TwoStep Cluster Analysis
TwoStep Cluster Analysis
The division is accomplished on the basis of similarity of the objects across a set of specified characteristics. Outliers are a problem with this technique, often caused by too many irrelevant variables. The sample should be representative of the population, and it is desirable to have uncorrelated factors. There are three main clustering methods: hierarchical, which is a treelike process appropriate for smaller data sets; nonhierarchical, which requires specification of the number of clusters a priori, and a combination of both. There are 4 main rules for developing clusters: the clusters should be different, they should be reachable, they should be measurable, and the clusters should be profitable (big enough to matter). This is a great tool for market segmentation.
classification issues from two varying perspectives. When considering groups of objects in a multivariate data set, two situations can arise. Given a data set containing measurements on individuals, in some cases we want to see if some natural groups or classes of individuals exist, and in other cases, we want to classify the individuals according to a set of existing groups. Cluster analysis develops tools and methods concerning the former case, that is, given a data matrix containing multivariate measurements on a large number of individuals (or objects), the objective is to build some natural subgroups or clusters of individuals. This is done by grouping individuals that are \similar" according to some appropriate criterion. Once the clusters are obtained, it is generally useful to describe each group using some descriptive tool from Chapters 1, 8 or 9 to create a better understanding of the differences that exist among the formulated groups. Cluster analysis is applied in many fields such as the natural sciences, the medical sciences, economics, marketing, etc. In marketing, for instance, it is useful to build and describe the different segments of a market from a survey on potential consumers. An insurance company, on the other hand, might be interested in the distinction among classes of potential customers so that it can derive optimal prices for its services. Other examples are provided below. Discriminant analysis presented in Chapter 12 addresses the other issue of classification. It focuses on situations where the different groups are known a priori. Decision rules are provided in classifying a multivariate observation into one of the known groups
Cluster analysis is a set of tools for building groups (clusters) from multivariate data objects. The aim is to construct groups with homogeneous properties out of heterogeneous large 302 11 Cluster Analysis samples. The groups or clusters should be as homogeneous as possible and the differences among the various groups as large as possible. Cluster analysis can be divided into two fundamental steps. 1. Choice of a proximity measure: One checks each pair of observations (objects) for the similarity of their values. A similarity (proximity) measure is defined to measure the \closeness" of the objects. The \closer" they are, the more homogeneous they are. 2. Choice of group-building algorithm:
On the basis of the proximity measures the objects assigned to groups so that differences between groups become large and observations in a group become as close as possible. The starting point of a cluster analysis is a data matrix X(n _ p) with n measurements(objects) of p variables. The proximity (similarity) among objects is described by amatrix
The matrix D contains measures of similarity or dissimilarity among the n objects. If the Values dij are distances, then they measure dissimilarity. The greater the distance, the less similar are the objects. If the values dij are proximity measures, then the opposite is true, i.e., the greater the proximity value, the more similar are the objects. A distance matrix, for example, could be defined by the L2-norm: dij= kxixjk2, where xi and xj denote the rows of the data matrix X. Distance and similarity are of course dual. If dijis a distance, then d0ij = maxi;j fdijgdijis a proximity measure. The nature of the observations plays an important role in the choice of proximity measure. Nominal values (like binary variables) lead in general to proximity values, whereas metric values lead (in general) to distance matrices. We first present possibilities for D in the binary case and then consider the continuous case.
where_ and _ are weighting factors. Table 11.2 shows some similarity measures for given weighting factors. These measures provide alternative ways of weighting mis matching and positive (presence of a common character) or negative (absence of a common character) matchings. In principle, we could also consider the Euclidian distance. However, the disadvantage of this distance is that it treats the observations 0 and 1 in the same way. If xik= 1 denotes, say, knowledge of a certain language, then the contrary, xik= 0 (not knowing the language) should eventually be treated differently.
Cluster analysis is concerned with group identification. The goal of cluster analysis is to partition a set of observations into a distinct number of unknown groups or clusters in such a manner that all observations within a group are similar, while observations in different groups are not similar. If data are represented as an n p Matrix Y =yi j_ where
the goal of cluster analysis is to develop a classification scheme that will partition the rows of Y into k distinct groups (clusters). The rows of the matrix usually represent items or objects. To uncover the groupings in the data, a measure of nearness, also called a proximity measure needs to be defined. Two natural measures of nearness are the degree of distance or dissimilarity and the degree of association or similarity between groups. The choice of the proximity measure depends on the subject matter, scale of measurement (nominal, ordinal, interval, ratio), and type of variables (continuous, categorical) being analyzed. In many applications of cluster analysis, one begins with a proximity matrix rather than a data matrix. Given the proximity matrix of order (n n) say, the entries may represent dissimilarities[drs] or similarities [srs] between the r th and sth objects. Cluster analysis is a tool for classifying objects into groups and is not concerned with the geometric representation of the objects in a lowdimensional space. To explore the dimensionality of the space, one may use multidimensional scaling.
Proximity Measures Proximity measures are used to represent the nearness of two objects. If a proximity measure represents similarity, the value of the measure increases as two objects become more similar. Alternatively, if the proximity measure represents dissimilarity the value of the measure decreases in value as two objects become more alike. Letting yr and ys represent two objects in a p-variate space, an example of a dissimilarity measure is the Euclidean distance between yr and ys. As a measure of similarity, one may use the proportion of the elements in the two vectors that match. More formally, one needs to establish a set of mathematical axioms to create dissimilarity and similarity measures. Dissimilarity Measures Given two objects yr and ys in a p-dimensional space, a dissimilarity measure satisfies the following conditions.
