Chapter 11. Cluster Analysis Advanced Methods.ppt

1 Data Mining: Concepts and Techniques (3rd ed.) Chapter 11 1 Subrata Kumer Paul Assistant Professor, Dept. of CSE, BAUET sksubrata96@gmail.com

2 Review: Basic Cluster Analysis Methods (Chap. 10)  Cluster Analysis: Basic Concepts  Group data so that object similarity is high within clusters but low across clusters  Partitioning Methods  K-means and k-medoids algorithms and their refinements  Hierarchical Methods  Agglomerative and divisive method, Birch, Cameleon  Density-Based Methods  DBScan, Optics and DenCLu  Grid-Based Methods  STING and CLIQUE (subspace clustering)  Evaluation of Clustering  Assess clustering tendency, determine # of clusters, and measure clustering quality 2

K-Means Clustering K=2 Arbitrarily partition objects into k groups Update the cluster centroids Update the cluster centroids Reassign objects Loop if needed 3 The initial data set  Partition objects into k nonempty subsets  Repeat  Compute centroid (i.e., mean point) for each partition  Assign each object to the cluster of its nearest centroid  Until no change

Hierarchical Clustering  Use distance matrix as clustering criteria. This method does not require the number of clusters k as an input, but needs a termination condition Step 0 Step 1 Step 2 Step 3 Step 4 b d c e a a b d e c d e a b c d e Step 4 Step 3 Step 2 Step 1 Step 0 agglomerative (AGNES) divisive (DIANA) 4

Distance between Clusters  Single link: smallest distance between an element in one cluster and an element in the other, i.e., dist(Ki, Kj) = min(tip, tjq)  Complete link: largest distance between an element in one cluster and an element in the other, i.e., dist(Ki, Kj) = max(tip, tjq)  Average: avg distance between an element in one cluster and an element in the other, i.e., dist(Ki, Kj) = avg(tip, tjq)  Centroid: distance between the centroids of two clusters, i.e., dist(Ki, Kj) = dist(Ci, Cj)  Medoid: distance between the medoids of two clusters, i.e., dist(Ki, Kj) = dist(Mi, Mj)  Medoid: a chosen, centrally located object in the cluster X X 5

BIRCH and the Clustering Feature (CF) Tree Structure CF1 child1 CF3 child3 CF2 child2 CF6 child6 CF1 child1 CF3 child3 CF2 child2 CF5 child5 CF1 CF2 CF6 prev next CF1 CF2 CF4 prev next B = 7 L = 6 Root Non-leaf node Leaf node Leaf node 6 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 CF = (5, (16,30),(54,190)) (3,4) (2,6) (4,5) (4,7) (3,8)

Overall Framework of CHAMELEON Construct (K-NN) Sparse Graph Partition the Graph Merge Partition Final Clusters Data Set K-NN Graph P and q are connected if q is among the top k closest neighbors of p Relative interconnectivity: connectivity of c1 and c2 over internal connectivity Relative closeness: closeness of c1 and c2 over internal closeness 7

Density-Based Clustering: DBSCAN  Two parameters:  Eps: Maximum radius of the neighbourhood  MinPts: Minimum number of points in an Eps- neighbourhood of that point  NEps(p): {q belongs to D | dist(p,q) ≤ Eps}  Directly density-reachable: A point p is directly density- reachable from a point q w.r.t. Eps, MinPts if  p belongs to NEps(q)  core point condition: |NEps (q)| ≥ MinPts MinPts = 5 Eps = 1 cm p q 8

9 Density-Based Clustering: OPTICS & Its Applications

DENCLU: Center-Defined and Arbitrary 10

STING: A Statistical Information Grid Approach  Wang, Yang and Muntz (VLDB’97)  The spatial area is divided into rectangular cells  There are several levels of cells corresponding to different levels of resolution 11

Evaluation of Clustering Quality  Assessing Clustering Tendency  Assess if non-random structure exists in the data by measuring the probability that the data is generated by a uniform data distribution  Determine the Number of Clusters  Empirical method: # of clusters ≈√n/2  Elbow method: Use the turning point in the curve of sum of within cluster variance w.r.t # of clusters  Cross validation method  Measuring Clustering Quality  Extrinsic: supervised  Compare a clustering against the ground truth using certain clustering quality measure  Intrinsic: unsupervised  Evaluate the goodness of a clustering by considering how well the clusters are separated, and how compact the clusters are 12

13 Outline of Advanced Clustering Analysis  Probability Model-Based Clustering  Each object may take a probability to belong to a cluster  Clustering High-Dimensional Data  Curse of dimensionality: Difficulty of distance measure in high-D space  Clustering Graphs and Network Data  Similarity measurement and clustering methods for graph and networks  Clustering with Constraints  Cluster analysis under different kinds of constraints, e.g., that raised from background knowledge or spatial distribution of the objects

14 Chapter 11. Cluster Analysis: Advanced Methods  Probability Model-Based Clustering  Clustering High-Dimensional Data  Clustering Graphs and Network Data  Clustering with Constraints  Summary 14

Fuzzy Set and Fuzzy Cluster  Clustering methods discussed so far  Every data object is assigned to exactly one cluster  Some applications may need for fuzzy or soft cluster assignment  Ex. An e-game could belong to both entertainment and software  Methods: fuzzy clusters and probabilistic model-based clusters  Fuzzy cluster: A fuzzy set S: FS : X → [0, 1] (value between 0 and 1)  Example: Popularity of cameras is defined as a fuzzy mapping  Then, A(0.05), B(1), C(0.86), D(0.27) 15

