ECE471-571 Lecture 1 - Introduction

ECE471-571 Pattern Recognition Lecture 12 Unsupervised Learning (Clustering) Hairong Qi, Gonzalez Family Professor Electrical Engineering and Computer Science University of Tennessee, Knoxville http://www.eecs.utk.edu/faculty/qi Email: [email protected] Pattern Classification Statistical Approach Supervised Basic concepts: Baysian decision rule (MPP, LR, Discri.)

Non-Statistical Approach Unsupervised Basic concepts: Distance Agglomerative method Parameter estimate (ML, BL) k-means Non-Parametric learning (kNN) Winner-takes-all LDF (Perceptron)

Kohonen maps NN (BP) Mean-shift Decision-tree Syntactic approach Support Vector Machine Deep Learning (DL) Dimensionality Reduction FLD, PCA

Performance Evaluation ROC curve (TP, TN, FN, FP) cross validation Stochastic Methods local opt (GD) global opt (SA, GA) Classifier Fusion majority voting NB, BKS Review - Bayes Decision Rule P (w j | x)= Maximum Posterior

Probability Likelihood Ratio Discriminant Function p(x | w j )P (w j ) p(x) For a given x, if P(w1|x) > P(w2|x), then x belongs to class 1, otherw If , then x belongs to class 1, otherwise, 2. The classifier will assign a feature vector x to class w i if gi (x)> gj (x) Case 1: Minimum Euclidean Distance (Linear Machine), i=2I

Case 2: Minimum Mahalanobis Distance (Linear Machine), i = Case 3: Quadratic classifier , i = arbitrary Non-parametric kNN For a given x, if k1/k > k2/k, then x belongs to class 1, otherwise 2 Estimate Gaussian (Maximum Likelihood Estimation, MLE), Two-modal Gaussian Dimensionality reduction Performance evaluation ROC curve 3

Unsupervised Learning Whats unknown? In the training set, which class does each sample belong to? For the problem in general, how many classes (clusters) is appropriate? 4 Clustering Algorithm Agglomerative clustering Step1: assign each data point in the training set to a separate cluster Step2: merge the two closest clusters Step3: repeat step2 until you get the number of

clusters you want or the appropriate cluster number The result is highly dependent on the measure of cluster distance 5 Distance from a Point to a Cluster Euclidean distance deuc (x, A)= x - m A City block distance Squared Mahalanobis distance T

dmah (x, A)=(x - m A ) -A1 (x - m A ) 6 Distance between Clusters The centroid distance dmean (A, B)= m A - m B Nearest neighbor measure dmin (A, B)=min a,b deuc (a, b) for a A, b B Furthest neighbor measure dmax (A, B)=max a,b deuc (a, b) for a A, b B 7

Example dmax dmin C C B B A A

8 Minimum Spanning Tree Step1: compute all edges in the graph Step2: sort the edges by length Step3: beginning with the shortest edge, for each edge between nodes u and v, perform the following operations: Step3.1: A = find(u) (A is the cluster where u is in) Step3.2: B = find(v) (B is the cluster where v is in) Step3.3: if (A!=B) C=union(A, B), and erase sets A and B 9

Comparison of Shape of Clusters dmin tends to choose clusters which are ?? dmax tends to choose clusters which are ?? 10 Example 11 The k-means Algorithm Step1: Begin with an arbitrary assignment of samples to clusters or begin with an arbitrary set of cluster centers and assign samples to nearest clusters

Step2: Compute the sample mean of each cluster Step3: Reassign each sample to the cluster with the nearest mean Step4: If the classification of all samples has not changed, stop; else go to step 2. 12 Winner-take-all Approach Begin with an arbitrary set of cluster centers wi For each sample x, find the nearest cluster center w, which is called the winner. Modify w using wnew = wold + (x - wold) is known as a learning parameter. Typical values of this parameter are small, on the order of 0.01.

13 Winner-take-all x x-w w 14 *Kohonen Feature Maps (NN) An extension of the winner-take-all algorithm. Also called self-organizing feature maps A problem-dependent topological distance is assumed to exist between each pair of the cluster centers

When the winning cluster center is updated, so are its neighbors in the sense of this topological distance. 15 *SOM A Demo x x-w w w w w

w 16 SOM The Algorithm The winning cluster center and its neighbors are trained based on the following formula wrk+1 =wrk + (k)F(k)(x - wrk ) ware the cluster centers r gware k the coordinate of r

the cluster centers Learning Rate min kmax (as k inc, dec, (k)= max gw winneris the coordinate of more stable) max the winner g - g wr w winner F(k)=exp 2

2 2 The closer (topological closeness) the

neighbor, the more it will be affected. 17 SOM - An Example http://www.ai-junkie.com/ann/som/som1.html 18 *Mean Shift Clustering Originally proposed in 1975 by Fukunaga and Hostetler for mode detection Chengs generalization to solve clustering problems in 1995 Non-parametric no prior knowledge needed about

number of clusters Key parameter: window size Challenging issue: How to determine the right window size? Slow convergence 19 Mean Shift Clustering 1. Initialization: Choose a window/kernel of size h, e.g., a flat kernel, and apply the window on each data point, x 2. Mean calculation: Within each window centered at x, compute the mean of data, where x is the set of points enclosed within window h 3. Mean shift: Shift the window to the mean, i.e., x=m(x), where the

difference m(x)-x is referred to as the mean shift. 4. If ||m(x)-x||>, go back to step 2. 20 Original | RGB plot k=10, 24, 64 (kmeans) Init (random | uniform), k=6 Mean shift (h=30 14 clusters) Mean shift (h=20 35 clusters) 21