Segmentation

Dendrogram. The dendrogram is a tree diagram to illustrate the arrangement of clusters produce Scree plot The screeplot displays, for each cluster solution, a measure of within-cluster heterogene observations that are widely different (which will happen if the number of clusters is too variability in the data), the value will be high. A good cluster solution might be where the screeplot displays an 'elbow', that is, where i clusters beyond a certain point does not dramatically decreases within-cluster heterogen The measure displayed in the screeplot is related, but not equivalent, to the distance rep

ed by hierarchical clustering, and how the observations are incrementally clustered together. eity. If clusters group o small to capture the increasing the number of neity. ported in the dendrogram.

Segment differences per segment. Cell colors indicate to what extent a segment is statistically diff Segmentation space The chart below is a graphical representation of the various segments, segment member is obtained by plotting the first two dimensions of a principal component analysis perform segmentation data, on top of which segment information has been overlaid. Because only the first two dimensions of the PCA are displayed, and these two dimensio variance in the data, some differences between segments might not appear here. Note t variance, if any, have been excluded. Two clusters that appear to overlap on the first two dimensions might be distinct on othe this chart is a useful guide, for checking which variables are correlated, but may be misle optimal number of segments.

65 2 66 1 67 2 68 2 69 1 70 2 71 2 72 2 73 2 74 2 75 1 76 1 77 2 78 2 79 2 80 2 81 1 82 1 83 2 84 1 85 2 86 1 87 1 88 1 89 2 90 1 91 2 92 2 93 2 94 2 95 1 96 2 97 2 98 1 99 2 100 2 101 1 102 2 103 2 104 1 105 2 106 2 107 2 108 2 109 1 110 2 111 2

206 2 207 2 208 2 209 2 210 2 211 2 212 2 213 2 214 2 215 2 216 2 217 2 218 2 219 2 220 2 221 2

component analysis. only available in the Excel formatted output.

Spider chart Spider chart comparing the averages of the segmentation variables across all segments. Spider chart. Segment 1 profile The following charts represent the profile of each segment. These charts are only availab standardized, hence the model assumes that all segmentation variables use the same sca - For each segment, the segmentation variables are ordered in decreasing order of mag - The colored dots represent the average of the segment. - The horizontal lines represent the standard deviations within that segment.

Segment 1 profile. Segment 2 profile - The vertical, gray lines represent the averages of the rest of the population, after exclu

Segment 2 profile.

ble when the data are not ale. gnitude.

uding members of the segment under scrutiny.

Descriptors This table reports the descriptor averages of each segment. The more differences can be predict segment membership based on descriptors alone. Descriptor data per segment Average value of each descriptor, overall and within each cluster. Descriptors that are statistically Population Segment 1 Fst2Try 2.67 2.16 Fashnista 3.34 2.69 PplThnkMe 3.63 3.05 PplAskMe 3.35 2.54 FashnMzne 3.93 2.75 NoBgDeal 2.95 3.30 NotBld 4.29 4.38 OnlyBest 5.12 4.72 ExclsvShop 3.12 2.07 WhrFrndShop 2.41 1.56 OtherPpl 3.72 2.64 MagznNews 3.90 3.15 Catlg 4.48 3.75 Internet 2.25 1.62 Excitbl 4.16 4.39 Thrfty 4.00 3.57 Rationl 3.50 3.03 Yuthfl 4.34 4.80 Orthdx 5.04 5.03 Modst 3.32 3.00 Age 2.45 2.84 ApparelEx 2.03 1.64 Income 2.70 2.54 Edu 1.81 1.78

Descriptor differences per segment. Cell colors indicate to what extent the distribution of a descr Descriptor space The chart below is a graphical representation of the various segments, segment member by outputing the first two dimensions of a principal component analysis performed on th top of which segment information has been overlayed. Because only the first two dimensions of the PCA are displayed, and these two dimensio variance in the data, some differences between segments might not appear here. Note t if any, have been excluded. If two or more segments fully overlap, it is unlikely that they could be clearly separated b However, two segments that seem to overlap on two dimensions may be more clearly se Consequently, the confusion matrix is a better guide to assess the quality of segment cla descriptors.

Descriptor space. Spatial representation of segments and their descriptors, using principal compo

e found, the easier it will be to y different from the rest of the population are highlighted in red (lower) or green (higher). Segment 2 2.87 3.59 3.86 3.66 4.38 2.82 4.26 5.28 3.52 2.74 4.14 4.19 4.75 2.49 4.08 4.17 3.68 4.17 5.04 3.44 2.30 2.18 2.76 1.82

riptor in a segment is statistically different from the rest of the population. rs, and descriptors. It is obtained he (standardized) descriptors, on ons capture only 31.2% of the that descriptors with no variance, based on descriptors alone. eparated on other dimensions. assification based on

onent analysis.

