N. Kumar, Asst. Professor of Marketing Database Marketing Cluster Analysis
Feb 24, 2016
N. Kumar, Asst. Professor of Marketing
Database Marketing
Cluster Analysis
N. Kumar, Asst. Professor of Marketing
2
AgendaDiscussion of the first Assignment
Motivation for conducting Cluster AnalysisBenefit Segmentation
Cluster AnalysisBasic ConceptsHierarchical/Non- Hierarchical Clustering
Implementation in SAS and interpreting the output
N. Kumar, Asst. Professor of Marketing
Voter ProfilingWhat are the different voting segments out there? What do they want to hear i.e. issues they care about?
What should I say?
N. Kumar, Asst. Professor of Marketing
Ad CampaignHow many customer segments are there?
How many do I want to target?
How should I target – what message should I communicate to each segment?
N. Kumar, Asst. Professor of Marketing
Promotional StrategiesCoupon Drops – who should they be targeted at?
Catalog Example – should the catalog be accompanied with a $5 coupon or a $10 coupon or no coupon?
N. Kumar, Asst. Professor of Marketing
What is Cluster Analysis? Cluster Analysis is a technique for combining observations into groups or clusters such that:
Each group is homogenous with respect to certain characteristics (that you specify)Each group is different from the other groups with respect to the same characteristics
N. Kumar, Asst. Professor of Marketing
DataConsumer Income ($ 1000s) Education (years)
1 5 5
2 6 6
3 15 14
4 16 15
5 25 19
6 30 20
N. Kumar, Asst. Professor of Marketing
Geometrical View of Cluster Analysis Education
Income
N. Kumar, Asst. Professor of Marketing
Similarity MeasuresWhy are consumers 1 and 2 similar?
Distance(1,2) = (5-6)2 + (5-6)2
More generally, if there are p variables: Distance(i,j) = (xik - xjk)2
N. Kumar, Asst. Professor of Marketing
Similarity MatrixC1 C2 C3 C4 C5 C6
C1 0 2 181 221 625 821
C2 2 0 145 181 557 745
C3 181 145 0 2 136 250
C4 221 181 2 0 106 212
C5 625 557 136 106 0 26
C6 821 745 250 212 26 0
N. Kumar, Asst. Professor of Marketing
Clustering Techniques
Hierarchical Clustering
Non-Hierarchical Clustering
N. Kumar, Asst. Professor of Marketing
Hierarchical ClusteringDistance(1,2) = 2 = Distance(3,4)Say, we group 1 and 2 together and leave the others as isHow do we compute the distance between a group that has two (or more) members and the others?
N. Kumar, Asst. Professor of Marketing
Hierarchical Clustering Algorithms
Centroid MethodNearest-Neighbor or Single-LinkageFarthest-Neighbor or Complete-LinkageAverage-LinkageWard’s Method
N. Kumar, Asst. Professor of Marketing
Centroid MethodEach group is replaced by an average consumerCluster 1 – average income = 5.5 and average education = 5.5
N. Kumar, Asst. Professor of Marketing
Data for Five ClustersCluster Members Income Education
1 C1&C2 5.5 5.5
2 C3 15 14
3 C4 16 15
4 C5 25 20
5 C6 30 19
N. Kumar, Asst. Professor of Marketing
Similarity MatrixC1&C2 C3 C4 C5 C6
C1&C2 0
C3 162.5 0
C4 200.5 2 0
C5 590.5 135.96 106 0
C6 782.5 250 212 26 0
N. Kumar, Asst. Professor of Marketing
Data for Four ClustersCluster Members Income Education
1 C1&C2 5.5 5.5
2 C3&C4 15.5 14.5
3 C5 25 20
4 C6 30 19
N. Kumar, Asst. Professor of Marketing
Similarity MatrixC1&C2 C3&C4 C5 C6
C1&C2 0
C3&C4 181 0
C5 590 120.5 0
C6 782.5 230.5 26 0
N. Kumar, Asst. Professor of Marketing
Data for Three ClustersCluster Members Income Education
1 C1&C2 5.5 5.5
2 C3&C4 15.5 14.5
3 C5&C6 27.5 19.5
N. Kumar, Asst. Professor of Marketing
Similarity MatrixC1&C2 C3&C4 C5&C6
C1&C2 0
C3&C4 181 0
C5&C6 680 169 0
N. Kumar, Asst. Professor of Marketing
Dendogram for the Data
C1 C2 C3 C4 C5 C6
N. Kumar, Asst. Professor of Marketing
Single LinkageFirst Cluster is formed in the same fashionDistance between Cluster 1 comprising of customers 1 and 2 and customer 3 is the minimum of Distance(1,3) = 181 and Distance(2,3) = 145
N. Kumar, Asst. Professor of Marketing
Similarity MatrixC1&C2 C3 C4 C5 C6
C1&C2 0
C3 145 0
C4 181 2 0
C5 557 136 106 0
C6 745 250 212 26 0
N. Kumar, Asst. Professor of Marketing
Complete Linkage
Distance between Cluster 1 comprising of customers 1 and 2 and customer 3 is the maximum of Distance(1,3) = 181 and Distance(2,3) = 145
N. Kumar, Asst. Professor of Marketing
Similarity MatrixC1&C2 C3 C4 C5 C6
C1&C2 0
C3 181 0
C4 221 2 0
C5 625 136 106 0
C6 821 250 212 26 0
N. Kumar, Asst. Professor of Marketing
Average Linkage