Condition (3) implies that the measure is symmetric so that the dissimilarity measure that compares yr(object r ) with ys(object s) is the same as the comparison for object s versus object r . Condition (2) requires the measure to be zero whenever object r equals object s, Cluster Analysis To initiate a cluster analysis one constructs a proximity matrix. The proximity matrix represents the strength of the relationship between pairs of rows in Y_pnor the data matrix Ynp. Algorithms designed to perform cluster analysis are usually divided into two broad classes called hierarchical and nonhierarchical clustering methods. Generally speaking, hierarchical methods generate a sequence of cluster solutions beginning with clusters Agglomerative Hierarchical Clustering Methods Agglomerative hierarchical clustering methods use the elements of a proximity matrix to generate a tree diagram or dendogram as shown in Figure 9.3.1. To begin the process, we start with n = 5 clusters which are the branches of the tree. Combining item 1 with 2reduces the number of clusters by one, from 5 to 4. Joining items 3 and 4, results in 3clusters. Next, joining item 5 with the cluster (3, 4), results in 2 clusters. Finally, all items are combined to form a single cluster, the root of the tree Average Link Method When comparing two clusters of objects R and S, the single link and complete link methods of combining clusters depended only upon a single pair of objects within each cluster. Instead of using a minimum or maximum measure, the average link method calculates the distance between
two clusters using the average of the dissimilarities in each cluster where r R, s S, and nR and nS represent the number of objects in each cluster. Hence, the dissimilarities in Step 3 are replaced by an average of nR nS dissimilarities between all pairs of elements r R and s S. Centroid Method
In the average link method, the distance between two clusters is defined as an average of dissimilarity measures. Alternatively, suppose cluster R contains nR elements and cluster S contains nS elements. Then, the centroids for the two item clusters are
and the square of the Euclidean distance between the two clusters is d2rs= yr ys2 . For the centroid agglomerative process, one begins with any dissimilarity matrix D (in SAS, the distances are squared unless one uses the NOSQUARE option). Then, the two most similar clusters are combined using the weighted average of the two clusters. Letting T represent the new cluster, the centroid of T is
The centroid method is called the median method if an unweighted average of the centroidsis used, yt=_yr+ ys_/2. The median method is preferred when nR>>nS or nS & nR. Letting the dissimilarity matrix D =_d2r_where d2rs= _yr ys_2, suppose the element sr R and s S are combined into a cluster T where Then to calculate the square of the Euclidean distance between cluster T and the centroid yu of a third cluster U, the following formula may be used This is a special case of a general algorithm for updating proximity measures for the single link, complete link, average link, centroid and median methods developed by Williams and Lance (1977). Wards (Incremental Sum of Squares) Method Given n objects with p variables, the sum of squares within clusters where each object forms its own group is zero. For all objects in a single group, the sum of squares within clusters the sum of squares error is equal to the total sum of squares
Thus, the sum of squares within clusters is between zero and SSE. Wards method for forming clusters joins objects based upon minimizing the minimal increment in the within or error sum of squares. At each step of the process, n(n1)/2 pairs of clusters are formed and the two objects that increase the sum of squares for error least are joined. The processis continued until all objects are joined. The dendogram is constructed based upon the minimum increase in the sum of squares for error. To see how the process works, let
for clusters R and S. Combining cluster R and S to form cluster T , the error sum of squares for cluster T is whereyt=_nRyr+ nSys_/ (nR+ nS). Then, the incremental increase in joining R and S to form cluster T is SSEt (SSEr+ SSEs ). Or, letting SSEt be the total sum of squares and SSEr+ SSEs the within cluster sum of squares, the incremental increase in the errorsum of squares is no more than the between cluster sum of squares. The incremental between cluster sum of squares
(IBCSS) is For clusters with one object, (9.3.8) becomes d2rs/2. Hence, starting with a dissimilarity matrix D =_d2rs_where d2rs is the square of the Euclidean distances (the default in SAS for Wards method) between objects r and s, the two most similar objects are combined and the new incremental sum of squares proximity matrix has elements prs= d2rs/2. Combining objects r and s to form a new cluster with mean yt using (9.3.5), the incremental increase in the error sum of squares may be calculated using the formula developed by Williams and Lance (1977) as Cluster analysis is used to categorize objects such as patients, voters, products, institutions, countries, and cities among others into homogeneous groups based upon a vector of p variables. The clustering of objects into homogeneous groups depends on the scale of the variables, the algorithm used for clustering, and the criterion used to estimate the number of clusters. In general, variables with very large variances relative to others and outliers have an adverse effect on cluster analysis methods. They tend to dominate the proximity measure. Cluster analysis is an exploratory data analysis methodology. It tries to discover how objects may or may not be combined. The analysis depends on the amount of random noise in the data, the existence of outliers in the data, the variables selected for the analysis, the proximity measure used, the spatial properties of the data, and the clustering method employed.
