K-means with Effic.pptVIP

X-means: Extending K-means with Efficient Estimation of the Number of Clusters Dan Phelleg, Andrew Moore Carnegie Mellon University Published: ICML 2000 Presentation by: Payam Refaeilzadeh Problems with K-means Need to know K Searching for K is expensive Even K-means with fixed-K scales poorly Need to calculate the distance from each point to each centroid to find new cluster assignments Remedies Forward search for the appropriate value of k in a given range Recursively split each cluster and use BIC score to decide if we should keep each split Use kd-trees to accelerate individual rounds of K-means Splitting Use local BIC score to decide on keeping a split Use global BIC score to decide which K to output at the end BIC (Bayesian Information Criterion) Kd-trees Points to be clustered are put into a binary hierarchical structure Each node represents a subset of points and stores The minimal hyper-rectangle enclosing all points in the subset The vector-sum of all the points in the subset The number of points in the subset Using kd-trees For each centroid store a counter containing the vector sum of all the points belonging to it and the number of points Update the above by scanning the kd-tree only once Start with the root node and all centroids As you walk down the tree centroids start to get black-listed (when the points in that node could not possibly belong to a centroid) When only one centroid remains, the counter for that centroid can be updated using the statistics stored in the node At the end of the scan we have enough info to recalculate the centroid coordinates Results * * Adjusted Log-likelihood of the model. The likelihood that the data is “explained by” the clusters according to the spherical-Gaussian assumption of k-means * * *