advertisement

Assessing clustering quality Scoring clustering solutions by their biological relevance Common Summary Motivation. Existent solutions and their problems. Background (Average Silhouette, Jaccard coefficient, Homogeneity and Separation, ANOVA, KW-test). The Clustering Quality Score (CQS) method. Results (method implementation on real and simulated data). Conclusion. Motivation Different clustering algorithms yield different clustering solutions on the same data. The same algorithm yields different results for different parameter settings. Few works addressed the systematic comparison and evaluation of clustering results. There is no consensus on choosing among the clustering algorithms. Quality assessing - existent state Different measures for the quality of a clustering solution are applicable in different situations. It depends on the data and on the availability of the true solution. In case the true solution is known, and we wish to compare it to another solution, we can use the Minkowski measure or the Jaccard coefﬁcient. When the true solution is not known, there are some approaches (Homogeneity and Separation, Average silhouette) ,but there is no agreed-upon approach for evaluating the quality of a suggested solution. Homogeneity and Separation approach The method is based on intra-cluster homogeneity or inter-cluster separation . The problem that the homogeneity and separation criteria are inherently conﬂicting, as an improvement in one will usually correspond to worsening of the other. The ways to solve this problem are: - to ﬁx the number of clusters and seek a solution with maximum homogeneity (K-means algorithm). - present a curve of homogeneity versus separation (a curve can show that one algorithm dominates another if it provides better homogeneity for all separation values, but typically different algorithms will dominate in different value range). Quality assessing - existent state Clustering solutions can be assessed by applying standard statistical techniques: multivariate analysis of variance (MANOVA) for high-dimensional data and discriminant analysis for normally distributed data. For the case of non-normal data, there are several extensions that require the data to be either lowdimensional or continuous. None of these methods applicable when wishing to test the signiﬁcance of a clustering solution based on highdimensional vectors of dependent biological attributes that do not necessarily follow a normal distribution and may even be discrete. Background Definitions N - a set of n elements , C ={C1, ... , Cl } - a partition of these elements into l clusters. Mates - two elements from the same cluster (with respect to C). Homogeneity of C is the average distance between mates (similarity). Separation of C is the average distance between non-mates (dissimilarity). Jaccard coefficient Jaccard coefficient is the proportion of correctly identiﬁed mates out of the sum of the correctly identiﬁed mates plus the total number of disagreements (pairs of elements that are mates in exactly one of the two solutions). A perfect solution has score 1, and the higher the score – the better the solution. The method is useful only when the true solution is known. Average silhouette Deﬁnition of the silhouette of element j is (bj − aj)/ max(aj , bj) aj is the average distance of element j from other elements of its cluster, bjk is the average distance of element j from the members of cluster Ck , and bj = min {k : j not belongs to Ck } bjk. The average silhouette is the mean of this ratio over all elements (genes). This method performs well in general, but fails to detect fine cluster structures. The use of external information The main focus is the evaluation of clustering solutions using external information. Given an n × p attribute matrix A. The rows of A correspond to elements, and the i-th row vector is called the attribute vector of element i. Also given a clustering C ={C1, ... , Cl } of the elements, where Si =|Ci |. We index the attribute vectors by the clustering: as the vector of element j in cluster i. C is obtained without using the information in A. The goal is to evaluate C with respect to A. ANOVA test When p = 1. The attribute is normally distributed, The variances of the l population distributions are identical. We can use standard analysis of variance (ANOVA): - aij - the attribute of element j in cluster i. - a¯i - the mean of the elements in cluster i. - a¯- the total mean of all n elements. - H0: µ1 = µ2 = ··· = µl , where µi is the expectation of group i. ANOVA test The test statistic is FH: SSH = SSE = FH statistic has a F distribution with l −1 and n − l degrees of freedom. For the multidimensional case (p > 1), the MANOVA test applies the same objective function FH if the attribute matrix is multinormally distributed. Kruskal-Wallis (KW) test In case the attribute does not follow a normal distribution, we can use the KW test as a nonparametric ANOVA test. The test assumes that the clusters are independent and have similar shape. We shall denote by PKW(C, A) the p-value obtained by the KW test for a clustering C using the attribute A. CQS Method Will be introduced statistically based method for comparing clustering solutions according to prior biological knowledge. The solutions are ranked according to their correspondence to prior knowledge about the clustered elements. The method tests the dependency between the