Introduction Often, segmentation (needs) variables for each customer may not be available to manag customers may be available. In this section, we explore whether descriptors alone can predict segment membership w confusion matrix and hit rates (reported below) indicate whether the model is accurate e For member classification based on descriptors, Enginius uses a multinomial logit model predict 'choices between multiple alternatives (A/B/C)' in the predictive modeling modul The largest segment is selected as the default option (dummy), and the model identifies significant for predicting cluster memberships. If a descriptor is highly predictive, its p-va zero, and the cells will appear in green (or red). Model coefficients Model parameters Segment 2 is the model baseline. Segment summary(model)$coefficients (Intercept) 5.58 Fst2Try 0.038 Fashnista -0.046 PplThnkMe 0.017 PplAskMe 0.172 FashnMzne -0.435 NoBgDeal 0.164 NotBld -0.063 OnlyBest -0.068 ExclsvShop -0.329 WhrFrndShop -0.318 OtherPpl -0.299 MagznNews 0.113 Catlg -0.069 Internet -0.284 Excitbl -0.203 Thrfty -0.070 Rationl -0.220 Yuthfl 0.074 Orthdx 0.079 Modst -0.123 Age 0.117 ApparelEx -0.525 Income -0.064 Edu -0.006 P-values

p-values Probabilities that parameter estimates are different from zero only by chance. Segment summary(model)$coefficients (Intercept) 0.017 Fst2Try 0.855 Fashnista 0.808 PplThnkMe 0.935 PplAskMe 0.372 FashnMzne 0.006 NoBgDeal 0.320 NotBld 0.677 OnlyBest 0.633 ExclsvShop 0.021 WhrFrndShop 0.148 OtherPpl 0.035 MagznNews 0.539 Catlg 0.679 Internet 0.126 Excitbl 0.138 Thrfty 0.629 Rationl 0.081 Yuthfl 0.606 Orthdx 0.654 Modst 0.469 Age 0.652 ApparelEx 0.101 Income 0.684 Edu 0.984 Confusion matrix The confusion matrix compares actual segment membership (obtained from the segmen segmentation variables) and predicted segment membership (obtained from the in-samp descriptors alone). When actual and predicted segment memberships coincide, the diago large, indicating that the classification model based on available descriptors is accurate. Confusion matrix (count) The model has correctly classified 180 of the 221 observations. The off-diagonal elements are cla Predicted 1 Predicted 2 Segment 1 35 26 Segment 2 15 145 Total 50 171 Confusion matrix (%) The global hit rate of the model is 81%. The diagonal elements represent segment-specific hit rat Predicted 1 Predicted 2

Segment 1 57% 43% Segment 2 9% 91% Confusion matrix (plot). Graphic representation of the confusion matrix: actual segment member Model predictions Model predictions (in-sample) This table details the probabilities of each member of the segmentation dataset to belong to each Prob(cluster 1) Prob(cluster 2) 1 77% 23% 2 61% 39% 3 96% 4%

4 62% 38% 5 37% 63% 6 93% 7% 7 5% 95% 8 77% 23% 9 63% 37% 10 18% 82% 11 62% 38% 12 71% 29% 13 0% 100% 14 3% 98% 15 42% 58% 16 84% 16% 17 63% 37% 18 2% 98% 19 52% 48% 20 35% 65% 21 49% 51% 22 50% 50% 23 6% 94% 24 55% 45% 25 63% 37% 26 24% 76% 27 32% 68% 28 16% 84% 29 68% 32% 30 70% 30% 31 25% 75% 32 7% 93% 33 70% 30% 34 78% 22% 35 78% 22% 36 64% 36% 37 38% 62% 38 3% 97% 39 40% 60% 40 10% 90% 41 72% 28% 42 3% 97% 43 10% 90% 44 7% 93% 45 13% 87% 46 9% 91% 47 77% 23% 48 79% 21% 49 9% 91% 50 34% 66%

51 9% 91% 52 1% 99% 53 17% 83% 54 13% 87% 55 40% 60% 56 18% 82% 57 78% 22% 58 87% 13% 59 30% 70% 60 8% 92% 61 6% 94% 62 59% 41% 63 9% 91% 64 2% 98% 65 19% 81% 66 93% 7% 67 1% 99% 68 4% 96% 69 44% 56% 70 3% 97% 71 4% 96% 72 15% 85% 73 40% 60% 74 4% 96% 75 63% 37% 76 61% 39% 77 63% 37% 78 30% 70% 79 6% 94% 80 62% 38% 81 78% 22% 82 89% 11% 83 15% 85% 84 17% 83% 85 13% 87% 86 93% 7% 87 47% 53% 88 8% 92% 89 0% 100% 90 23% 77% 91 15% 85% 92 7% 93% 93 0% 100% 94 4% 96% 95 78% 22% 96 3% 97% 97 4% 96%

