SVM Classification Activity

After importing the data using the ODMrImport Wizard, a Classification Activity using SVM is initiated, setting CHAS as the target attribute.  Note that there are two classes: 1 if the tract bounds the Charles River and 0 if not. 

Of the total 506 properties in the dataset, only 7% are next to the Charles River.  The unique identifier in the case table is OBS after changing the “OBS.” in the original dataset to “OBS” in order to eliminate problems with having a “.” in the column headings.  The preferred target value is 1, and we’ll keep all the default advanced settings.

The results of the SVM classification activity show that the model predictive accuracy  is in the best range of 69%, with 80% of the preferred target class 1 correctly classified, and 89% of the class 0 correct. 

Note that the SVM algorithm chose the Gaussian kernel function, so there are no rules to examine.  We’ll re-build the model and pick the Linear kernel, and compare the two results. 

Now, we click on “Activity”, and build another SVM classificationmodel as above.  This time, however, after completing the New Activity Wizard, choose Advanced Settings, and under the “Build” tab and select “Algorithm” settings and pick “linear” as the kernel function, keeping the default settings for tolerance, complexity factors, and Active Learning.  Finally, click “OK” and then “Finish”, completing the Build Activity. 

Interpreting the SVM Results

Examination of the Test metrics result shows that the Predictive Confidenceis good at 57%, slightly less than the model built using the Gaussian kernel.  Click on Build Result to see the coefficients and values of the attributes used to build the model.  You can see that NOX, a measure of air pollution, was the topmost attribute, and the towns of Dedham, Waltham, Dover, Watertown, Newton, Wellesley and Boston following next.  

The positive values of the coefficientsmean that these towns are highly likely have properties bordering the Charles River, whereas towns like Brookline and Belmont with coefficients of -1 are very unlikely to be near the Charles. 

A Google Map of the Dedham area shows that indeed there are many residential areas bordering the Charles River. 

Refining the SVM Model

But wait a minute!  What if you were searching the area for housing for yourself or a client?  You are concerned about the NOX having the highest coefficient of 3.97.  This model does not give any indication of whether the air pollution index is higher or lower for properties around the Charles River, only that it is an important factor.  If you look at the Boston case dataset by right-clicking the table name and choosing “Show Summary Single Record”, you’ll find that NOX ranges from 0.38 to 0.87 with mean of 0.55 and variance of 0.01. 

NOXis a continuous as opposed to categorical variable, meaning that there are an infinite number of possible values between the minimum and maximum.   ODMrthe continuous variables as FLOAT data type, which is seen when you click on data summary on Step 3 of the Activity Wizard.  Discrete or categorical variables can possess only exact values, and intermediate values are not possible.  To examine the effect of NOX on our Charles River target attribute, we have a couple of options.  One is to discretize NOX into High, Low, and Medium values.  ODMr has a discretize transformation that we’ll explore in Chapter 5.  For now, let’s look at the regression capabilities of the SVM algorithmin modeling continuous variables.