Fuzzy (Soft) Clustering  Example: Let cluster features be  C1 :“digital camera” and “lens”  C2: “computer“  Fuzzy clustering  k fuzzy clusters C1, …,Ck ,represented as a partition matrix M = [wij]  P1: for each object oi and cluster Cj, 0 ≤ wij ≤ 1 (fuzzy set)  P2: for each object oi, , equal participation in the clustering  P3: for each cluster Cj , ensures there is no empty cluster  Let c1, …, ck as the center of the k clusters  For an object oi, sum of the squared error (SSE), p is a parameter:  For a cluster Ci, SSE:  Measure how well a clustering fits the data: 16

Probabilistic Model-Based Clustering  Cluster analysis is to find hidden categories.  A hidden category (i.e., probabilistic cluster) is a distribution over the data space, which can be mathematically represented using a probability density function (or distribution function).  Ex. 2 categories for digital cameras sold  consumer line vs. professional line  density functions f1, f2 for C1, C2  obtained by probabilistic clustering  A mixture model assumes that a set of observed objects is a mixture of instances from multiple probabilistic clusters, and conceptually each observed object is generated independently  Out task: infer a set of k probabilistic clusters that is mostly likely to generate D using the above data generation process 17

18 Model-Based Clustering  A set C of k probabilistic clusters C1, …,Ck with probability density functions f1, …, fk, respectively, and their probabilities ω1, …, ωk.  Probability of an object o generated by cluster Cj is  Probability of o generated by the set of cluster C is  Since objects are assumed to be generated independently, for a data set D = {o1, …, on}, we have,  Task: Find a set C of k probabilistic clusters s.t. P(D|C) is maximized  However, maximizing P(D|C) is often intractable since the probability density function of a cluster can take an arbitrarily complicated form  To make it computationally feasible (as a compromise), assume the probability density functions being some parameterized distributions

19 Univariate Gaussian Mixture Model  O = {o1, …, on} (n observed objects), Θ = {θ1, …, θk} (parameters of the k distributions), and Pj(oi| θj) is the probability that oi is generated from the j-th distribution using parameter θj, we have  Univariate Gaussian mixture model  Assume the probability density function of each cluster follows a 1- d Gaussian distribution. Suppose that there are k clusters.  The probability density function of each cluster are centered at μj with standard deviation σj, θj, = (μj, σj), we have

The EM (Expectation Maximization) Algorithm  The k-means algorithm has two steps at each iteration:  Expectation Step (E-step): Given the current cluster centers, each object is assigned to the cluster whose center is closest to the object: An object is expected to belong to the closest cluster  Maximization Step (M-step): Given the cluster assignment, for each cluster, the algorithm adjusts the center so that the sum of distance from the objects assigned to this cluster and the new center is minimized  The (EM) algorithm: A framework to approach maximum likelihood or maximum a posteriori estimates of parameters in statistical models.  E-step assigns objects to clusters according to the current fuzzy clustering or parameters of probabilistic clusters  M-step finds the new clustering or parameters that maximize the sum of squared error (SSE) or the expected likelihood 20

Fuzzy Clustering Using the EM Algorithm  Initially, let c1 = a and c2 = b  1st E-step: assign o to c1,w. wt =   1st M-step: recalculate the centroids according to the partition matrix, minimizing the sum of squared error (SSE)  Iteratively calculate this until the cluster centers converge or the change is small enough

22 Univariate Gaussian Mixture Model  O = {o1, …, on} (n observed objects), Θ = {θ1, …, θk} (parameters of the k distributions), and Pj(oi| θj) is the probability that oi is generated from the j-th distribution using parameter θj, we have  Univariate Gaussian mixture model  Assume the probability density function of each cluster follows a 1- d Gaussian distribution. Suppose that there are k clusters.  The probability density function of each cluster are centered at μj with standard deviation σj, θj, = (μj, σj), we have

23 Computing Mixture Models with EM  Given n objects O = {o1, …, on}, we want to mine a set of parameters Θ = {θ1, …, θk} s.t.,P(O|Θ) is maximized, where θj = (μj, σj) are the mean and standard deviation of the j-th univariate Gaussian distribution  We initially assign random values to parameters θj, then iteratively conduct the E- and M- steps until converge or sufficiently small change  At the E-step, for each object oi, calculate the probability that oi belongs to each distribution,  At the M-step, adjust the parameters θj = (μj, σj) so that the expected likelihood P(O|Θ) is maximized

Advantages and Disadvantages of Mixture Models  Strength  Mixture models are more general than partitioning and fuzzy clustering  Clusters can be characterized by a small number of parameters  The results may satisfy the statistical assumptions of the generative models  Weakness  Converge to local optimal (overcome: run multi-times w. random initialization)  Computationally expensive if the number of distributions is large, or the data set contains very few observed data points  Need large data sets  Hard to estimate the number of clusters 24

26 Clustering High-Dimensional Data  Clustering high-dimensional data (How high is high-D in clustering?)  Many applications: text documents, DNA micro-array data  Major challenges:  Many irrelevant dimensions may mask clusters  Distance measure becomes meaningless—due to equi-distance  Clusters may exist only in some subspaces  Methods  Subspace-clustering: Search for clusters existing in subspaces of the given high dimensional data space  CLIQUE, ProClus, and bi-clustering approaches  Dimensionality reduction approaches: Construct a much lower dimensional space and search for clusters there (may construct new dimensions by combining some dimensions in the original data)  Dimensionality reduction methods and spectral clustering