attributes and the grouping of the elements. The method computes a quality score for the functional enrichment of attribute classes among each solution’s clusters. The results are represented as the CQS (Clustering Quality Score) of the clustering. Computing a linear combination of the attributes Each element is assigned a real value, which is a weighted sum of its attributes. An attribute’s weight is its coefﬁcient in the linear combination. The proposition is to use weights that maximize the ability to discriminate between the clusters using the one-dimensional data. Finding the weights will be done in the same manner as in Linear Discriminant Analysis (LDA). LDA creates a linear combination by maximizing the ratio of between-groups-variance to within-groups-variance. Computing a linear combination of the attributes The statistic being maximized is the ratio of MSH to MSE: The mean vector of cluster i The total mean vector The p-dimensional vector of weights Computing a linear combination of the attributes The maximum value of F(w) is proportional to the greatest root of the equation | H – λE | = 0 . H is a p ×p matrix containing the between-groups sum of square: E is a p ×p matrix of the sum of squared errors: The desired combination w is the eigenvector corresponding to the greatest root. This result holds without assuming any prior distribution on the attributes. The total mean of attribute r The mean of attribute r in cluster i Projection Applying the linear combination w to the attribute vectors , thereby projecting these vectors onto the real line: Computing CQS using the projected values We have to evaluate the clustering vis-á-vis the projected attributes using the KW test. The value of CQS = (– log p). p = PKW(C, Z), - the p-value assigned to the clustering by the KW test. The p-value is the probability that all values in this particular projection have been taken from the same population. CQS favors clustering solutions whose best discriminating weights enable significant grouping. Estimating confidence Estimation of the accuracy of the scores and the signiﬁcance by evaluating the sensitivity of CQS to small modiﬁcations of the clustering solution. Each alternative solution is obtained by introducing k exchanges of random pairs of elements from different clusters of the original solution (k is elements). The larger the inﬂuence of small perturbations in the clustering on the CQS value - the smaller the conﬁdence we have in the CQS. The CQS conﬁdence is the standard deviation of CQS for the group of alternative clustering solutions. The overall procedure 1. Finding the eigenvector w corresponding to the greatest root of the system of equations |H −λE|=0. 2. For each attribute vector aij set zij = ∑t aijwt 3. Compute p = PKW(C, Z); let CQS(C, A) = −log p. 4. Estimate the statistical conﬁdence of the result by perturbations on C. Results on simulated data The effect of one-dimensional projection. The effect of solution accuracy on CQS. Sensitivity of CQS to the number of clusters. CQS ability to detect fine clustering structures. Simulation 80 binary attributes. 5 groups of n=50 genes each. For each attribute randomly selected one group in which its frequency will be r. In the other 4 groups its freq. was set to r0. The set of r (r0) genes with that attribute was randomly selected from the relevant groups. The larger the difference between r and r0, the easier the distinction between the groups. The effect of one-dimensional projection Was simulated data with r = 6,10,15,20 and 25. Was computed the ratio of separation to homogeneity of the true clustering on the original data (S/ H) and on the projected data (S*/ H*). The procedure was repeated 10 times. The effect of one-dimensional projection - results Was found the monotonicity of the homogeneity, separation and their ratio as a function of r on the original data and on the reduced data. CQS improves monotonically with r. Box plots for the projection of 5 simulated clusters after dimensionality reduction The projected data with r = 6. The clusters look very similar, even though the data were reduced using the best separating linear combination. The projected data with r = 25. Inter-cluster separation of most clusters is clearly visible. The y-axis is the real-valued projection of the elements. Each box-plot depicts the median of the distribution (dot), 0.1 and 0.9 distribution quantiles (white box), and the maximum and minimum values. The effect of solution accuracy on CQS - results The accuracy of different clustering solutions is measured by the number of inter-cluster exchanges introduced in the original solution. X-axis: number of exchanges. Y -axis: CQS (left scale) and Jaccard coefﬁcient (right scale). Was compared CQS of the true partition with that of other, similar and remote partitions. Those were produced by starting with the true solution and repeatedly exchanging a randomly chosen pair of elements from different clusters. CQS is highest for the true partition and decreases with the number of exchanges applied (200 exchanges generate an essentially random partition). The Jaccard coefﬁcients decrease with the number of exchanges. Sensitivity of CQS to the number of clusters (splitting) How CQS changes when splitting clusters? Was compared the true 5-cluster solution with a 25-cluster solution obtained by randomly splitting each of the 5 clusters into 5 equal-size sub-clusters. This