Swap-based Clustering for Location-based Services

Clustering is an unsupervised learning method widely used in many fields, such as machine learning, pattern recognition, data mining and image analysis. The goal of this study is to investigate swap-based clustering and its application to location-based services. Swap-based clustering is a local search heuristic trying to find the optimal centroid locations by performing a sequence of centroid swaps between existing centroids and a set of candidate centroids. Firstly, the thesis presents several swap-based clustering algorithms, such as random swap, deterministic swap and hybrid swap which is a combination of random and deterministic swap. Then we propose a simple and efficient swap-based clustering algorithm, called smart swap. It performs the swap by finding the nearest pair among the centroids and sorting the clusters by their distortion values, and then it swaps one of the nearest pair centroids to any position in the cluster from the clusters list sorted by distortion value. K-means iteration is employed to repartition the dataset and fine-tune the swapped solution. Experimental results of swap-based clustering methods on both synthetic datasets and real datasets are provided and analyzed. Finally, we study location-based services and in one specific application, MOPSI project. We then apply the clustering in the MOPSI applications to reduce the clutter problem in map visualization in different scales, using a split smart swap clustering method to cluster the user locations and using a grid-based clustering with bounding box method to cluster the photo collections. Experimental results in the studied web applications show that the split smart swap method works in real-time but is slow for large dataset, and grid-based clustering method works with good clustering result and significant fast speed.