Starting this session, we are going to introduce grid-based clustering methods. What is grid-based clustering method? Essentially, it is you consider the whole space. Can be partitioned into multi-resolution grid structure. Then you work on the cells in this grid structure to perform multi-resolution clustering. That means we can partition the data space into a finite number of cells to form a grid structure. For example, on the plane you may be able to, to partition this plane into a 10 by 10 or 100 by 100, these kind of grid structure. Then you may find a cluster, so dense regions from the cells in the grid structure. That means you have a higher or lower resolution, or refined resolution or cross resolution in your clustering. So, a typical clustering algorithm have the following features and challenges. The first thing is very obvious. It is efficient and scalable in the sense, once you partition the number of cells, usually number of cells, even you say 10 by 10, you only have 100 cells. It is much smaller the number of data points, it could be minutes. The second one is its uniformity. It means it is uniform, because you get a ten by ten. It's a 3D uniform structure, but it, it's hard and are highly irregular data distribution. The third feature is locality. That means it's limited by the predefined cell size and borders and the predefined density threshold. The first one is a curse of dimensionality, means it's hard to cluster high dimensional data. We're going to introduce two methods in this lecture, one called STING, called a Statistic Information Grid Approach, developed by Wei Wang, Joe Yang, and Dick Muntz at UCLA, published in 1997, VLDB Conference. Another one, CLIQUE, which is both grid-based and a subspace clustering, an interesting methodology for subpace clustering developed by Rakesh Agrawal, Johannes Gehrke, Gunopulos, and Raghavan at IBM, published in 1998 SIGMOD Conference. [MUSIC]