Traditional Distance Measures May Not Be Effective on High-D Data  Traditional distance measure could be dominated by noises in many dimensions  Ex. Which pairs of customers are more similar?  By Euclidean distance, we get,  despite Ada and Cathy look more similar  Clustering should not only consider dimensions but also attributes (features)  Feature transformation: effective if most dimensions are relevant (PCA & SVD useful when features are highly correlated/redundant)  Feature selection: useful to find a subspace where the data have nice clusters 27

28 The Curse of Dimensionality (graphs adapted from Parsons et al. KDD Explorations 2004)  Data in only one dimension is relatively packed  Adding a dimension “stretch” the points across that dimension, making them further apart  Adding more dimensions will make the points further apart—high dimensional data is extremely sparse  Distance measure becomes meaningless—due to equi-distance

29 Why Subspace Clustering? (adapted from Parsons et al. SIGKDD Explorations 2004)  Clusters may exist only in some subspaces  Subspace-clustering: find clusters in all the subspaces

Subspace Clustering Methods  Subspace search methods: Search various subspaces to find clusters  Bottom-up approaches  Top-down approaches  Correlation-based clustering methods  E.g., PCA based approaches  Bi-clustering methods  Optimization-based methods  Enumeration methods

Subspace Clustering Method (I): Subspace Search Methods  Search various subspaces to find clusters  Bottom-up approaches  Start from low-D subspaces and search higher-D subspaces only when there may be clusters in such subspaces  Various pruning techniques to reduce the number of higher-D subspaces to be searched  Ex. CLIQUE (Agrawal et al. 1998)  Top-down approaches  Start from full space and search smaller subspaces recursively  Effective only if the locality assumption holds: restricts that the subspace of a cluster can be determined by the local neighborhood  Ex. PROCLUS (Aggarwal et al. 1999): a k-medoid-like method 31

32 Salary (10,000 ) 20 30 40 50 60 age 5 4 3 1 2 6 7 0 20 30 40 50 60 age 5 4 3 1 2 6 7 0 Vacation (week) age Vacatio n 30 50  = 3 CLIQUE: SubSpace Clustering with Aprori Pruning

Subspace Clustering Method (II): Correlation-Based Methods  Subspace search method: similarity based on distance or density  Correlation-based method: based on advanced correlation models  Ex. PCA-based approach:  Apply PCA (for Principal Component Analysis) to derive a set of new, uncorrelated dimensions,  then mine clusters in the new space or its subspaces  Other space transformations:  Hough transform  Fractal dimensions 33

Subspace Clustering Method (III): Bi-Clustering Methods  Bi-clustering: Cluster both objects and attributes simultaneously (treat objs and attrs in symmetric way)  Four requirements:  Only a small set of objects participate in a cluster  A cluster only involves a small number of attributes  An object may participate in multiple clusters, or does not participate in any cluster at all  An attribute may be involved in multiple clusters, or is not involved in any cluster at all 34  Ex 1. Gene expression or microarray data: a gene sample/condition matrix.  Each element in the matrix, a real number, records the expression level of a gene under a specific condition  Ex. 2. Clustering customers and products  Another bi-clustering problem

Types of Bi-clusters  Let A = {a1, ..., an} be a set of genes, B = {b1, …, bn} a set of conditions  A bi-cluster: A submatrix where genes and conditions follow some consistent patterns  4 types of bi-clusters (ideal cases)  Bi-clusters with constant values:  for any i in I and j in J, eij = c  Bi-clusters with constant values on rows:  eij = c + αi  Also, it can be constant values on columns  Bi-clusters with coherent values (aka. pattern-based clusters)  eij = c + αi + βj  Bi-clusters with coherent evolutions on rows  eij (ei1j1− ei1j2)(ei2j1− ei2j2) ≥ 0  i.e., only interested in the up- or down- regulated changes across genes or conditions without constraining on the exact values 35

Bi-Clustering Methods  Real-world data is noisy: Try to find approximate bi-clusters  Methods: Optimization-based methods vs. enumeration methods  Optimization-based methods  Try to find a submatrix at a time that achieves the best significance as a bi-cluster  Due to the cost in computation, greedy search is employed to find local optimal bi-clusters  Ex. δ-Cluster Algorithm (Cheng and Church, ISMB’2000)  Enumeration methods  Use a tolerance threshold to specify the degree of noise allowed in the bi-clusters to be mined  Then try to enumerate all submatrices as bi-clusters that satisfy the requirements  Ex. δ-pCluster Algorithm (H. Wang et al.’ SIGMOD’2002, MaPle: Pei et al., ICDM’2003) 36

37 Bi-Clustering for Micro-Array Data Analysis  Left figure: Micro-array “raw” data shows 3 genes and their values in a multi-D space: Difficult to find their patterns  Right two: Some subsets of dimensions form nice shift and scaling patterns  No globally defined similarity/distance measure  Clusters may not be exclusive  An object can appear in multiple clusters

Bi-Clustering (I): δ-Bi-Cluster  For a submatrix I x J, the mean of the i-th row:  The mean of the j-th column:  The mean of all elements in the submatrix is  The quality of the submatrix as a bi-cluster can be measured by the mean squared residue value  A submatrix I x J is δ-bi-cluster if H(I x J) ≤ δ where δ ≥ 0 is a threshold. When δ = 0, I x J is a perfect bi-cluster with coherent values. By setting δ > 0, a user can specify the tolerance of average noise per element against a perfect bi-cluster  residue(eij) = eij − eiJ − eIj + eIJ 38

