Zeidat&Eick, MLMTA, Las Vegas K-medoid-style Clustering Algorithms for Supervised Summary Generation Nidal Zeidat & Christoph F. Eick Dept. of Computer Science University of Houston

Eick&Zeidat, MLMTA, Las Vegas Talk Outline 1. What is Supervised Clustering? 2. Representative-based Clustering Algorithms 3. Benefits of Supervised Clustering 4. Algorithms for Supervised Clustering 5. Empirical Results 6. Conclusion and Areas of Future Work

Eick&Zeidat, MLMTA, Las Vegas (Traditional) Clustering Partition a set of objects into groups of similar objects. Each group is called cluster Clustering is used to “detect classes” in data set (“unsupervised learning”) Clustering is based on a fitness function that relies on a distance measure and usually tries to minimize distance between objects within a cluster.

Eick&Zeidat, MLMTA, Las Vegas (Traditional) Clustering… (continued) A CB Attribute2 Attribute1

Eick&Zeidat, MLMTA, Las Vegas Supervised Clustering Assumes that clustering is applied to classified examples. The goal of supervised clustering is to identify class-uniform clusters that have a high probability density.  prefers clusters whose members belong to single class (low impurity) We would, also, like to keep the number of clusters low (small number of clusters).

Eick&Zeidat, MLMTA, Las Vegas Supervised Clustering … (continued) Attribute 1 Attribute 2 Traditional ClusteringSupervised Clustering

Eick&Zeidat, MLMTA, Las Vegas A Fitness Function for Supervised Clustering q(X) := Impurity(X) + β*Penalty(k) k: number of clusters used n: number of examples the dataset c: number of classes in a dataset. β: Weight for Penalty(k), 0< β ≤2.0

Eick&Zeidat, MLMTA, Las Vegas Representative-Based Supervised Clustering (RSC) Aims at finding a set of objects among all objects (called representatives) in the data set that best represent the objects in the data set. Each representative corresponds to a cluster. The remaining objects in the data set are, then, clustered around these representatives by assigning objects to the cluster of the closest representative. Remark: The popular k-medoid algorithm, also called PAM, is a representative-based clustering algorithm.

Eick&Zeidat, MLMTA, Las Vegas Representative Based Supervised Clustering … (Continued) Attribute2 Attribute1

Eick&Zeidat, MLMTA, Las Vegas Representative Based Supervised Clustering … (Continued) Attribute2 Attribute

Eick&Zeidat, MLMTA, Las Vegas Representative Based Supervised Clustering … (Continued) Attribute2 Attribute Objective of RSC: Find a subset O R of O such that the clustering X obtained by using the objects in O R as representatives minimizes q(X).

Eick&Zeidat, MLMTA, Las Vegas Why do we use Representative-Based Clustering Algorithms? Representatives themselves are useful: – can be used for summarization – can be used for dataset compression Smaller search space if compared with algorithms, such as k-means. Less sensitive to outliers Can be applied to datasets that contain nominal attributes (not feasible to compute means)

Eick&Zeidat, MLMTA, Las Vegas Applications of Supervised Clustering Enhance classification algorithms. – Use SC for Dataset Editing to enhance NN-classifiers [ICDM04] – Improve Simple Classifiers [ICDM03] Learn Sub-classes / Summary Generation Distance Function Learning Dataset Compression/Reduction For Measuring the Difficulty of a Classification Task

Eick&Zeidat, MLMTA, Las Vegas Representative Based Supervised Clustering  Dataset Editing A C E a. Dataset clustered using supervised clustering. b. Dataset edited using cluster representatives. Attribute1 D B Attribute2 F Attribute1

Eick&Zeidat, MLMTA, Las Vegas Representative Based Supervised Clustering  Enhance Simple Classifiers Attribute1 Attribute2

Eick&Zeidat, MLMTA, Las Vegas Representative Based Supervised Clustering  Learning Sub-classes Attribute2 Ford Trucks Attribute1 Ford Trucks Ford Vans GMC Trucks GMC Van :Ford :GMC

Eick&Zeidat, MLMTA, Las Vegas Clustering Algorithms Currently Investigated 1. Partitioning Around Medoids (PAM).  Traditional 2. Supervised Partitioning Around Medoids (SPAM). 3. Single Representative Insertion/Deletion Steepest Decent Hill Climbing with Randomized Restart (SRIDHCR). 4. Top Down Splitting Algorithm (TDS). 5. Supervised Clustering using Evolutionary Computing (SCEC).

Zeidat&Eick, MLMTA, Las Vegas Algorithm SRIDHCR REPEAT r TIMES curr := a randomly created set of representatives (with size between c+1 and 2*c) WHILE NOT DONE DO 1.Create new solutions S by adding a single non- representative to curr and by removing a single representative from curr. 2.Determine the element s in S for which q(s) is minimal (if there is more than one minimal element, randomly pick one). 3.IF q(s)<q(curr) THEN curr:=s ELSE IF q(s)=q(curr) AND |s|>|curr| THEN Curr:=s ELSE terminate and return curr as the solution for this run. Report the best out of the r solutions found.