test was repeated 10 times. Was observed a decrease of the clustering quality measures in all runs. The decrease of S/H is maintained in CQS and on the reduced data (S*/H* ). The diagram shows CQS, S*/H* and S/H (y-axis) of the true solution (gray) and the modiﬁed solution (black) Sensitivity of CQS to the number of clusters (merging) How CQS changes when merging clusters? Was simulated two 5-cluster data sets using n = 25. Then combined these data sets into a single data set whose true solution consists of 10 equal size clusters with 25 genes each. Was merged pairs of clusters, one from each original data set, to form in total 5 clusters with 50 genes each. These 5 clusters comprised the alternative (merged) solution. All the measures decrease due to the merging (as in splitting test). The decrease of S/H is maintained and enlarged in S*/H* and CQS. The diagram shows CQS, S*/H* and S/H (y-axis) of the true solution (gray) and the modiﬁed solution (black) Sensitivity of CQS to the number of clusters CQS and Jaccard coefficient ? Was simulated 5-cluster data and applied K-means to the data, with K = 2, ... ,15. Was computed CQS and Jaccard coefficient for each clustering solution. Sensitivity of CQS to the number of clusters - results CQS behaves as the Jaccard coefﬁcient and S/H , with a maximum at K = 5 (the true number of clusters). The ranking of all 14 solutions according to the Jaccard score (which is based on the true solution) and according to CQS (which is based on the attributes only) are virtually identical. The ratio score also does quite well, with a maximum at K = 5. However, the ranking of solutions by this score does not agree with the Jaccard score. CQS ability to detect fine clustering structures Proﬁles of 30 binary attributes were generated for four clusters of n = 50 genes each. For each attribute, its frequencies in clusters 1, 2, 3 and 4 were set to 2, b,50−b and 48, respectively. Was simulated data sets with b = 3, 5, 10, 15, 20. For each data set, scored two clustering solutions: the original 4cluster solution, and a 2-cluster solution obtained by merging cluster 1 with 2 and merging cluster 3 with 4. For large values of b we expect the 4-cluster solution to score higher than the 2-cluster solution. For each data set and each of the two solutions, was computed S/H , CQS and the average silhouette score. CQS ability to detect fine clustering structures - results CQS S/H Silhouette The ratios are increasing with b in all scores. The silhouette for the 2-cluster solution is always greater than for the corresponding 4-cluster solution. For b=3,5,10,15, S/H is greater for the incorrect 2-cluster solution. CQS is able to identify the ﬁne structure in the data: for all b values except b = 3. For b = 3, the 2-cluster CQS is higher than the 4-cluster CQS, since there is almost no difference between the clusters with 2 or 3 occurrences of attributes, and between the clusters with 47 or 48 occurrences. Yeast cell-cycle data The tested data is from yeast cell-cycle data. The data set contains 698 genes and 72 conditions. Each row of the 698×72 matrix was normalized to have mean 0 and variance 1. We expect to ﬁnd in the data 5 main clusters. The 698×72 data set was clustered using four clustering methods: K-means, SOM, CAST and CLICK. Yeast cell-cycle data The results are: K-means : 5 clusters SOM : 6 clusters CAST : 5 clusters CLICK : 6 clusters + 23 singletons Also, the genes was manually divided into 5 groups using their peak of expression – the ‘true’ solution. Was scored a random clustering of the data into 5 equal-size clusters. Yeast cell-cycle data As gene attributes was used the GO classes and MIPS annotation. Overall was used: 51 GO process attributes, 37 GO function attributes, 27 GO component attributes and 59 MIPS attributes. CQS was computed 3 times: using the GO process attributes only, all GO attributes, and the MIPS attributes only. The experiment shows that different biological attributes lead to different evaluations of clustering solutions. Yeast cell-cycle data Clustering solutions is computed using GO level 5 process attributes only: Yeast cell-cycle data Clustering solutions is computed using all GO level 5 attributes : Yeast cell-cycle data Clustering solutions is computed using MIPS level 4 attributes : Yeast cell-cycle data The 22 most enriched attributes for CLICK solution using all GO attributes: Yeast cell-cycle data The distribution and co-occurrence of the attributes ‘DNA metabolism’, ‘DNA replication’ and ‘Chromosome organization’ in the 6 clusters of the CLICK solution : Conclusion CQS method based on biological relevance using attributes of the clustered elements, which are available independently from the data used to generate the clusters. The method can be applied : - to compare the functional enrichment of many biological attributes simultaneously in different clustering solutions. - to optimize the parameters of a clustering algorithm . Conclusion The method outperforms previous numeric methods (the S/H ratio and the average silhouette measure). CQS is sensitive to small modifications of the clustering solution and to changes in the simulation setting. The attribute weights were computed using information about all the attributes together, without assuming that the attributes are independent. CQS has the advantage that it can use such continuous data without any assumption on the data distribution.