encountering their very own paradise or damnation, having a good time or a prophetically calamitous occasion. This information, with more conventional government informational indexes, reveals the manner in which the money-related occurrences are linked with these applications, because more check-ins occur at places with higher business importance.
Figure 1. Visualization of a LBSN with common check-ins and venue-locations
A sample LBSN visualization is shown in Figure 1, wherein social-users sharing common interests stamp their check-ins at the venue-locations of their choice. This check-in information can be utilized in order to predict the location of a social-user. Apart from predicting locations of social-users, the proposed model can be very useful in the expert and intelligent systems environments, where the system needs to detect malicious users present in various domains like credit card transactions, spammers, etc. It helps commercial-intelligences, where valuable insights for the high-level decision makers can be provided from the electronic information. It also helps security agencies and government firms in tracking crime suspects.
LBSN datasets contain various types of features. In addition, the presence of redundant and irrelevant feature-values makes it hard to train the classification models to obtain higher accuracies. Thus, feature analysis in terms of their relevance in classification models is the foremost important pre-processing step in the data mining process (Yang & Olafsson, 2006). Feature selection helps to choose best features by eliminating unwanted features. Sometimes, this improves the accuracy of classifiers. This pre-processing step improves the learning speed and overall accuracy of the classification model.
In literature most of the researchers found that individual classifiers performance is moderate as compared to ensemble approaches. Hence, ensemble approaches are gaining researcher’s attention in recent studies. Bagging, Boosting, Random Sub-space, etc. (Dietterich, 2000) are commonly used ensemble frameworks. The ensemble of classifiers is expected to reduce the estimated error variance and improve the overall classification accuracy. Various machine learning techniques like Support Vector Machines, Decision Tree, Naïve Bayes algorithm, etc. are used extensively for classification purposes. But the results obtained show lack of an ideal classifier that does not produce errors. Thus, a small improvement in classification accuracy can be very important for security purposes where mistakes can cause a lot of commercial-damage to businesses and government agencies. Combining multiple classification techniques into ensembles, greatly improves the classification performance (Dietterich, 2000). For creating ensemble classifier model, a particular set of base-classifiers are trained and placed in a specific order, based on their individual performances.