98 47% 53% 99 32% 68% 100 48% 52% 101 83% 17% 102 31% 69% 103 4% 96% 104 31% 69% 105 26% 74% 106 49% 51% 107 9% 91% 108 29% 71% 109 85% 15% 110 7% 93% 111 28% 72% 112 7% 93% 113 2% 98% 114 30% 70% 115 7% 93% 116 35% 65% 117 4% 96% 118 54% 46% 119 35% 65% 120 46% 54% 121 2% 98% 122 12% 88% 123 62% 38% 124 77% 23% 125 4% 96% 126 6% 94% 127 42% 58% 128 2% 98% 129 4% 96% 130 2% 98% 131 0% 100% 132 63% 37% 133 22% 78% 134 1% 99% 135 1% 99% 136 0% 100% 137 10% 90% 138 12% 88% 139 19% 81% 140 40% 60% 141 5% 95% 142 68% 32% 143 97% 3% 144 70% 30%

145 0% 100% 146 42% 58% 147 12% 88% 148 14% 86% 149 85% 15% 150 2% 98% 151 5% 95% 152 35% 65% 153 3% 97% 154 8% 92% 155 56% 44% 156 20% 80% 157 6% 94% 158 3% 97% 159 12% 88% 160 1% 99% 161 1% 99% 162 2% 98% 163 1% 99% 164 1% 99% 165 20% 80% 166 5% 95% 167 95% 5% 168 20% 80% 169 7% 93% 170 87% 13% 171 1% 99% 172 74% 26% 173 0% 100% 174 27% 73% 175 17% 83% 176 45% 55% 177 28% 72% 178 18% 82% 179 31% 69% 180 13% 87% 181 23% 77% 182 13% 87% 183 2% 98% 184 12% 88% 185 2% 98% 186 28% 72% 187 19% 81% 188 2% 98% 189 2% 98% 190 3% 97% 191 6% 94%

192 2% 98% 193 2% 98% 194 0% 100% 195 1% 99% 196 12% 88% 197 63% 37% 198 10% 90% 199 2% 98% 200 4% 96% 201 2% 98% 202 1% 99% 203 11% 89% 204 14% 86% 205 6% 94% 206 28% 72% 207 3% 97% 208 41% 59% 209 5% 95% 210 1% 99% 211 8% 92% 212 13% 87% 213 0% 100% 214 2% 98% 215 9% 91% 216 8% 92% 217 19% 81% 218 6% 94% 219 2% 98% 220 16% 84% 221 1% 99%

gers, but descriptors variables for with sufficient accuracy. The enough. (similar to the one used to le. which descriptors are the most alues will be close to

ntation analysis and the original ple classification analysis and the onal elements will be comparatively assification errors. Total 61 160 221 tes. Total

100% 100% rship versus predicted segment membership. Bubbles in the diagonale represent correct classification. h cluster (as predicted by the in-sample classification model and the descriptors alone). The segment with the highest proba Predicted Actual Correct 1 1 1 1 2 0 1 1 1

1 1 1 2 2 1 1 1 1 2 2 1 1 1 1 1 1 1 2 2 1 1 1 1 1 1 1 2 2 1 2 2 1 2 2 1 1 1 1 1 1 1 2 2 1 1 1 1 2 2 1 2 1 0 2 1 0 2 2 1 1 2 0 1 2 0 2 2 1 2 2 1 2 2 1 1 2 0 1 1 1 2 1 0 2 1 0 1 1 1 1 1 1 1 1 1 1 2 0 2 2 1 2 2 1 2 2 1 2 1 0 1 1 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 1 1 1 1 1 1 2 2 1 2 1 0

2 2 1 2 2 1 2 1 0 2 2 1 2 2 1 2 1 0 1 1 1 1 1 1 2 1 0 2 2 1 2 2 1 1 2 0 2 2 1 2 2 1 2 2 1 1 1 1 2 2 1 2 2 1 2 1 0 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 1 1 1 1 1 1 1 2 0 2 2 1 2 2 1 1 2 0 1 1 1 1 1 1 2 2 1 2 1 0 2 2 1 1 1 1 2 1 0 2 1 0 2 2 1 2 1 0 2 2 1 2 2 1 2 2 1 2 2 1 1 1 1 2 2 1 2 2 1

2 1 0 2 2 1 2 2 1 1 1 1 2 2 1 2 2 1 2 1 0 2 2 1 2 2 1 2 2 1 2 2 1 1 1 1 2 2 1 2 2 1 2 1 0 2 2 1 2 2 1 2 2 1 2 1 0 2 2 1 1 1 1 2 2 1 2 2 1 2 2 1 2 2 1 1 2 0 1 2 0 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 1 2 0 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 1 0 2 2 1 2 2 1 2 2 1 1 1 1 1 1 1 1 2 0

2 2 1 2 1 0 2 2 1 2 1 0 1 1 1 2 2 1 2 2 1 2 1 0 2 2 1 2 2 1 1 2 0 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 1 1 1 2 2 1 2 2 1 1 1 1 2 2 1 1 2 0 2 2 1 2 2 1 2 2 1 2 1 0 2 2 1 2 1 0 2 1 0 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1

2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 1 2 0 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 1 0 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1 2 2 1

ability is retained, and is compared to the actual segment membership to measure model accuracy and classification errors.

.