Chapter Four Basic techniques for cluster detection Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Chapter Overview • • • • • • • The problem of cluster detection Measuring proximity between data objects The K-means cluster detection method The agglomeration cluster detection method Performance issues of the basic methods Cluster evaluation and interpretation Undertaking a clustering task in Weka Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Problem of Cluster Detection • What is cluster detection? – – – – – Cluster: a group of objects known as members The centre of a cluster is known as the centroid Members of a cluster are similar to each other Members of different clusters are different Clustering is a process of discovering clusters : centroids Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Problem of Cluster Detection • Outputs of cluster detection process – Assigned cluster tag for members of a cluster – Cluster summary: size, centroid, variations, etc. s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 s11 125 178 178 180 167 170 173 135 120 145 125 61 90 92 83 85 89 98 40 35 70 50 Cluster Tag 2 1 1 1 1 1 1 2 2 2 2 100 90 80 Body Weight SubjectID Body Height Body Weight Cluster 2: Size: 5 Centroid:(130, 51) Variation: bodyHeight = 10, bodyWeight = 14.48 70 60 Cluster 1: Size: 6 Centroid:(154, 90) Variation: bodyHeight = 5.16 bodyWeight = 5.32 50 40 30 100 110 120 130 140 150 Body Height Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning 160 170 180 190 Problem of Cluster Detection • Basic elements of a clustering solution – A sensible measure for similarity, e.g. Euclidean – An effective and efficient clustering algorithm, e.g. K-means – A goodness-of-fit function for evaluating the quality of resulting clusters, e.g. SSE ? ? ? Internal variation Inter-cluster distance Good or Bad? Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Problem of Cluster Detection • Requirements for clustering solutions – – – – – – – – – Scalability Able to deal with different types of attributes Able to discover clusters of arbitrary shapes Minimal requirements for domain knowledge to determine input parameters Able to deal with noise and outliers Insensitive to order of input data records Able to deal with high dimensionality Incorporation of user-specified constraints Interpretability and usability Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Basics – Proximity between two data objects is represented by either similarity or dissimilarity – Similarity: a numeric measure of the degree of alikeness, dissimilarity: numeric measure of the degree of difference between two objects – Similarity measure and dissimilarity measure are often convertible; normally dissimilarity is preferred – Measure of dissimilarity: • Measuring the difference between values of the corresponding attributes • Combining the measures of the differences Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Distance function – Metric properties of function d: • d(x, y) 0 and d(x, x) = 0, for all data objects x and y • d(x, y) = d(y, x), for all data objects x and y • d(x, y) d(x, z) + d(z, y), for all data objects x, y and z – Difference of values for a single attribute is directly related to the domain type of the attribute. – It is important to consider which operations are applicable. – Some measure is better than no measure at all. Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Difference between Attribute Values – Difference between nominal values • If two names are the same, the difference is 0; otherwise the maximum e.g. diff(“John”, “John”) = 0, diff(“John”, “Mary”) = • Same for difference between binary values e.g. diff(Yes, No) = – Difference between ordinal values • Different degree of proximity can be compared e.g. diff(A, B) < diff(A, D). • Converting ordinal values to consecutive integers e.g. A: 5, B: 4, C: 3, D: 2, E:1. A – B 1 and A – D 3 – Distance measure for interval and ratio attributes – Difference between values that may be unknown diff(NULL, v) = |v|, diff(NULL, NULL) = Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Distance between data objects – Ratio of mismatched features for nominal attributes Given two data objects i and j of p nominal attributes. Let m represent the number of attributes where the values of the two objects match. m d (i, j) p p e.g. Body Weight heavy heavy normal heavy low low normal low heavy low heavy Body Height short short tall tall medium tall medium short tall medium medium Blood Pressure high high normal normal normal normal high high high normal normal Blood Sugar 3 1 3 2 2 1 3 2 2 3 3 Habit smoker nonsmoker nonsmoker smoker nonsmoker nonsmoker smoker smoker nonsmoker smoker nonsmoker Class P P N N N P P P P P N d (row1, row2) 64 1 6 3 d (row1, row3) 6 1 5 6 6 Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Distance between data objects – Minkowski function for interval/ratio attributes q q q d (i, j) q (| x x | + | x x | +...