Handling of categorical and continuous variables. By assuming variables to be independent, a joint multinomial-normal distribution can be placed on categorical and continuous variables. Automatic selection of number of clusters. By comparing the values of a model-choice criterion across different clustering solutions, the procedure can automatically determine the optimal number of clusters. Scalability. By constructing a cluster features (CF) tree that summarizes the records, the TwoStep algorithm allows you to analyze large data files. Show me Example. Retail and consumer product companies regularly apply clustering techniques to data that describe their customers' buying habits, gender, age, income level, etc. These companies tailor their marketing and product development strategies to each consumer group to increase sales and build brand loyalty. Distance Measure. This selection determines how the similarity between two clusters is computed. Log-likelihood. The likelihood measure places a probability distribution on the variables. Continuous variables are assumed to be normally distributed, while categorical variables are assumed to be multinomial. All variables are assumed to be independent. Euclidean. The Euclidean measure is the "straight line" distance between two clusters. It can be used only when all of the variables are continuous. Number of Clusters. This selection allows you to specify how the number of clusters is to be determined. Determine automatically. The procedure will automatically determine the "best" number of clusters, using the criterion specified in the Clustering Criterion group. Optionally, enter a positive integer specifying the maximum number of clusters that the procedure should consider. Specify fixed. Allows you to fix the number of clusters in the solution. Enter a positive integer. Count of Continuous Variables. This group provides a summary of the continuous variable standardization specifications made in the Options dialog box. See the topic TwoStep Cluster Analysis Options for more information. Clustering Criterion. This selection determines how the automatic clustering algorithm determines the number of clusters. Either the Bayesian Information Criterion (BIC) or the Akaike Information Criterion (AIC) can be specified. Data. This procedure works with both continuous and categorical variables. Cases represent objects to be clustered, and the variables represent attributes upon which the clustering is based.
Case Order. Note that the cluster features tree and the final solution may depend on the order of cases. To minimize order effects, randomly order the cases. You may want to obtain several different solutions with cases sorted in different random orders to verify the stability of a given solution. In situations where this is difficult due to extremely large file sizes, multiple runs with a sample of cases sorted in different random orders might be substituted. Assumptions. The likelihood distance measure assumes that variables in the cluster model are independent. Further, each continuous variable is assumed to have a normal (Gaussian) distribution, and each categorical variable is assumed to have a multinomial distribution. Empirical internal testing indicates that the procedure is fairly robust to violations of both the assumption of independence and the distributional assumptions, but you should try to be aware of how well these assumptions are met. Use the Bivariate Correlations procedure to test the independence of two continuous variables. Use the Crosstabs procedure to test the independence of two categorical variables. Use the Means procedure to test the independence between a continuous variable and categorical variable. Use the Explore procedure to test the normality of a continuous variable. Use the ChiSquare Test procedure to test whether a categorical variable has a specified multinomial distribution.
Outlier Treatment. This group allows you to treat outliers specially during clustering if the cluster features (CF) tree fills. The CF tree is full if it cannot accept any more cases in a leaf node and no leaf node can be split. If you select noise handling and the CF tree fills, it will be regrown after placing cases in sparse leaves into a "noise" leaf. A leaf is considered sparse if it contains fewer than the specified percentage of cases of the maximum leaf size. After the tree is regrown, the outliers will be placed in the CF tree if possible. If not, the outliers are discarded. If you do not select noise handling and the CF tree fills, it will be regrown using a larger distance change threshold. After final clustering, values that cannot be assigned to a cluster are labeled outliers. The outlier cluster is given an identification number of 1 and is not included in the count of the number of clusters. Memory Allocation. This group allows you to specify the maximum amount of memory in megabytes (MB) that the cluster algorithm should use. If the procedure exceeds this maximum, it will use the disk to store information that will not fit in memory. Specify a number greater than or equal to 4. Consult your system administrator for the largest value that you can specify on your system. The algorithm may fail to find the correct or desired number of clusters if this value is too low. Variable standardization. The clustering algorithm works with standardized continuous variables. Any continuous variables that are not standardized should be left as variables in the To be Standardized list. To save some time and computational effort, you can select any continuous variables that you have already standardized as variables in the Assumed Standardized list.
Advanced Options
CF Tree Tuning Criteria. The following clustering algorithm settings apply specifically to the cluster features (CF) tree and should be changed with care: Initial Distance Change Threshold. This is the initial threshold used to grow the CF tree. If inserting a given case into a leaf of the CF tree would yield tightness less than the threshold, the leaf is not split. If the tightness exceeds the threshold, the leaf is split. Maximum Branches (per leaf node). The maximum number of child nodes that a leaf node can have. Maximum Tree Depth. The maximum number of levels that the CF tree can have. Maximum Number of Nodes Possible. This indicates the maximum number of CF tree nodes that could potentially be generated by the procedure, based on the function (bd+1 1) / (b 1), where b is the maximum branches and d is the maximum tree depth. Be aware that an overly
large CF tree can be a drain on system resources and can adversely affect the performance of the procedure. At a minimum, each node requires 16 bytes. Cluster Model Update. This group allows you to import and update a cluster model generated in a prior analysis. The input file contains the CF tree in XML format. The model will then be updated with the data in the active file. You must select the variable names in the main dialog box in the same order in which they were specified in the prior analysis. The XML file remains unaltered, unless you specifically write the new model information to the same filename. See the topic TwoStep Cluster Analysis Output for more information. If a cluster model update is specified, the options pertaining to generation of the CF tree that were specified for the original model are used. More specifically, the distance measure, noise handling, memory allocation, or CF tree tuning criteria settings for the saved model are used, and any settings for these options in the dialog boxes are ignored. Note: When performing a cluster model update, the procedure assumes that none of the selected cases in the active dataset were used to create the original cluster model. The procedure also assumes that the cases used in the model update come from the same population as the cases used to create the original model; that is, the means and variances of continuous variables and levels of categorical variables are assumed to be the same across both sets of cases. If your "new" and "old" sets of cases come from heterogeneous populations, you should run the TwoStep Cluster Analysis procedure on the combined sets of cases for the best results.