Bi-Clustering (I): The δ-Cluster Algorithm  Maximal δ-bi-cluster is a δ-bi-cluster I x J such that there does not exist another δ-bi-cluster I′ x J′ which contains I x J  Computing is costly: Use heuristic greedy search to obtain local optimal clusters  Two phase computation: deletion phase and additional phase  Deletion phase: Start from the whole matrix, iteratively remove rows and columns while the mean squared residue of the matrix is over δ  At each iteration, for each row/column, compute the mean squared residue:  Remove the row or column of the largest mean squared residue  Addition phase:  Expand iteratively the δ-bi-cluster I x J obtained in the deletion phase as long as the δ-bi-cluster requirement is maintained  Consider all the rows/columns not involved in the current bi-cluster I x J by calculating their mean squared residues  A row/column of the smallest mean squared residue is added into the current δ-bi-cluster  It finds only one δ-bi-cluster, thus needs to run multiple times: replacing the elements in the output bi-cluster by random numbers 39

Bi-Clustering (II): δ-pCluster  Enumerating all bi-clusters (δ-pClusters) [H. Wang, et al., Clustering by pattern similarity in large data sets. SIGMOD’02]  Since a submatrix I x J is a bi-cluster with (perfect) coherent values iff ei1j1 − ei2j1 = ei1j2 − ei2j2. For any 2 x 2 submatrix of I x J, define p-score  A submatrix I x J is a δ-pCluster (pattern-based cluster) if the p-score of every 2 x 2 submatrix of I x J is at most δ, where δ ≥ 0 is a threshold specifying a user's tolerance of noise against a perfect bi-cluster  The p-score controls the noise on every element in a bi-cluster, while the mean squared residue captures the average noise  Monotonicity: If I x J is a δ-pClusters, every x x y (x,y ≥ 2) submatrix of I x J is also a δ-pClusters.  A δ-pCluster is maximal if no more row or column can be added into the cluster and retain δ-pCluster: We only need to compute all maximal δ-pClusters. 40

MaPle: Efficient Enumeration of δ- pClusters  Pei et al., MaPle: Efficient enumerating all maximal δ- pClusters. ICDM'03  Framework: Same as pattern-growth in frequent pattern mining (based on the downward closure property)  For each condition combination J, find the maximal subsets of genes I such that I x J is a δ-pClusters  If I x J is not a submatrix of another δ-pClusters  then I x J is a maximal δ-pCluster.  Algorithm is very similar to mining frequent closed itemsets  Additional advantages of δ-pClusters:  Due to averaging of δ-cluster, it may contain outliers but still within δ-threshold  Computing bi-clusters for scaling patterns, take logarithmic on will lead to the p-score form 41   yb xb ya xa d d d d / /

Dimensionality-Reduction Methods  Dimensionality reduction: In some situations, it is more effective to construct a new space instead of using some subspaces of the original data 42  Ex. To cluster the points in the right figure, any subspace of the original one, X and Y, cannot help, since all the three clusters will be projected into the overlapping areas in X and Y axes.  Construct a new dimension as the dashed one, the three clusters become apparent when the points projected into the new dimension  Dimensionality reduction methods  Feature selection and extraction: But may not focus on clustering structure finding  Spectral clustering: Combining feature extraction and clustering (i.e., use the spectrum of the similarity matrix of the data to perform dimensionality reduction for clustering in fewer dimensions)  Normalized Cuts (Shi and Malik, CVPR’97 or PAMI’2000)  The Ng-Jordan-Weiss algorithm (NIPS’01)

Spectral Clustering: The Ng-Jordan-Weiss (NJW) Algorithm  Given a set of objects o1, …, on, and the distance between each pair of objects, dist(oi, oj), find the desired number k of clusters  Calculate an affinity matrix W, where σ is a scaling parameter that controls how fast the affinity Wij decreases as dist(oi, oj) increases. In NJW, set Wij = 0  Derive a matrix A = f(W). NJW defines a matrix D to be a diagonal matrix s.t. Dii is the sum of the i-th row of W, i.e., Then, A is set to  A spectral clustering method finds the k leading eigenvectors of A  A vector v is an eigenvector of matrix A if Av = λv, where λ is the corresponding eigen-value  Using the k leading eigenvectors, project the original data into the new space defined by the k leading eigenvectors, and run a clustering algorithm, such as k-means, to find k clusters  Assign the original data points to clusters according to how the transformed points are assigned in the clusters obtained 43

Spectral Clustering: Illustration and Comments  Spectral clustering: Effective in tasks like image processing  Scalability challenge: Computing eigenvectors on a large matrix is costly  Can be combined with other clustering methods, such as bi-clustering 44

Clustering Graphs and Network Data  Applications  Bi-partite graphs, e.g., customers and products, authors and conferences  Web search engines, e.g., click through graphs and Web graphs  Social networks, friendship/coauthor graphs  Similarity measures  Geodesic distances  Distance based on random walk (SimRank)  Graph clustering methods  Minimum cuts: FastModularity (Clauset, Newman & Moore, 2004)  Density-based clustering: SCAN (Xu et al., KDD’2007) 46

