1. K-MEANS Michael Jones ENEE698Q Fall 2011

2. Overview  Introduction  Problem Formulation  How K-Means Works  Pros and Cons of Using K-Means  How to Improve K-Means  K-Means on a Manifold  Vector Quantization

3. Introduction  K-means was first proposed by Stuart Lloyd in 1957 as a technique for pulse-code modulation.  “Least square quantization in PCM”, Bell Telephone Laboratories Paper.  Groups data into K clusters and attempts to group data points to minimize the sum of squares distance to their central mean.  Algorithm works by iterating between two stages until the data points converge.

4. Problem Formulation  Given a data set of {x 1,…,x N } which consists of N random instances of a random D-dimensional Euclidean variable x.  Introduce a set of K prototype vectors, µ k where k=1,…,K and µ k corresponds to the mean of the k th cluster.  Goal is to find a grouping of data points and prototype vectors that minimizes the sum of squares distance of each data point.

5. Problem Formulation (cont.)  This can be formalized by introduce a indicator variable for each data point:  r nk is {0,1}, and k=1,…,K  Our objective function becomes: 

6. How K-Means works  Algorithm initializes the K prototype vectors to K distinct random data points.  Cycles between two stages until convergence is reached.  1. For each data point, determine r nk where:  2. Update µ k :

7. How K-Means works (cont)  K-Means follows the Expectation Maximization algorithm.  Stage 1 is the E step.  Stage 2 is the M step.  If K and D are fixed, the clustering can be performed in time.

8. How K-Means works (example

9. Pros and Cons of K-Means  Convergence: J may converge to a local minima and not the global minimum. May have to repeat algorithm multiple times.  Inter-Vector Relationships: Works well for Euclidian data but cannot make use of inter-vector relationships with each x.  With a large data set, the Euclidian distance calculations can be slow.  K is an input parameter. If K is inappropriately chosen it may yield poor results.

10. How to Improve K-Means  The E step can modified to have a general dissimilarity measure which leads to the K- medoids algorithm.   Can speed up K-means through various methods:  Pre-compute a tree where near by points are in the same sub tree. (Ramas. And Paliwal, 1990)  Use triangle inequality for computing distances. (Hodgson,1998).

11. Vector Quantization  Proposed by Robert M. Gray  Algorithm is nearly identical to K-Means  “Step 0. Given: A training sequence and an initial decoder.  Step 1. Encode the training sequence into a sequence of channel symbols using the given decoder minimum distortion rule. If the average distortion is small enough, quit.  Step 2. Replace the old reproduction codeword of the decoder for each channel symbol v by the centriod of all training vectors which mapped into v in Step 1. Go to Step 1.”

12. K-Means on a Manifold  K-Means can be performed on a manifold if one can compute the mean of the data.  Fletcher et al. introduced the notion of computing means on Riemannian manifolds.  Turaga et al. performed such an experiment applying K-Means on Riemannian manifolds.  Used iterative algorithm to find the sample Karcher mean  Used the dissimilarity measure:

13. Sources  Bishop C., “K-Means Clustering” in Pattern Recognition and Machine Learning, 2006, 423-430.  Fletcher, P., Lu, C., Pitzer, M., Joshi, & S., “Principal Geodesic Analysis for the Study of Nonlinear Statistics of Shape” from IEEE Transactions on Medical Imaging, VOL. 23, NO. 8, August 2004, 995-1005.  Gray, M., “Vector Quantization” in IEEE ASSP Magazine, pp. 4- 29, April 1984.  Turaga, P., Veeraraghavan, A., Srivastava, A. & Chellappa, R., “Statistical Computations on Grassman and Stiefel manifolds for Image and Video-Based Recognition” in IEEE PAMI, accepted 2010.

14. Questions? July, 2010