Export final model. The final cluster model is exported to the specified file in XML (PMML) format. You can use this model file to apply the model information to other data files for scoring purposes. See the topic Scoring Wizard for more information. Export CF tree. This option allows you to save the current state of the cluster tree and update it later using newer data. See TwoStep Cluster Analysis Options for more information on reading this file.
The Model Summary view shows a snapshot, or summary, of the cluster model, including a Silhouette measure of cluster cohesion and separation that is shaded to indicate poor, fair, or good results. This snapshot enables you to quickly check if the quality is poor, in which case you may decide to return to the modeling node to amend the cluster model settings to produce a better result. The results of poor, fair, and good are based on the work of Kaufman and Rousseeuw (1990) regarding interpretation of cluster structures. In the Model Summary view, a good result equates to data that reflects Kaufman and Rousseeuw's rating as either reasonable or strong evidence of cluster structure, fair reflects their rating of weak evidence, and poor reflects their rating of no significant evidence. The silhouette measure averages, over all records, (BA) / max(A,B), where A is the record's distance to its cluster center and B is the record's distance to the nearest cluster center that it doesn't belong to. A silhouette coefficient of 1 would mean that all cases are located directly on their cluster centers. A value of 1 would mean all cases are located on the cluster centers of some other cluster. A value of 0 means, on average, cases are equidistant between their own cluster center and the nearest other cluster. The summary includes a table that contains the following information: Algorithm. The clustering algorithm used, for example, "TwoStep". Input Features. The number of fields, also known as inputs or predictors. Clusters. The number of clusters in the solution The Clusters view contains a cluster-by-features grid that includes cluster names, sizes, and profiles for each cluster. The columns in the grid contain the following information: Cluster. The cluster numbers created by the algorithm. Label. Any labels applied to each cluster (this is blank by default). Double-click in the cell to enter a label that describes the cluster contents; for example, "Luxury car buyers". Description. Any description of the cluster contents (this is blank by default). Double-click in the cell to enter a description of the cluster; for example, "55+ years of age, professionals, earning over $100,000". Size. The size of each cluster as a percentage of the overall cluster sample. Each size cell within the grid displays a vertical bar that shows the size percentage within the cluster, a size percentage in numeric format, and the cluster case counts.
Features. The individual inputs or predictors, sorted by overall importance by default. If any columns have equal sizes they are shown in ascending sort order of the cluster numbers. Overall feature importance is indicated by the color of the cell background shading; the most important feature is darkest; the least important feature is unshaded. A guide above the table indicates the importance attached to each feature cell color. When you hover your mouse over a cell, the full name/label of the feature and the importance value for the cell is displayed. Further information may be displayed, depending on the view and feature type. In the Cluster Centers view, this includes the cell statistic and the cell value; for example: Mean: 4.32. For categorical features the cell shows the name of the most frequent (modal) category and its percentage. Within the Clusters view, you can select various ways to display the cluster information: Transpose clusters and features. See the topic Transpose Clusters and Features for more information. Sort features. See the topic Sort Features for more information. Sort clusters. See the topic Sort Clusters for more information. Select cell contents. See the topic Cell Contents for more information.
Sort Features
The Sort Features By buttons enable you to select how feature cells are displayed: Overall Importance. This is the default sort order. Features are sorted in descending order of overall importance, and sort order is the same across clusters. If any features have tied importance values, the tied features are listed in ascending sort order of the feature names. Within-Cluster Importance. Features are sorted with respect to their importance for each cluster. If any features have tied importance values, the tied features are listed in ascending sort order of the feature names. When this option is chosen the sort order usually varies across clusters.
Name. Features are sorted by name in alphabetical order. Data order. Features are sorted by their order in the dataset.
Sort Clusters
By default clusters are sorted in descending order of size. The Sort Clusters By buttons enable you to sort them by name in alphabetical order, or, if you have created unique labels, in alphanumeric label order instead. Features that have the same label are sorted by cluster name. If clusters are sorted by label and you edit the label of a cluster, the sort order is automatically updated.
Cell Contents
The Cells buttons enable you to change the display of the cell contents for features and evaluation fields. Cluster Centers. By default, cells display feature names/labels and the central tendency for each cluster/feature combination. The mean is shown for continuous fields and the mode (most frequently occurring category) with category percentage for categorical fields. Absolute Distributions. Shows feature names/labels and absolute distributions of the features within each cluster. For categorical features, the display shows bar charts overlaid with categories ordered in ascending order of the data values. For continuous features, the display shows a smooth density plot which use the same endpoints and intervals for each cluster. The solid red colored display shows the cluster distribution, whilst the paler display represents the overall data. Relative Distributions. Shows feature names/labels and relative distributions in the cells. In general the displays are similar to those shown for absolute distributions, except that relative distributions are displayed instead. The solid red colored display shows the cluster distribution, while the paler display represents the overall data.
Basic View. Where there are a lot of clusters, it can be difficult to see all the detail without scrolling. To reduce the amount of scrolling, select this view to change the display to a more compact version of the table.