Similarity Measure (I): Geodesic Distance  Geodesic distance (A, B): length (i.e., # of edges) of the shortest path between A and B (if not connected, defined as infinite)  Eccentricity of v, eccen(v): The largest geodesic distance between v and any other vertex u ∈ V − {v}.  E.g., eccen(a) = eccen(b) = 2; eccen(c) = eccen(d) = eccen(e) = 3  Radius of graph G: The minimum eccentricity of all vertices, i.e., the distance between the “most central point” and the “farthest border”  r = min v∈V eccen(v)  E.g., radius (g) = 2  Diameter of graph G: The maximum eccentricity of all vertices, i.e., the largest distance between any pair of vertices in G  d = max v∈V eccen(v)  E.g., diameter (g) = 3  A peripheral vertex is a vertex that achieves the diameter.  E.g., Vertices c, d, and e are peripheral vertices 47

SimRank: Similarity Based on Random Walk and Structural Context  SimRank: structural-context similarity, i.e., based on the similarity of its neighbors  In a directed graph G = (V,E),  individual in-neighborhood of v: I(v) = {u | (u, v) ∈ E}  individual out-neighborhood of v: O(v) = {w | (v, w) ∈ E}  Similarity in SimRank:  Initialization:  Then we can compute si+1 from si based on the definition  Similarity based on random walk: in a strongly connected component  Expected distance:  Expected meeting distance:  Expected meeting probability: 48 P[t] is the probability of the tour

Graph Clustering: Sparsest Cut  G = (V,E). The cut set of a cut is the set of edges {(u, v) ∈ E | u ∈ S, v ∈ T } and S and T are in two partitions  Size of the cut: # of edges in the cut set  Min-cut (e.g., C1) is not a good partition  A better measure: Sparsity:  A cut is sparsest if its sparsity is not greater than that of any other cut  Ex. Cut C2 = ({a, b, c, d, e, f, l}, {g, h, i, j, k}) is the sparsest cut  For k clusters, the modularity of a clustering assesses the quality of the clustering:  The modularity of a clustering of a graph is the difference between the fraction of all edges that fall into individual clusters and the fraction that would do so if the graph vertices were randomly connected  The optimal clustering of graphs maximizes the modularity li: # edges between vertices in the i-th cluster di: the sum of the degrees of the vertices in the i-th cluster 49

Graph Clustering: Challenges of Finding Good Cuts  High computational cost  Many graph cut problems are computationally expensive  The sparsest cut problem is NP-hard  Need to tradeoff between efficiency/scalability and quality  Sophisticated graphs  May involve weights and/or cycles.  High dimensionality  A graph can have many vertices. In a similarity matrix, a vertex is represented as a vector (a row in the matrix) whose dimensionality is the number of vertices in the graph  Sparsity  A large graph is often sparse, meaning each vertex on average connects to only a small number of other vertices  A similarity matrix from a large sparse graph can also be sparse 50

Two Approaches for Graph Clustering  Two approaches for clustering graph data  Use generic clustering methods for high-dimensional data  Designed specifically for clustering graphs  Using clustering methods for high-dimensional data  Extract a similarity matrix from a graph using a similarity measure  A generic clustering method can then be applied on the similarity matrix to discover clusters  Ex. Spectral clustering: approximate optimal graph cut solutions  Methods specific to graphs  Search the graph to find well-connected components as clusters  Ex. SCAN (Structural Clustering Algorithm for Networks)  X. Xu, N. Yuruk, Z. Feng, and T. A. J. Schweiger, “SCAN: A Structural Clustering Algorithm for Networks”, KDD'07 51

SCAN: Density-Based Clustering of Networks  How many clusters?  What size should they be?  What is the best partitioning?  Should some points be segregated? 52 An Example Network  Application: Given simply information of who associates with whom, could one identify clusters of individuals with common interests or special relationships (families, cliques, terrorist cells)?

A Social Network Model  Cliques, hubs and outliers  Individuals in a tight social group, or clique, know many of the same people, regardless of the size of the group  Individuals who are hubs know many people in different groups but belong to no single group. Politicians, for example bridge multiple groups  Individuals who are outliers reside at the margins of society. Hermits, for example, know few people and belong to no group  The Neighborhood of a Vertex 53 v  Define () as the immediate neighborhood of a vertex (i.e. the set of people that an individual knows )

Structure Similarity  The desired features tend to be captured by a measure we call Structural Similarity  Structural similarity is large for members of a clique and small for hubs and outliers | ) ( || ) ( | | ) ( ) ( | ) , ( w v w v w v        54 v

Structural Connectivity [1]  -Neighborhood:  Core:  Direct structure reachable:  Structure reachable: transitive closure of direct structure reachability  Structure connected: } ) , ( | ) ( { ) (        w v v w v N       | ) ( | ) ( , v N v CORE ) ( ) ( ) , ( , , v N w v CORE w v DirRECH         ) , ( ) , ( : ) , ( , , , w u RECH v u RECH V u w v CONNECT           [1] M. Ester, H. P. Kriegel, J. Sander, & X. Xu (KDD'96) “A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases 55

Structure-Connected Clusters  Structure-connected cluster C  Connectivity:  Maximality:  Hubs:  Not belong to any cluster  Bridge to many clusters  Outliers:  Not belong to any cluster  Connect to less clusters ) , ( : , , w v CONNECT C w v     C w w v REACH C v V w v       ) , ( : , ,  hub outlier 56

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 57

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 0.63 58

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 0.75 0.67 0.82 59

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 60

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 0.67 61

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 0.73 0.73 0.73 62

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 63

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 0.51 64

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 0.68 65

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 0.51 66

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 67

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 0.51 0.51 0.68 68

13 9 10 11 7 8 12 6 4 0 1 5 2 3 Algorithm  = 2  = 0.7 69

Running Time  Running time = O(|E|)  For sparse networks = O(|V|) [2] A. Clauset, M. E. J. Newman, & C. Moore, Phys. Rev. E 70, 066111 (2004). 70