Distance between Cluster 1 comprising of customers 1 and 2 and customer 3 is the average of Distance(1,3) = 181 and Distance(2,3) = 145
N. Kumar, Asst. Professor of Marketing
Similarity MatrixC1&C2 C3 C4 C5 C6
C1&C2 0
C3 163 0
C4 201 2 0
C5 591 136 106 0
C6 783 250 212 26 0
N. Kumar, Asst. Professor of Marketing
Ward’s MethodDoes not compute distance between clustersForms clusters by maximizing within-cluster homogeneity or minimizing error sum of squares (ESS)ESS for cluster with two observations (say, C1 and C2) = (5-5.5)2 + (6-5.5)2 + (5-5.5)2 + (6-5.5)2
N. Kumar, Asst. Professor of Marketing
Ward’s MethodCL1 CL2 CL3 CL4 CL5 ESS
1 C1,C2 C3 C4 C5 C6 12 C1,C3 C2 C4 C5 C6 90.53 C1,C4 C2 C3 C5 C6 110.54 C1,C5 C2 C3 C4 C6 312.55 C1,C6 C2 C3 C4 C5 410.56 C2,C3 C1 C4 C5 C6 72.57 C2,C4 C1 C3 C5 C6 90.5
N. Kumar, Asst. Professor of Marketing
Non-Hierarchical ClusteringData are grouped into K clusters
Requires a priori knowledge of K
N. Kumar, Asst. Professor of Marketing
Basic Steps in Non-Hierarchical Clustering
Select K initial cluster centroidsAssign each observation to the cluster to which it is closestReassign or reallocate each observation to one of the K clusters according to a pre-determined stopping ruleStop if there is no reallocationApproaches differ in Step 1 and/or step 3
N. Kumar, Asst. Professor of Marketing
Algorithm ISelects first K observations as cluster centers
N. Kumar, Asst. Professor of Marketing
Initial Cluster CentroidsVariable CL1 CL2 CL3
Income 5 6 15
Education 5 6 14
N. Kumar, Asst. Professor of Marketing
Initial AssignmentDistance from C1
Distance from C2
Distance from C3
Assigned to CL
C1 0 2 181 1
C2 2 0 145 2
C3 181 145 0 3
C4 221 181 2 3
C5 625 557 136 3
C6 821 745 250 3
N. Kumar, Asst. Professor of Marketing
New Cluster CentroidsVariable CL1 CL2 CL3
Income 5 6 21.5
Education 5 6 17
N. Kumar, Asst. Professor of Marketing
Distance MatrixDistance from CL1
Distance from CL2
Distance from CL3
Previous Assignment
Current Assignment
C1 0 2 416.15 1 1
C2 2 0 316.25 2 2
C3 181 145 51.25 3 3
C4 221 181 34.25 3 3
C5 625 557 21.25 3 3
C6 821 990 76.25 3 3
N. Kumar, Asst. Professor of Marketing
Algorithm IIDiffers from Algorithm I in how the initial seeds are modifiedAs before first K observations are selected as the initial cluster seedsA seed that is a candidate for replacement is from one of the two seeds that are closest to each otherAn observation qualifies to replace one of the two candidates if the distance between the seeds is less than the distance between the observation and the closest seed
N. Kumar, Asst. Professor of Marketing
Algorithm II …contd.C1, C2 and C3 are the initial seedsThe smallest distance between the seeds is between C1 and C2Observation C4 does not qualify as a replacement as Distance(C1,C2) > Distance(C4 and the nearest seed C3)Observation C5 does qualify as a replacement as Distance(C1,C2) < Distance(C5 and the nearest seed C3): replace C2 with C5
N. Kumar, Asst. Professor of Marketing
Initial AssignmentDistance from C1
Distance from C2
Distance from C3
Assigned to CL
C1 0 181 625 1
C2 2 145 557 1
C3 181 0 136 2
C4 221 2 106 2
C5 625 136 0 3
C6 821 250 26 3
N. Kumar, Asst. Professor of Marketing
New Cluster CentroidsVariable CL1 CL2 CL3
Income 5.5 15.5 27.5
Education 5.5 14.5 19.5
N. Kumar, Asst. Professor of Marketing
Distance MatrixDistance from CL1
Distance from CL2
Distance from CL3
Previous Assignment
Current Assignment
C1 0.5 200.5 716.5 1 1
C2 0.5 162.5 644.5 1 1
C3 162.5 0.5 186.5 2 2
C4 200.5 0.5 152.5 2 2
C5 590.5 120.5 6.5 3 3
C6 600.50 230.5 6.5 3 3
N. Kumar, Asst. Professor of Marketing
Hierarchical vs. Non-Hierarchical Clustering
Hierarchical clustering does not require a priori knowledge of the number of clustersAssignments are staticUse hierarchical clustering for exploratory purposesNon-Hierarchical Methods can be viewed as a complementary rather than a competing method
N. Kumar, Asst. Professor of Marketing
Voter ProfilingSurvey of voters concerns may help us group customers with similar concerns – perhaps they all live in a certain area?
Target ads/mailings with customized messages
N. Kumar, Asst. Professor of Marketing
Ad CampaignUse attitudinal data to segment customers
Target message appropriately
N. Kumar, Asst. Professor of Marketing
Promotional StrategiesUse transaction data to group customers into those that are more prone to purchasing the product on deal
Give a stronger incentive to the price sensitive segment