The Cluster Sizes view shows a pie chart that contains each cluster. The percentage size of each cluster is shown on each slice; hover the mouse over each slice to display the count in that slice.
Below the chart, a table lists the following size information: The size of the smallest cluster (both a count and percentage of the whole). The size of the largest cluster (both a count and percentage of the whole). The ratio of size of the largest cluster to the smallest cluster.
Select a field or cluster to view more details. Compare clusters to select items of interest. Alter the display. Transpose axes.
Filtering Records
f you want to know more about the cases in a particular cluster or group of clusters, you can select a subset of records for further analysis based on the selected clusters. Select the clusters in the Cluster view of the Cluster Viewer. To select multiple clusters, use Ctrl-click. From the menus choose: Generate > Filter Records...
Enter a filter variable name. Records from the selected clusters will receive a value of 1 for this field. All other records will receive a value of 0 and will be excluded from subsequent analyses until you change the filter status. Click OK.
Binary. Available alternatives are Euclidean distance, squared Euclidean distance, size difference, pattern difference, variance, dispersion, shape, simple matching, phi 4-point correlation, lambda, Anderberg's D, dice, Hamann, Jaccard, Kulczynski 1, Kulczynski 2, Lance and Williams, Ochiai, Rogers and Tanimoto, Russel and Rao, Sokal and Sneath 1, Sokal and Sneath 2, Sokal and Sneath 3, Sokal and Sneath 4, Sokal and Sneath 5, Yule's Y, and Yule's Q. Transform Values. Allows you to standardize data values for either cases or values before computing proximities (not available for binary data). Available standardization methods are z scores, range 1 to 1, range 0 to 1, maximum magnitude of 1, mean of 1, and standard deviation of 1. Transform Measures. Allows you to transform the values generated by the distance measure. They are applied after the distance measure has been computed. Available alternatives are absolute values, change sign, and rescale to 01 range. This feature requires the Statistics Base option. From the menus choose: Analyze > Classify > Hierarchical Cluster... In the Hierarchical Cluster Analysis dialog box, click Method.
Minkowski. The pth root of the sum of the absolute differences to the pth power between the values for the items. Customized. The rth root of the sum of the absolute differences to the pth power between the values for the items.
Simple matching. This is the ratio of matches to the total number of values. Equal weight is given to matches and nonmatches. Phi 4-point correlation. This index is a binary analog of the Pearson correlation coefficient. It has a range of 1 to 1. Lambda. This index is Goodman and Kruskal's lambda. Corresponds to the proportional reduction of error (PRE) using one item to predict the other (predicting in both directions). Values range from 0 to 1. Anderberg's D. Similar to lambda, this index corresponds to the actual reduction of error using one item to predict the other (predicting in both directions). Values range from 0 to 1. Dice. This is an index in which joint absences are excluded from consideration, and matches are weighted double. Also known as the Czekanowski or Sorensen measure. Hamann. This index is the number of matches minus the number of nonmatches, divided by the total number of items. It ranges from 1 to 1. Jaccard. This is an index in which joint absences are excluded from consideration. Equal weight is given to matches and nonmatches. Also known as the similarity ratio. Kulczynski 1. This is the ratio of joint presences to all nonmatches. This index has a lower bound of 0 and is unbounded above. It is theoretically undefined when there are no nonmatches; however, the software assigns an arbitrary value of 9999.999 when the value is undefined or is greater than this value. Kulczynski 2. This index is based on the conditional probability that the characteristic is present in one item, given that it is present in the other. The separate values for each item acting as a predictor of the other are averaged to compute this value. Lance and Williams. Computed from a fourfold table as (b+c)/(2a+b+c), where a represents the cell corresponding to cases present on both items, and b and c represent the diagonal cells corresponding to cases present on one item but absent on the other. This measure has a range of 0 to 1. (Also known as the Bray-Curtis nonmetric coefficient.) Ochiai. This index is the binary form of the cosine similarity measure. It has a range of 0 to 1. Rogers and Tanimoto. This is an index in which double weight is given to nonmatches. Russel and Rao. This is a binary version of the inner (dot) product. Equal weight is given to matches and nonmatches. This is the default for binary similarity data. Sokal and Sneath 1. This is an index in which double weight is given to matches.
Sokal and Sneath 2. This is an index in which double weight is given to nonmatches, and joint absences are excluded from consideration. Sokal and Sneath 3. This is the ratio of matches to nonmatches. This index has a lower bound of 0 and is unbounded above. It is theoretically undefined when there are no nonmatches; however, the software assigns an arbitrary value of 9999.999 when the value is undefined or is greater than this value. Sokal and Sneath 4. This index is based on the conditional probability that the characteristic in one item matches the value in the other. The separate values for each item acting as a predictor of the other are averaged to compute this value. Sokal and Sneath 5. This index is the squared geometric mean of conditional probabilities of positive and negative matches. It is independent of item coding. It has a range of 0 to 1. Yule's Y. This index is a function of the cross-ratio for a 2 x 2 table and is independent of the marginal totals. It has a range of 1 to 1. Also known as the coefficient of colligation. Yule's Q. This index is a special case of Goodman and Kruskal's gamma. It is a function of the cross-ratio and is independent of the marginal totals. It has a range of 1 to 1. You may optionally change the Present and Absent fields to specify the values that indicate that a characteristic is present or absent. The procedure will ignore all other values.
Additionally, you can choose how standardization is done. Alternatives are By variable or By case.