Zeidat&Eick, MLMTA, Las Vegas Set of Medoids after adding one non-medoid q(X)Set of Medoids after removing a medoid q(X) (Initial solution) ……....…… …………… Trials in first part (add a non-medoid)Trials in second part (drop a medoid) RunSet of Medoids producing lowest q(X) in the runq(X)Purity (Init. Solution)

Zeidat&Eick, MLMTA, Las Vegas Algorithm SPAM Build Initial Solution curr: ( given # of clusters k ) 1.Determine the medoid of the most frequent class in the dataset. Insert that object m into curr. 2.For k-1 times, add an object v in the dataset to curr (that is not already in curr) that gives the lowest value for q(X) for curr  {v}. Improve Initial Solution curr: DO FOREVER FOR ALL representative objects r in curr DO FOR ALL non-representatives objects o in dataset DO 1.Create a new solution v by clustering the dataset around the representative set curr  {r }  {o} and insert v into S. 2.Calculate q(v) for this clustering. Determine the element s in S for which q(s) is minimal (if there is more than one minimal element, randomly pick one). IF q(s)<q(curr) THEN curr:=s ELSE TERMINATE returning curr as the final solution.

Eick&Zeidat, MLMTA, Las Vegas Differences between SPAM and SRIDHCR 1. SPAM tries to improve the current solution by replacing a representative by a non-representative, whereas SRIDHCR improves the current solution by removing a representative/by inserting a non-representative. 2. SPAM is run keeping the number of clusters k fixed, whereas SRIDHCR searches for a “good” value of k, therefore exploring a larger solution space. However, in the case of SRIDHCR which choices for k are good is somewhat restricted by the selection of the parameter . 3. SRIDHCR is run r times starting from a random initial solution, SPAM is only run once.

Eick&Zeidat, MLMTA, Las Vegas Performance Measures for the Experimental Evaluation The investigated algorithms were evaluated based on the following performance measures: Cluster Purity (Majority %). Value of the fitness function q(X). Average dissimilarity between all objects and their representatives (cluster tightness). Wall-Clock Time (WCT). Actual time, in seconds, that the algorithm took to finish the clustering task.

Zeidat&Eick, MLMTA, Las Vegas AlgorithmPurityq(X)Tightness(X). Iris-Plants data set, # clusters=3 PAM SRIDHCR SPAM Vehicle data set, # clusters =65 PAM SRIDHCR SPAM Image-Segment data set, # clusters =53 PAM SRIDHCR SPAM Pima-Indian Diabetes data set, # clusters =45 PAM SRIDHCR SPAM % 7% Table 4: Traditional vs. Supervised Clustering (β=0.1)

Zeidat&Eick, MLMTA, Las Vegas Algorithmq(X)PurityTightness (X) WCT (Sec.) IRIS-Flowers Dataset, # clusters=3 PAM SRIDHCR SPAM Vehicle Dataset, # clusters=65 PAM SRIDHCR SPAM Segmentation Dataset, # clusters=53 PAM SRIDHCR SPAM Pima-Indians-Diabetes, # clusters=45 PAM SRIDHCR SPAM Table 5: Comparative Performance of the Different Algorithms, β=0.1

AlgorithmAvg. PurityTightness(X)Avg.WCT (Sec.) IRIS-Flowers Dataset, # clusters=3 PAM SRIDHCR SPAM Vehicle Dataset, # clusters=56 PAM SRIDHCR SPAM Segmentation Dataset, # clusters=32 PAM SRIDHCR SPAM Pima-Indians-Diabetes, # clusters=2 PAM SRIDHCR SPAM Table 6: Average Comparative Performance of the Different Algorithms, β=0.4

Eick&Zeidat, MLMTA, Las Vegas Why is SRIDHCR performing so much better than SPAM? SPAM is relatively slow compared with a single run of SRIDHCR allowing for 5-30 restarts of SRIDHCR using the same resources. This enables SRIDHCR to conduct a more balanced exploration of the search space. Fitness landscape induced by q(X) contains many plateau-like structures (q(X1)=q(X2)) and many local minima and SPAM seems to get stuck more easily. The fact that SPAM uses a fixed k-value does not seem beneficiary for finding good solutions, e.g.: SRIDHCR might explore {u1,u2,u3,u4}  …  {u1,u2,u3,u4,v1,v2}  …  {u3,u4,v1,v2}, whereas SPAM might terminate with the sub- optimal solution {u1,u2,u3,u4}, if neither the replacement of u1 through v1 nor the replacement of u2 by v2 enhances q(X).

Zeidat&Eick, MLMTA, Las Vegas DatasetkβTies % Using q(X)Ties % Using Tightness(X) Iris-Plants Iris-Plants Iris-Plants Iris-Plants Vehicle Vehicle Vehicle Vehicle Segmentation Segmentation Segmentation Segmentation Diabetes Diabetes Diabetes Diabetes Table 7: Ties distribution

Zeidat&Eick, MLMTA, Las Vegas Figure 2: How Purity and k Change as β Increases

Eick&Zeidat, MLMTA, Las Vegas Conclusions 1. As expected, supervised clustering algorithms produced significantly better cluster purity than traditional clustering. Improvements range between 7% and 19% for different data sets. 2. Algorithms that too greedily explore the search space, such as SPAM, do not seem to be very suitable for supervised clustering. In general, algorithms that explore the search space more randomly seem to be more suitable for supervised clustering. 3. Supervised clustering can be used to enhance classifiers, dataset summarization, and generate better distance functions.

Eick&Zeidat, MLMTA, Las Vegas Future Work 1. Continue work on supervised clustering algorithms – Find better solutions – Faster – Explain some observations 2. Using supervised clustering for summary generation/learning subclasses 3. Using supervised clustering to find “compressed” nearest neighbor classifiers. 4. Using supervised clustering to enhance simple classifiers 5. Distance function learning

Eick&Zeidat, MLMTA, Las Vegas K-Means Algorithm Attribute2 Attribute

Eick&Zeidat, MLMTA, Las Vegas K-Means Algorithm Attribute2 Attribute