![]() | Only 14 pages are availabe for public view |
Abstract Understanding business behaviors requires acquiring huge amounts of data from diverse field studies. The additive growing use of mobile devices in social media, especially in recent years, provides large amounts of data transactions that can help in understanding business behaviors replacing the data acquired by the exhaustive field studies. Location Based Social Networks (LBSN) are considered as a solution providing such data used in urban analysis for economic reasons. Towards more insight for business behavior in this dissertation, a suggestion of global perspective exploiting data collected from LBSNs is introduced in order to predict business behavior according to the business geographical location. Moreover, business behavior prediction in LBSNs is studied in this research for big data application. Prediction of customers’ presence rates for business venues is introduced to be implemented using machine learning techniques. Machine learning techniques are investigated for both static and dynamic business predictions in LBSNs. Spatial regression models are thoroughly presented as static machine learning techniques. A comparative study is attained in this dissertation for suitability to model the relationships in LBSNs in order to be used for prediction. Geographically Weighted Regression (GWR) model proved to be the appropriate model in handling the sparse geographical distribution imposed by the LBSNs data. A proposed enhancement over the GWR model is introduced through a distributed training process that is integrated into a partitioned-GWR architecture. The proposed architecture includes a three blocks processes that are designed to deal with LBSNs data heterogeneity pursuing more enhanced predictions for business behavior. For dynamic business predictions, spatial interpolation techniques are unprecedentedly proposed to be used in LBSNs. Enhancements over the spatial interpolation techniques are proposed in this study. A Local Filtered (LF) spatial interpolation is proposed to handle the data instabilities occurrences in LBSNs, to increase business prediction accuracy. As a second enhancement, a Similarity Embedded (SE) spatial interpolation is introduced to consider the diversity of features provided by the LBSNs’ data in the prediction process. The proposed SE spatial interpolation suggested a hybrid feature similarity and distance measurements involvement in the prediction process. Moreover, a design of a Filtered Similarity Embedded (FSE) spatial interpolation is proposed. The FSE spatial interpolation pursue additive accurate prediction results through a fusion of the two previously proposed spatial interpolation enhancements; LF spatial interpolation and SE spatial interpolation. Since big data analytics are one of the most important topics of social media related research nowadays, an Iterative Nearest Neighbors First (INNF) search method is designed for timely efficient implementation of the proposed interpolation techniques over big datasets. The proposed INNF search method design the timely efficient solution using the geo indexing property provided by Not Only Structured Query Language (NOSQL) big databases. For assessing both prediction accuracy and implementation time efficiency of the previously proposed techniques, extensive experiments are implemented over data extracted from Foursquare. Data is collected and analyzed for observation about venues registered in Foursquare residing in Texas State in the United States of America. The experiments results show noticeable improvements for prediction accuracies of the spatial interpolation enhancements proposed in this study. The LF spatial interpolation increases the prediction accuracies with 61% and 54% compared to the classical K Nearest Neighbors (KNN) and Inverse Distance Weighting (IDW) spatial interpolation techniques, respectively. While the SE spatial interpolation increases the prediction accuracies with 44% and 37% compared to the KNN and IDW spatial interpolation techniques, respectively. Moreover, the FSE spatial interpolation increases the prediction accuracies with 67% and 63% compared to the KNN and IDW spatial interpolation techniques, respectively. The proposed INNF search method shows remarkable improvements in experiments. The INNF search method reduces the implementation runtime of the SE spatial interpolation in terms of milliseconds compared to the tens of seconds classical implementations in the experiments attained over the Foursquare dataset. Moreover the INNF search method shows steady runtimes for successive implementations over increased sized several synthetic datasets. |