From the menus choose: Analyze > Classify > Hierarchical Cluster... In the Hierarchical Cluster Analysis dialog box, click Save.
Note: These options are available only if you select the Iterate and classify method from the KMeans Cluster Analysis dialog box. Maximum Iterations. Limits the number of iterations in the k-means algorithm. Iteration stops after this many iterations even if the convergence criterion is not satisfied. This number must be between 1 and 999. To reproduce the algorithm used by the Quick Cluster command prior to version 5.0, set Maximum Iterations to 1. Convergence Criterion. Determines when iteration ceases. It represents a proportion of the minimum distance between initial cluster centers, so it must be greater than 0 but not greater than 1. If the criterion equals 0.02, for example, iteration ceases when a complete iteration does not move any of the cluster centers by a distance of more than 2% of the smallest distance between any initial cluster centers. Use running means. Allows you to request that cluster centers be updated after each case is assigned. If you do not select this option, new cluster centers are calculated after all cases have been assigned.
Initial cluster centers. First estimate of the variable means for each of the clusters. By default, a number of well-spaced cases equal to the number of clusters is selected from the data. Initial cluster centers are used for a first round of classification and are then updated. ANOVA table. Displays an analysis-of-variance table which includes univariate F tests for each clustering variable. The F tests are only descriptive and the resulting probabilities should not be interpreted. The ANOVA table is not displayed if all cases are assigned to a single cluster. Cluster information for each case. Displays for each case the final cluster assignment and the Euclidean distance between the case and the cluster center used to classify the case. Also displays Euclidean distance between final cluster centers.
Missing Values. Available options are Exclude cases listwise or Exclude cases pairwise.
Exclude cases listwise. Excludes cases with missing values for any clustering variable from the analysis. Exclude cases pairwise. Assigns cases to clusters based on distances that are computed from all variables with nonmissing values. This feature requires the Statistics Base option. From the menus choose: Analyze > Classify > K-Means Cluster... In the K-Means Cluster dialog box, click Options.
Distances
This procedure calculates any of a wide variety of statistics measuring either similarities or dissimilarities (distances), either between pairs of variables or between pairs of cases. These similarity or distance measures can then be used with other procedures, such as factor analysis, cluster analysis, or multidimensional scaling, to help analyze complex datasets. Example. Is it possible to measure similarities between pairs of automobiles based on certain characteristics, such as engine size, MPG, and horsepower? By computing similarities between autos, you can gain a sense of which autos are similar to each other and which are different from
each other. For a more formal analysis, you might consider applying a hierarchical cluster analysis or multidimensional scaling to the similarities to explore the underlying structure. Statistics. Dissimilarity (distance) measures for interval data are Euclidean distance, squared Euclidean distance, Chebychev, block, Minkowski, or customized; for count data, chi-square or phi-square; for binary data, Euclidean distance, squared Euclidean distance, size difference, pattern difference, variance, shape, or Lance and Williams. Similarity measures for interval data are Pearson correlation or cosine; for binary data, Russel and Rao, simple matching, Jaccard, dice, Rogers and Tanimoto, Sokal and Sneath 1, Sokal and Sneath 2, Sokal and Sneath 3, Kulczynski 1, Kulczynski 2, Sokal and Sneath 4, Hamann, Lambda, Anderberg's D, Yule's Y, Yule's Q, Ochiai, Sokal and Sneath 5, phi 4-point correlation, or dispersion.
The Transform Measures group allows you to transform the values generated by the distance measure. They are applied after the distance measure has been computed. Available options are absolute values, change sign, and rescale to 01 range. This feature requires the Statistics Base option. From the menus choose: Analyze > Correlate > Distances... With Dissimilarities selected, click Measures. From the Measure group, select the alternative that corresponds to your type of data. From the drop-down list, select a measure that corresponds to that type of measure.
Chi-square measure. This measure is based on the chi-square test of equality for two sets of frequencies. This is the default for count data. Phi-square measure. This measure is equal to the chi-square measure normalized by the square root of the combined frequency.
Interval data. Pearson correlation or cosine. Binary data. Russell and Rao, simple matching, Jaccard, Dice, Rogers and Tanimoto, Sokal and Sneath 1, Sokal and Sneath 2, Sokal and Sneath 3, Kulczynski 1, Kulczynski 2, Sokal and Sneath 4, Hamann, Lambda, Anderberg's D, Yule's Y, Yule's Q, Ochiai, Sokal and Sneath 5, phi 4-point correlation, or dispersion. (Enter values for Present and Absent to specify which two values are meaningful; Distances will ignore all other values.) The Transform Values group allows you to standardize data values for either cases or variables before computing proximities. These transformations are not applicable to binary data. Available standardization methods are z scores, range 1 to 1, range 0 to 1, maximum magnitude of 1, mean of 1, and standard deviation of 1. The Transform Measures group allows you to transform the values generated by the distance measure. They are applied after the distance measure has been computed. Available options are absolute values, change sign, and rescale to 01 range. This feature requires the Statistics Base option. From the menus choose: Analyze > Correlate > Distances... With Similarities selected, click Measures. From the Measure group, select the alternative that corresponds to your type of data. From the drop-down list, select a measure that corresponds to that type of measure.