Chapter 11. Cluster Analysis: Advanced Methods  Probability Model-Based Clustering  Clustering High-Dimensional Data  Clustering Graphs and Network Data  Clustering with Constraints  Summary 71

72 Why Constraint-Based Cluster Analysis?  Need user feedback: Users know their applications the best  Less parameters but more user-desired constraints, e.g., an ATM allocation problem: obstacle & desired clusters

73 Categorization of Constraints  Constraints on instances: specifies how a pair or a set of instances should be grouped in the cluster analysis  Must-link vs. cannot link constraints  must-link(x, y): x and y should be grouped into one cluster  Constraints can be defined using variables, e.g.,  cannot-link(x, y) if dist(x, y) > d  Constraints on clusters: specifies a requirement on the clusters  E.g., specify the min # of objects in a cluster, the max diameter of a cluster, the shape of a cluster (e.g., a convex), # of clusters (e.g., k)  Constraints on similarity measurements: specifies a requirement that the similarity calculation must respect  E.g., driving on roads, obstacles (e.g., rivers, lakes)  Issues: Hard vs. soft constraints; conflicting or redundant constraints

74 Constraint-Based Clustering Methods (I): Handling Hard Constraints  Handling hard constraints: Strictly respect the constraints in cluster assignments  Example: The COP-k-means algorithm  Generate super-instances for must-link constraints  Compute the transitive closure of the must-link constraints  To represent such a subset, replace all those objects in the subset by the mean.  The super-instance also carries a weight, which is the number of objects it represents  Conduct modified k-means clustering to respect cannot-link constraints  Modify the center-assignment process in k-means to a nearest feasible center assignment  An object is assigned to the nearest center so that the assignment respects all cannot-link constraints

Constraint-Based Clustering Methods (II): Handling Soft Constraints  Treated as an optimization problem: When a clustering violates a soft constraint, a penalty is imposed on the clustering  Overall objective: Optimizing the clustering quality, and minimizing the constraint violation penalty  Ex. CVQE (Constrained Vector Quantization Error) algorithm: Conduct k-means clustering while enforcing constraint violation penalties  Objective function: Sum of distance used in k-means, adjusted by the constraint violation penalties  Penalty of a must-link violation  If objects x and y must-be-linked but they are assigned to two different centers, c1 and c2, dist(c1, c2) is added to the objective function as the penalty  Penalty of a cannot-link violation  If objects x and y cannot-be-linked but they are assigned to a common center c, dist(c, c′), between c and c′ is added to the objective function as the penalty, where c′ is the closest cluster to c that can accommodate x or y 75

76 Speeding Up Constrained Clustering  It is costly to compute some constrained clustering  Ex. Clustering with obstacle objects: Tung, Hou, and Han. Spatial clustering in the presence of obstacles, ICDE'01  K-medoids is more preferable since k-means may locate the ATM center in the middle of a lake  Visibility graph and shortest path  Triangulation and micro-clustering  Two kinds of join indices (shortest-paths) worth pre-computation  VV index: indices for any pair of obstacle vertices  MV index: indices for any pair of micro- cluster and obstacle indices

77 An Example: Clustering With Obstacle Objects Taking obstacles into account Not Taking obstacles into account

78 User-Guided Clustering: A Special Kind of Constraints name office position Professor course-id name area course semester instructor office position Student name student course semester unit Register grade professor student degree Advise name Group person group Work-In area year conf Publication title title Publish author Target of clustering User hint Course Open-course  X. Yin, J. Han, P. S. Yu, “Cross-Relational Clustering with User's Guidance”, KDD'05  User usually has a goal of clustering, e.g., clustering students by research area  User specifies his clustering goal to CrossClus

79 Comparing with Classification  User-specified feature (in the form of attribute) is used as a hint, not class labels  The attribute may contain too many or too few distinct values, e.g., a user may want to cluster students into 20 clusters instead of 3  Additional features need to be included in cluster analysis All tuples for clustering User hint

80 Comparing with Semi-Supervised Clustering  Semi-supervised clustering: User provides a training set consisting of “similar” (“must-link) and “dissimilar” (“cannot link”) pairs of objects  User-guided clustering: User specifies an attribute as a hint, and more relevant features are found for clustering All tuples for clustering Semi-supervised clustering All tuples for clustering User-guided clustering x

81 Why Not Semi-Supervised Clustering?  Much information (in multiple relations) is needed to judge whether two tuples are similar  A user may not be able to provide a good training set  It is much easier for a user to specify an attribute as a hint, such as a student’s research area Tom Smith SC1211 TA Jane Chang BI205 RA Tuples to be compared User hint

82 CrossClus: An Overview  Measure similarity between features by how they group objects into clusters  Use a heuristic method to search for pertinent features  Start from user-specified feature and gradually expand search range  Use tuple ID propagation to create feature values  Features can be easily created during the expansion of search range, by propagating IDs  Explore three clustering algorithms: k-means, k-medoids, and hierarchical clustering

83 Multi-Relational Features  A multi-relational feature is defined by:  A join path, e.g., Student → Register → OpenCourse → Course  An attribute, e.g., Course.area  (For numerical feature) an aggregation operator, e.g., sum or average  Categorical feature f = [Student → Register → OpenCourse → Course, Course.area, null] Tuple Areas of courses DB AI TH t1 5 5 0 t2 0 3 7 t3 1 5 4 t4 5 0 5 t5 3 3 4 areas of courses of each student Tuple Feature f DB AI TH t1 0.5 0.5 0 t2 0 0.3 0.7 t3 0.1 0.5 0.4 t4 0.5 0 0.5 t5 0.3 0.3 0.4 Values of feature f f(t1) f(t2) f(t3) f(t4) f(t5) DB AI TH

