1. Chapter 4 –Dimension Reduction Data Mining for Business Analytics Shmueli, Patel & Bruce

2. Dimensionality It is the number of variables in the data set It can grow due to derived variables Too many variables can lead to overfitting Accuracy and reliability can suffer Can lead to multicollinearity problem Can pose computational problem When the model is deployed can lead to unnecessary costs of collecting and processing data

3. Curse of dimensionality Data required to densely populate space increases exponentially as the dimension increases In 100- dimensional space, they would be like distant galaxies Eight points fill the one- dimensional space Become more separated as the dimension increases

4. Practical Considerations Which variables are most important for the task? Which variables are likely to be useless? Which variables are likely to contain much error? Which variables can be measured in the future? Which variables can be measured before the outcome?

5. Exploring the data Statistical summary of data: common metrics Average Median Minimum Maximum Standard deviation Counts & percentages

6. Summary Statistics – Boston Housing

7. Summarize Using Pivot Tables Counts & percentages are useful for summarizing categorical data Boston Housing example: 471 neighborhoods border the Charles River (1) 35 neighborhoods do not (0)

8. Pivot Tables - cont. Boston Housing example: Compare average home values in neighborhoods that border Charles River (1) and those that do not (0) Averages are useful for summarizing grouped numerical data

9. Pivot Tables, cont. Group by multiple criteria: By # rooms and location E.g., neighborhoods on the Charles with 6-7 rooms have average house value of 25.92 ($000)

10. Pivot Table - Hint To get counts, drag any variable (e.g. “ID”) to the data area Select “settings” then change “sum” to “count”

11. Correlations Between Pairs of Variables: Correlation Matrix from Excel CRIMZNINDUSCHASNOXRMAGEDISRADTAXPTRATIOBLSTATMEDV CRIM1.00 ZN-0.201.00 INDUS0.41-0.531.00 CHAS-0.06-0.040.061.00 NOX0.42-0.520.760.091.00 RM-0.220.31-0.390.09-0.301.00 AGE0.35-0.570.640.090.73-0.241.00 DIS-0.380.66-0.71-0.10-0.770.21-0.751.00 RAD0.63-0.310.60-0.010.61-0.210.46-0.491.00 TAX0.58-0.310.72-0.040.67-0.290.51-0.530.911.00 PTRATIO0.29-0.390.38-0.120.19-0.360.26-0.230.46 1.00 B-0.390.18-0.360.05-0.380.13-0.270.29-0.44 -0.181.00 LSTAT0.46-0.410.60-0.050.59-0.610.60-0.500.490.540.37-0.371.00 MEDV-0.390.36-0.480.18-0.430.70-0.380.25-0.38-0.47-0.510.33-0.741.00

12. Reducing Categories A single categorical variable with m categories is typically transformed into m-1 dummy variables Each dummy variable takes the values 0 or 1 0 = “no” for the category 1 = “yes” Problem: Can end up with too many variables Solution: Reduce by combining categories that are close to each other Use pivot tables to assess outcome variable sensitivity to the dummies Exception: Naïve Bayes can handle categorical variables without transforming them into dummies

13. Combining Categories Many zoning categories are the same or similar with respect to CATMEDV ZN = 12.5, 18, 21, 25, 28, 30, 70 and 85 all have CAT.MEDV = 0 ZN = 17.5, 90, 95 and 100 all have CAT.MEDV = 1

14. Principal Components Analysis Goal: Reduce a set of numerical variables. The idea: Remove the overlap of information between these variable. [“Information” is measured by the sum of the variances of the variables.] Final product: A smaller number of numerical variables that contain most of the information

15. Principal Components Analysis How does PCA do this? Create new variables that are linear combinations of the original variables (i.e., they are weighted averages of the original variables). These linear combinations are uncorrelated (no information overlap), and only a few of them contain most of the original information. The new variables are called principal components.