The following similarity measures are available for binary data: Russel and Rao. This is a binary version of the inner (dot) product. Equal weight is given to matches and nonmatches. This is the default for binary similarity data. Simple matching. This is the ratio of matches to the total number of values. Equal weight is given to matches and nonmatches. Jaccard. This is an index in which joint absences are excluded from consideration. Equal weight is given to matches and nonmatches. Also known as the similarity ratio. Dice. This is an index in which joint absences are excluded from consideration, and matches are weighted double. Also known as the Czekanowski or Sorensen measure. Rogers and Tanimoto. This is an index in which double weight is given to nonmatches. Sokal and Sneath 1. This is an index in which double weight is given to matches. Sokal and Sneath 2. This is an index in which double weight is given to nonmatches, and joint absences are excluded from consideration. Sokal and Sneath 3. This is the ratio of matches to nonmatches. This index has a lower bound of 0 and is unbounded above. It is theoretically undefined when there are no nonmatches; however, Distances assigns an arbitrary value of 9999.999 when the value is undefined or is greater than this value. Kulczynski 1. This is the ratio of joint presences to all nonmatches. This index has a lower bound of 0 and is unbounded above. It is theoretically undefined when there are no nonmatches; however, Distances assigns an arbitrary value of 9999.999 when the value is undefined or is greater than this value. Kulczynski 2. This index is based on the conditional probability that the characteristic is present in one item, given that it is present in the other. The separate values for each item acting as predictor of the other are averaged to compute this value. Sokal and Sneath 4. This index is based on the conditional probability that the characteristic in one item matches the value in the other. The separate values for each item acting as predictor of the other are averaged to compute this value. Hamann. This index is the number of matches minus the number of nonmatches, divided by the total number of items. It ranges from -1 to 1. Lambda. This index is Goodman and Kruskal's lambda. Corresponds to the proportional reduction of error (PRE) using one item to predict the other (predicting in both directions). Values range from 0 to 1.
Anderberg's D. Similar to lambda, this index corresponds to the actual reduction of error using one item to predict the other (predicting in both directions). Values range from 0 to 1. Yule's Y. This index is a function of the cross-ratio for a 2 x 2 table, and is independent of the marginal totals. It has a range of -1 to 1. Also known as the coefficient of colligation. Yule's Q. This index is a special case of Goodman and Kruskal's gamma. It is a function of the cross-ratio and is independent of the marginal totals. It has a range of -1 to 1. Ochiai. This index is the binary form of the cosine similarity measure. It has a range of 0 to 1. Sokal and Sneath 5. This index is the squared geometric mean of conditional probabilities of positive and negative matches. It is independent of item coding. It has a range of 0 to 1. Phi 4-point correlation. This index is a binary analogue of the Pearson correlation coefficient. It has a range of -1 to 1. Dispersion. This index has a range of -1 to 1. You can optionally change the Present and Absent fields to specify the values that indicate that a characteristic is present or absent. The procedure will ignore all other values.
CLUSTER provides three procedures for clustering: Hierarchical Clustering, K-Clustering and Additive Trees. The Hierarchical Clustering procedure comprises hierarchical linkage methods. The K-Clustering procedure splits a set of objects into a selected number of groups by maximizing between-cluster variation and minimizing within-cluster variation. The Additive Trees Clustering procedure produces a Sattath-Tversky additive tree clustering.
Hierarchical Clustering clusters cases, variables, individually or both cases and variables simultaneously; K-Clustering clusters cases only; and Additive Trees clusters a similarity or dissimilarity matrix. Several distance metrics are available with Hierarchical Clustering and K-Clustering including metrics for binary, quantitative and frequency count data. Hierarchical Clustering has ten methods for linking clusters and displays the results as a tree (dendrogram) or a polar dendrogram. When the MATRIX option is used to cluster cases and variables, SYSTAT uses a gray-scale or color spectrum to represent the values. Resampling procedures are available only in Hierarchical Clustering. SYSTAT further provides five indices, viz., statistical criteria by which an appropriate number of clusters can be chosen from the Hierarchical Tree. Options for cutting (or pruning) and coloring the hierarchical tree are also provided. In the K-Clustering procedure SYSTAT offers two algorithms, KMEANS and KMEDIANS, for partitioning. Further, SYSTAT provides nine methods for selecting initial seeds for both KMEANS and KMEDIANS.
Hierarchical clustering produces hierarchical clusters that are displayed in a tree. Initially, each object (case or variable) is considered a separate cluster. SYSTAT begins by joining the two closest objects as a cluster and continues (in a stepwise manner) joining an object with another object, an object with a cluster, or a cluster with another cluster until all objects are combined into one cluster. You must select the elements of the data file to cluster (Join):
Rows. Rows (cases) of the data matrix are clustered. Columns. Columns (variables) of the data matrix are clustered. Matrix. Rows and columns of the data matrix are clusteredthey are permuted to bring similar rows and columns next to one another.
Linkage allows you to specify the type of joining algorithm used to amalgamate clusters (that is, define how distances between clusters are measured).
Average. Average linkage averages all distances between pairs of objects in different clusters to decide how far apart they are. Centroid. Centroid linkage uses the average value of all objects in a cluster (the cluster centroid) as the reference point for distances to other objects or clusters. Complete. Complete linkage uses the most distant pair of objects in two clusters to compute between-cluster distances. This method tends to produce compact, globular clusters. If you use a similarity or dissimilarity matrix from a SYSTAT file, you get Johnsons max method.