+ | x x | ) i1 j1 i2 j2 ip jp Special cases: Manhattan distance (q = 1) d (i, j) | x x | + | x x | +...+ | x x | i1 j1 i2 j2 ip jp Euclidean distance (q = 2) d (i, j) (| x x |2 + | x x |2 +...+ | x x |2 ) i1 j1 i2 j2 ip jp it jt . Supremum/Chebyshev (q = ) d (i, j ) max t Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Distance between data objects – Minkowski function for interval/ratio attributes (example) customerID 101 102 103 104 105 106 107 No of Trans 30 40 35 20 50 80 10 Revenue 1000 400 300 1000 500 100 1000 Tenure(Months) 20 30 30 35 1 10 2 Manhattan Euclidean Chebyshev Tenure 40 30 d1 (cust101, cust102) | 30 40 | + | 1000 400| + | 20 30 | 620 20 10 d 2 (cust101, cust102) (30 40) + (1000 400) + (20 30) 600.16 2 d max (cust101, cust102) | 1000 400| 600 2 2 200 400 600 800 1000 Revenue Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning 10 20 30 40 50 No. of Trans Measures of Proximity • Distance between data objects – For binary attributes • Given two data objects i and j of p binary attributes, – f00 : the number of attributes where i is 0 and j is 0 – f01 : the number of attributes where i is 0 and j is 1 – f10 : the number of attributes where i is 1 and j is 0 – f11 : the number of attributes where i is 1 and j is 1 • Simple mismatch coefficient (SMC) for symmetric values: SMC (i, j ) f 01 + f10 f 00 + f 01 + f10 + f11 • Jaccard coefficient is defined for asymmetric values: JC (i, j ) f 01 + f10 f 01 + f10 + f11 Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Distance between data objects – For binary attributes (example) DocumentID Query d1 d2 d3 SMC(d1, d 2) 1 0 0 Database Programming Interface User 1 1 1 0 1 0 0 0 0 0 0 0 Usability Network Web 0 0 0 0 0 0 0 0 0 GUI 0 0 0 HTML 0 0 0 f 01 + f10 f 01 + f10 1+1 2 1+1 2 JC(d1, d 2) f 00 + f 01 + f10 + f11 7 + 1 + 1 + 1 10 f 01 + f10 + f11 1 + 1 + 1 3 SMC not that different; JC very different: two-word (out of 3) difference SMC(d1, d 3) f 01 + f10 1 1 f 00 + f 01 + f10 + f11 8 + 1 + 1 10 JC(d1, d 3) f 01 + f10 1 1 f 01 + f10 + f11 1 + 1 2 SMC very similar; JC still quite different: one word (out of 2) difference Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Similarity between data objects – Cosine similarity function • • • • Treating two data objects as vectors Similarity is measured as the angle between the two vectors Similarity is 1 when = 0, and 0 when = 90 Similarity function: i j cos(i, j ) || i || || j || i j n i k 1 k jk || i || n k 1 Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning ik 2 j i Measures of Proximity • Similarity between data objects – Cosine similarity function (illustrated) Given two data objects: x = (3, 2, 0, 5), and y = (1, 0, 0, 0) Since, x y = 3*1 + 2*0 + 0*0 + 5*0 = 3 ||x|| = sqrt(32 + 22 + 02 + 52) 6.16 ||y|| = sqrt(12 + 02 + 02 + 02) = 1 Then, the similarity between x and y: cos(x, y) = 3/(6.16 * 1) = 0.49 The dissimilarity between x and y: 1 – cos(x,y) = 0.51 Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Distance between data objects – Combining heterogeneous attributes • Based on the principle of ratio of mismatched features • For the kth attribute, compute the dissimilarity dk in [0,1] • Set the indicator variable k as follows: – k = 0, if the kth attribute is an asymmetric binary attribute and both objects have value 0 for the attribute – k = 1, otherwise • Compute the overall distance between i and j as: n d (i, j ) k dk k 1 n k k 1 Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Measures of Proximity • Distance between data objects – Attribute scaling • When: – on the same attribute when data from different data sources are merged – on different attributes when data is projected into the N-space • Normalising variables into comparable ranges: – divide each value by the mean – divide each value by the range – z-score – Attribute weighting n d (i, j ) • The weighted overall dissimilarity function: Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning w k 1 k k dk n k 1 k K-means, a Basic Clustering Method • Outline of main steps 1. Define the number of clusters (k) 2. Choose k data objects randomly to serve as the initial centroids for the k clusters 3. Assign each data object to the cluster represented by its nearest centroid 4. Find a new centroid for each cluster by calculating the mean vector of its members 5. Undo the memberships of all data objects. Go back to Step 3 and repeat the process until cluster membership no longer changes or a maximum number of iterations is reached. Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means, a Basic Clustering Method • Illustration of the method: Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means, a Basic Clustering Method • Strengths & weaknesses – Strengths • Simple and easy to implement • Quite efficient – Weaknesses • Need to specify the value of k, but we may not know what the value should be beforehand • Sensitive to the choice of initial k centroids: the result can be non-deterministic • Sensitive to noise • Applicable only when mean is meaningful to the given data set Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means, a Basic Clustering Method • Overcoming the weaknesses: – Using cluster quality to determine the value of k – Improving how the initial k centroids are chosen • Running the clustering a number of times and select the result with highest quality • Using hierarchical clustering to locate the centres • Finding centres that are farther apart – Dealing with noise • Removing outliers before clustering? • K-medoid method, using the nearest data object to the virtual centre as the centroid. – When mean cannot be defined, • K-mode method, calculating mode instead of mean for the centre of the cluster. Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means, a Basic Clustering Method Cluster errors (e.g. SSE) • Value of k and cluster quality Scree plot Number of clusters Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means, a Basic Clustering Method • Choosing initial k centroids – Running the clustering many times (only trial and error) – Using hierarchical clustering to locate the centres (why partition based?) – Finding centres that are farther apart Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means, a Basic Clustering Method • K-medoid: • Bisecting K-means Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning The Agglomeration Method • Outline of main steps 1. Take all n data objects as individual clusters and build a n x n dissimilarity matrix. The matrix stores the distance between any pair of data objects. 2. While the number of clusters > 1 do: i. Find a pair of data objects/clusters with the minimum distance ii. Merge the two data objects/clusters into a bigger cluster iii.Replace the entries in the matrix for the original clusters or objects by the cluster tag of the newly formed cluster iv.Re-calculate relevant distances and update the matrix Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning The Agglomeration Method • Illustration of the method Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning 27 The Agglomeration Method • Illustration of the method (dendrogram) # of clusters 1 2 3 4 5 6 7 8 9 10 Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning The Agglomeration Method • Agglomeration schemes – Single link: the distance between two closest points – Complete link: the distance between two farthest points – Group average: the average of all pair-wise distances – Centroids: the distance between the centroids Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning The Agglomeration Method • Strengths and weaknesses – Strengths • • • • Deterministic results Multiple possible versions of clustering No need to specify the value of a k beforehand Can create clusters of arbitrary shapes (single-link) – Weaknesses • Does not scale up for large data sets • Cannot undo membership like the K-means • Problems with agglomeration schemes (see Chapter 5) Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Cluster Evaluation & Interpretation • Cluster quality – Principle: • High-level similarity/low-level variation within a cluster • High-level dissimilarity between clusters – The measures • • • Cohesion: sum of squared errors (SSE), and sum of SSEs for all clusters (WC) Separation: sum of distances between clusters (BC) Combining the cohesion and separation, the ratio BC/WC is a good indicator of overall quality. Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Ck: cluster k rk: centroid of Ck SSE(Ck ) d ( x, r ) k xCk K W C SSE(C k) k 1 BC d (r , r ) j 1 j k K Q BC WC k 2 2 Cluster Evaluation & Interpretation • Cluster quality illustrated 100 90 Cluster c2 80 Body Weight Cluster SubjectID Body Height Body Weight Tag s1 125 61 2 s2 178 90 1 s3 178 92 1 s4 180 83 1 s5 167 85 1 s6 170 89 1 s7 173 98 1 s8 135 40 2 s9 120 35 2 s10 145 70 2 s11 125 50 2 Cluster c1 70 60 50 40 30 100 110 120 130 140 150 160 170 180 Body Height SSE(C1 ) 274.83 SSE(C2 ) 1238.8 WC 274.83+1238.8 1513.63 C1 is a better quality cluster than C2. BC 3432.3 Q 3432.3 2.268 1513.63 Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning 190 Cluster Evaluation & Interpretation • Using cluster quality for clustering – With K-means: • Add an outer loop for different values of K (from low to high) • At an iteration, conduct K-means clustering using the current K • Measure the overall cluster quality and decide whether the resulting cluster quality acceptable • If not, increase the value of K by 1 and repeat the process – With agglomeration: • • • Traverse the hierarchy level by level from the root At a level, evaluate the overall quality of clusters If the quality is acceptable, take the clusters at the level as the final result. If not, move to the next level and repeat the process. Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Cluster Evaluation & Interpretation • Cluster tendency – Cluster tendency: do clusters really exist? – Measures for tendency: • Quality measure: when BC and WC are similar, it means clusters do not exist. • Use Hopkins statistic d ( p, t ) H ( P, S ) p ,t p S d ( m, t m)+ m ,t m P P: a set of n randomly generated data points S: a sample of n data points from the data set tp: the nearest neighbour of point p in S tm: the nearest neighbour of point m in P Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning p d ( p, t p ,t p S p) Cluster Evaluation & Interpretation • Cluster interpretation – Within cluster • How values of the clustering attributes are distributed • How values of supplementary attributes are distributed – Outside cluster • Exceptions and anomalies – Between cluster • Comparative view Value distributions for the cluster Value distributions for the population Value distributions for the population Value distributions for the cluster Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means & Agglomeration in Weka • Clustering in Weka: Preprocess page Specify “No Class” Specify all attributes for clustering Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means & Agglomeration in Weka • Clustering in Weka: Cluster page 2. Set parameters 1. Choose a Clustering Solution 4. Observe results 3. Execute the chosen solution 5. Select “Visualise Cluster Assignment” Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means & Agglomeration in Weka • Clustering in Weka: SimpleKMeans Specify the distance function used Specify the value of K Specify the max. number of iterations Specify the random seed affecting the initial random selection of K centroids Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means & Agglomeration in Weka • Clustering in Weka: SimpleKMeans Save membership into a file Visualise Cluster membership Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning K-means & Agglomeration in Weka • Clustering in Weka: Agglomeration Tree-shaped Dendrogram Select Cobweb Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning Chapter Summary • A clustering solution must provide a sensible proximity function, effective algorithm and a cluster evaluation function • Proximity is normally measured by a distance function that combines measures of value differences upon attributes • The K-Means method continues to refine prototype partitions until membership changes no longer occur • The agglomeration method constructs all possible groupings of individual data objects into a hierarchy of clusters • Good clustering results mean high similarity among members of a cluster and low similarity between members of different clusters • Normal procedure of clustering in Weka is explained Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning References Read Chapter 4 of Data Mining Techniques and Applications Useful further references • Tan, P-N., Steinbach, M. and Kumar, V. (2006), Introduction to Data Mining, Addison-Wesley, Chapters 2 (section 2.4) and 8. Data Mining Techniques and Applications, 1st edition Hongbo Du ISBN 978-1-84480-891-5 © 2010 Cengage Learning