16. Example – Breakfast Cereals

17. Description of Variables Name: name of cereal mfr: manufacturer type: cold or hot calories: calories per serving protein: grams fat: grams sodium: mg. fiber: grams carbo: grams complex carbohydrates sugars: grams potass: mg. vitamins: % FDA rec shelf: display shelf weight: oz. 1 serving cups: in one serving rating: consumer reports

18. Consider calories & ratings Total variance (=“information”) is sum of individual variances: 379.63 + 197.32 = 576.95 Calories accounts for 379.63/576.95 = 66% Rating accounts for 197.32/576.95 = 34% CaloriesRating Calories379.63-189.68 Rating-189.68197.32

19. First & Second Principal Components Z 1 and Z 2 are two linear combinations. Z 1 has the highest variation (spread of values) Z 2 has the lowest variation

20. PCA output for these 2 variables Top: weights to project original data onto Z 1 & Z 2 e.g. (-0.847, 0.532) are weights for Z 1 Bottom: reallocated variance for new variables Z 1 : 86% of total variance Z 2 : 14%

21. Computing the PC scores Let X 1 = Calories, X 2 = Rating Z 1 = -0.847(X 1 -  1 ) + 0.532(X 2 -  2 ) Z 2 = 0.532(X 1 -  1 ) + 0.847(X 2 -  2 )

22. Computing the PC scores Let X 1 = Calories, X 2 = Rating Z 1 = -0.847(X 1 -  1 ) + 0.532(X 2 -  2 ) Z 2 = 0.532(X 1 -  1 ) + 0.847(X 2 -  2 )  1 = 106.88,  2 = 42.67 NameCaloriesRatingZ1Z1 Z2Z2 100%_Bran7068.40-0.847(70 - 106.88) + 0.532(68.4 – 42.67) = 44.92 0.532(70 - 106.88) + 0.847(68.4 – 42.67) = 2.20 100%_Natural_Bran12033.98-0.847(120 - 106.88) + 0.532(33.98 – 42.67) = - 15.73 0.532(120 - 106.88) + 0.847(33.98 – 42.67) = 2.20

23. Principal Component Scores Weights are used to compute the above scores e.g., col. 1 scores are computed Z 1 scores using weights (-0.847, 0.532)

24. Properties of the resulting variables New distribution of information: New variances = 498 (for Z 1 ) and 79 (for Z 2 ) Sum of variances = sum of variances for original variables calories and ratings New variable Z 1 has most of the total variance, might be used as proxy for both calories and ratings Z 1 and Z 2 have correlation of zero (no information overlap)

25. Generalization X 1, X 2, X 3, … X p, original p variables Z 1, Z 2, Z 3, … Z p, weighted averages of original variables All pairs of Z variables have 0 correlation Order Z’s by variance (z 1 largest, Z p smallest) Usually the first few Z variables contain most of the information, and so the rest can be dropped.

26. PCA on full data set First 6 components shown First 2 capture 93% of the total variation Note: data differ slightly from text

27. Normalizing data In these results, sodium dominates first PC Just because of the way it is measured (mg), its scale is greater than almost all other variables Hence its variance will be a dominant component of the total variance Normalize each variable to remove scale effect Divide by std. deviation (may subtract mean first) Normalization (= standardization) is usually performed in PCA; otherwise measurement units affect results Note: In XLMiner, use correlation matrix option to use normalized variables

28. PCA using standardized variables First component accounts for smaller part of variance Need to use more components to capture same amount of information

29. PCA in Classification/Prediction Apply PCA to training data Decide how many PC’s to use Use variable weights in those PC’s with validation/new data This creates a new reduced set of predictors in validation/new data

30. Regression-Based Dimension Reduction Multiple Linear Regression or Logistic Regression Use subset selection Algorithm chooses a subset of variables This procedure is integrated directly into the predictive task

31. Summary Data summarization is an important for data exploration Data summaries include numerical metrics (average, median, etc.) and graphical summaries Data reduction is useful for compressing the information in the data into a smaller subset Categorical variables can be reduced by combining similar categories Principal components analysis transforms an original set of numerical data into a smaller set of weighted averages of the original data that contain most of the original information in less variables.