Flexibeta. Flexible beta linkage uses a weighted average distance between pairs of objects in different clusters to decide how far apart they are. You can choose the value of the weight . The range of is between 1 and 1. K-nbd.Kth nearest neighborhood method is a density linkage method. The estimated density is proportional to the number of cases in the smallest sphere containing the Kth nearest neighbor. A new dissimilarity matrix is then constructed using the density estimate. Finally the single linkage cluster analysis is performed. You can specify the number k; its range is between 1 and the total number of cases in the dataset. Median. Median linkage uses the median distances between pairs of objects in different clusters to decide how far apart they are. Single. Single linkage defines the distance between two objects or clusters as the distance between the two closest members of those clusters. This method tends to produce long, stringy clusters. If you use a SYSTAT file that contains a similarity or dissimilarity matrix, you get clustering via Johnsons min method. Uniform. Uniform Kernel method is a density linkage method. The estimated density is proportional to the number of cases in a sphere of radius r. A new dissimilarity matrix is then constructed using the density estimate. Finally single linkage cluster analysis is performed. You can choose the number r; its range is the positive real line. Ward. Wards method averages all distances between pairs of objects in different clusters, with adjustments for covariances, to decide how far apart the clusters are. Weighted. Weighted average linkage uses a weighted average distance between pairs of objects in different clusters to decide how far apart they are. The weights used are proportional to the size of the cluster.
For some data, some methods cannot produce a hierarchical tree with strictly increasing amalgamation distances. In these cases, you may see stray branches that do not connect to others. If this happens, you should consider Single or Complete linkage. For more information on these problems, see Fisher and Van Ness (1971). These reviewers concluded that these and other problems made Centroid, Average, Median, and Ward (as well as K-Means) inadmissible clustering procedures. In practice and in Monte Carlo simulations, however, they sometimes perform better than Single and Complete linkage, which Fisher and Van Ness considered admissible. Milligan (1980) tested all of the hierarchical joining methods in a large Monte Carlo simulation of clustering algorithms. Consult his paper for further details. In addition, the following options can be specified:
Distance. Specifies the distance metric used to compare clusters. Polar. Produces a polar (circular) cluster tree. Save. Save provides two options either to save cluster identifiers or to save cluster identifiers along with data. You can specify the number of clusters to identify for the saved file. If not specified, two clusters are identified. In the Mahalanobis Tab, you can specify the covariance matrix to compute Mahalanobis distance.
Covariance matrix. Specify the covariance matrix to compute the Mahalanobis distance. Enter the covariance matrix either through the keyboard or from a SYSTAT file. Otherwise, by default SYSTAT computes the matrix from the data. Select a grouping variable for inter-group distance measures.
Cut cluster tree at. You can choose the following options for cutting the cluster tree:
Height. Provides the option of cutting the cluster tree at a specified distance. Leaf nodes. Provides the option of cutting the cluster tree by number of leaf nodes.
Color clusters by. The colors in the cluster tree can be assigned by two different methods:
Length of terminal node. As you pass from node to node in order down the cluster tree, the color changes when the length of a node on the distance scale changes between less than and greater than the specified length of terminal nodes (on a scale of 0 to 1). Proportion of total nodes. Colors are assigned based on the proportion of members in a cluster.
Validity.Provides five validity indices to evaluate the partition quality. In particular, it is used to find out the appropriate number of clusters for the given data set.
RMSSTD. Provides root-mean-square standard deviation of the clusters at each step in hierarchical clustering. Pseudo F. Provides pseudo F-ratio for the clusters at each step in hierarchical clustering. Pseudo T-square. Provides pseudo T-square statistic for cluster assessment. DB. Provides Davies-Bouldins index for each hierarchy of clustering. This index is applicable for rectangular data only. Dunn. Provides Dunns cluster separation measure. Maximum groups. Performs the computation of indices up to this specified number of clusters. The default value is square root of number of objects.
K-Clustering dialog box provides the options for K-Means clustering and K-Medians clustering. Both clustering methods splits a set of objects into a selected number of groups by maximizing between-cluster variation relative to the within-cluster variation. It is similar to doing a one-way analysis of variance where the groups are unknown and the largest F value is sought by reassigning members to each group. By default, the algorithms start with one cluster and split it into two clusters by picking the case farthest from the center as a seed for a second cluster and assigning each case to the nearest center. It continues splitting one of the clusters into two (and reassigning cases) until a specified number of clusters are formed. The reassigning of cases continue until the within-groups sum of squares can no longer be reduced. The initial seeds or partitions can be chosen from a possible set of nine options. Algorithm. Provides K-Means and K-Medians clustering options.
Groups. Enter the number of desired clusters. Default number (Groups) is two. Iterations. Enter the maximum number of iterations. If not stated, the maximum is 20. Distance. Specifies the distance metric used to compare clusters. Save. Save provides three options to save either cluster identifiers, cluster identifiers along with data, or final cluster seeds, to a SYSTAT file. Click More above for descriptions of the distance metrics.
None. Starts with one cluster and splits it into two clusters by picking the case farthest from the center as a seed for the second cluster and then assigning each case optimally. It continues splitting and reassigning the cases until k clusters are formed. First k.Considers the first k non-missing cases as initial seeds. Last k.Considers the last k non-missing cases as initial seeds. Random k. Chooses randomly (without replacement) k non-missing cases as initial seeds. Random segmentation. Assigns each case to any of k partitions randomly. Computes seeds from each initial partition taking the mean or the median of the observations, whichever is applicable.