84 Representing Features  Similarity between tuples t1 and t2 w.r.t. categorical feature f  Cosine similarity between vectors f(t1) and f(t2)  Most important information of a feature f is how f groups tuples into clusters  f is represented by similarities between every pair of tuples indicated by f  The horizontal axes are the tuple indices, and the vertical axis is the similarity  This can be considered as a vector of N x N dimensions Similarity vector Vf                    L k k L k k L k k k f p t f p t f p t f p t f t t 1 2 2 1 2 1 1 2 1 2 1 . . . . , sim

85 Similarity Between Features Feature f (course) Feature g (group) DB AI TH Info sys Cog sci Theory t1 0.5 0.5 0 1 0 0 t2 0 0.3 0.7 0 0 1 t3 0.1 0.5 0.4 0 0.5 0.5 t4 0.5 0 0.5 0.5 0 0.5 t5 0.3 0.3 0.4 0.5 0.5 0 Values of Feature f and g Similarity between two features – cosine similarity of two vectors Vf Vg   g f g f V V V V g f sim   ,

86 Computing Feature Similarity Tuples Feature f Feature g DB AI TH Info sys Cog sci Theory Similarity between feature values w.r.t. the tuples sim(fk,gq)=Σi=1 to N f(ti).pk∙g(ti).pq DB Info sys       2 1 1 1 1 , , ,           l k m q q k N i N j j i g j i f g f g f sim t t sim t t sim V V Tuple similarities, hard to compute Feature value similarities, easy to compute DB AI TH Info sys Cog sci Theory Compute similarity between each pair of feature values by one scan on data

87 Searching for Pertinent Features  Different features convey different aspects of information  Features conveying same aspect of information usually cluster tuples in more similar ways  Research group areas vs. conferences of publications  Given user specified feature  Find pertinent features by computing feature similarity Research group area Advisor Conferences of papers Research area GPA Number of papers GRE score Academic Performances Nationality Permanent address Demographic info

88 Heuristic Search for Pertinent Features Overall procedure 1. Start from the user- specified feature 2. Search in neighborhood of existing pertinent features 3. Expand search range gradually name office position Professor office position Student name student course semester unit Register grade professor student degree Advise person group Work-In name Group area year conf Publication title title Publish author Target of clustering User hint course-id name area Course course semester instructor Open-course 1 2  Tuple ID propagation is used to create multi-relational features  IDs of target tuples can be propagated along any join path, from which we can find tuples joinable with each target tuple

89 Clustering with Multi-Relational Features  Given a set of L pertinent features f1, …, fL, similarity between two tuples  Weight of a feature is determined in feature search by its similarity with other pertinent features  Clustering methods  CLARANS [Ng & Han 94], a scalable clustering algorithm for non-Euclidean space  K-means  Agglomerative hierarchical clustering         L i i f weight f t t t t i 1 2 1 2 1 . , sim , sim

90 Experiments: Compare CrossClus with  Baseline: Only use the user specified feature  PROCLUS [Aggarwal, et al. 99]: a state-of-the-art subspace clustering algorithm  Use a subset of features for each cluster  We convert relational database to a table by propositionalization  User-specified feature is forced to be used in every cluster  RDBC [Kirsten and Wrobel’00]  A representative ILP clustering algorithm  Use neighbor information of objects for clustering  User-specified feature is forced to be used

91 Measure of Clustering Accuracy  Accuracy  Measured by manually labeled data  We manually assign tuples into clusters according to their properties (e.g., professors in different research areas)  Accuracy of clustering: Percentage of pairs of tuples in the same cluster that share common label  This measure favors many small clusters  We let each approach generate the same number of clusters

92 DBLP Dataset Clustering Accurarcy - DBLP 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 C o n f W o r d C o a u t h o r C o n f + W o r d C o n f + C o a u t h o r W o r d + C o a u t h o r A l l t h r e e CrossClus K-Medoids CrossClus K-Means CrossClus Agglm Baseline PROCLUS RDBC

94 Summary  Probability Model-Based Clustering  Fuzzy clustering  Probability-model-based clustering  The EM algorithm  Clustering High-Dimensional Data  Subspace clustering: bi-clustering methods  Dimensionality reduction: Spectral clustering  Clustering Graphs and Network Data  Graph clustering: min-cut vs. sparsest cut  High-dimensional clustering methods  Graph-specific clustering methods, e.g., SCAN  Clustering with Constraints  Constraints on instance objects, e.g., Must link vs. Cannot Link  Constraint-based clustering algorithms

95 References (I)  R. Agrawal, J. Gehrke, D. Gunopulos, and P. Raghavan. Automatic subspace clustering of high dimensional data for data mining applications. SIGMOD’98  C. C. Aggarwal, C. Procopiuc, J. Wolf, P. S. Yu, and J.-S. Park. Fast algorithms for projected clustering. SIGMOD’99  S. Arora, S. Rao, and U. Vazirani. Expander flows, geometric embeddings and graph partitioning. J. ACM, 56:5:1–5:37, 2009.  J. C. Bezdek. Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press, 1981.  K. S. Beyer, J. Goldstein, R. Ramakrishnan, and U. Shaft. When is ”nearest neighbor” meaningful? ICDT’99  Y. Cheng and G. Church. Biclustering of expression data. ISMB’00  I. Davidson and S. S. Ravi. Clustering with constraints: Feasibility issues and the k-means algorithm. SDM’05  I. Davidson, K. L. Wagstaff, and S. Basu. Measuring constraint-set utility for partitional clustering algorithms. PKDD’06  C. Fraley and A. E. Raftery. Model-based clustering, discriminant analysis, and density estimation. J. American Stat. Assoc., 97:611–631, 2002.  F. H¨oppner, F. Klawonn, R. Kruse, and T. Runkler. Fuzzy Cluster Analysis: Methods for Classification, Data Analysis and Image Recognition. Wiley, 1999.  G. Jeh and J. Widom. SimRank: a measure of structural-context similarity. KDD’02  H.-P. Kriegel, P. Kroeger, and A. Zimek. Clustering high dimensional data: A survey on subspace clustering, pattern-based clustering, and correlation clustering. ACM Trans. Knowledge Discovery from Data (TKDD), 3, 2009.  U. Luxburg. A tutorial on spectral clustering. Statistics and Computing, 17:395–416, 2007

References (II)  G. J. McLachlan and K. E. Bkasford. Mixture Models: Inference and Applications to Clustering. John Wiley & Sons, 1988.  B. Mirkin. Mathematical classification and clustering. J. of Global Optimization, 12:105–108, 1998.  S. C. Madeira and A. L. Oliveira. Biclustering algorithms for biological data analysis: A survey. IEEE/ACM Trans. Comput. Biol. Bioinformatics, 1, 2004.  A. Y. Ng, M. I. Jordan, and Y. Weiss. On spectral clustering: Analysis and an algorithm. NIPS’01  J. Pei, X. Zhang, M. Cho, H. Wang, and P. S. Yu. Maple: A fast algorithm for maximal pattern-based clustering. ICDM’03  M. Radovanović, A. Nanopoulos, and M. Ivanović. Nearest neighbors in high-dimensional data: the emergence and influence of hubs. ICML’09  S. E. Schaeffer. Graph clustering. Computer Science Review, 1:27–64, 2007.  A. K. H. Tung, J. Hou, and J. Han. Spatial clustering in the presence of obstacles. ICDE’01  A. K. H. Tung, J. Han, L. V. S. Lakshmanan, and R. T. Ng. Constraint-based clustering in large databases. ICDT’01  A. Tanay, R. Sharan, and R. Shamir. Biclustering algorithms: A survey. In Handbook of Computational Molecular Biology, Chapman & Hall, 2004.  K. Wagstaff, C. Cardie, S. Rogers, and S. Schrödl. Constrained k-means clustering with background knowledge. ICML’01  H. Wang, W. Wang, J. Yang, and P. S. Yu. Clustering by pattern similarity in large data sets. SIGMOD’02  X. Xu, N. Yuruk, Z. Feng, and T. A. J. Schweiger. SCAN: A structural clustering algorithm for networks. KDD’07  X. Yin, J. Han, and P.S. Yu, “Cross-Relational Clustering with User's Guidance”, KDD'05

Slides Not to Be Used in Class 97

98 Conceptual Clustering  Conceptual clustering  A form of clustering in machine learning  Produces a classification scheme for a set of unlabeled objects  Finds characteristic description for each concept (class)  COBWEB (Fisher’87)  A popular a simple method of incremental conceptual learning  Creates a hierarchical clustering in the form of a classification tree  Each node refers to a concept and contains a probabilistic description of that concept

99 COBWEB Clustering Method A classification tree

100 More on Conceptual Clustering  Limitations of COBWEB  The assumption that the attributes are independent of each other is often too strong because correlation may exist  Not suitable for clustering large database data – skewed tree and expensive probability distributions  CLASSIT  an extension of COBWEB for incremental clustering of continuous data  suffers similar problems as COBWEB  AutoClass (Cheeseman and Stutz, 1996)  Uses Bayesian statistical analysis to estimate the number of clusters  Popular in industry

101 Neural Network Approaches  Neural network approaches  Represent each cluster as an exemplar, acting as a “prototype” of the cluster  New objects are distributed to the cluster whose exemplar is the most similar according to some distance measure  Typical methods  SOM (Soft-Organizing feature Map)  Competitive learning  Involves a hierarchical architecture of several units (neurons)  Neurons compete in a “winner-takes-all” fashion for the object currently being presented

102 Self-Organizing Feature Map (SOM)  SOMs, also called topological ordered maps, or Kohonen Self- Organizing Feature Map (KSOMs)  It maps all the points in a high-dimensional source space into a 2 to 3-d target space, s.t., the distance and proximity relationship (i.e., topology) are preserved as much as possible  Similar to k-means: cluster centers tend to lie in a low-dimensional manifold in the feature space  Clustering is performed by having several units competing for the current object  The unit whose weight vector is closest to the current object wins  The winner and its neighbors learn by having their weights adjusted  SOMs are believed to resemble processing that can occur in the brain  Useful for visualizing high-dimensional data in 2- or 3-D space

103 Web Document Clustering Using SOM  The result of SOM clustering of 12088 Web articles  The picture on the right: drilling down on the keyword “mining”  Based on websom.hut.fi Web page

Chapter 11. Cluster Analysis Advanced Methods.ppt

More Related Content

What's hot

Similar to Chapter 11. Cluster Analysis Advanced Methods.ppt

More from Subrata Kumer Paul

Recently uploaded

In this document

Chapter 11. Cluster Analysis Advanced Methods.ppt