Building the SVM Activity

Now we’ll execute the Build Activity and choose the Support Vector Machinealgorithm under the Classification Function.  Select “season” as the target class, and pick your favorite season as the preferred target value.  All steps in the Activity Wizard are the same as the Naïve Bayes, Adaptive Bayes Network, and Decision Tree until we come to the Advanced Settings Dialog.  In the SVM algorithm, we have new tabs for Outlier Treatment, Missing Values, and Normalize. 

SVM may be adversely affected by extreme or “outlier” values in the case data table, so we need to get rid of them, and ODMrgives you options of how to handle these by specifying the number of standard deviations, the percent of upper and lower tailing values in the distribution, or by typing in an actual value for the cutoff point.   The “Replace with” option gives you the choice of either replacing or discarding the extreme values.  The default is to use standard deviation.

Missing Values in SVM Analysis

Missing values must also be addressed, and under the Missing Values tab you can replace numeric types of data with the mean, minimum value, maximum value, a custom value that you type in, or simply drop the attribute if the column is null.  For categorical data you can replace the value with the mode, which is the most frequently occurring value, or a custom value that you type in.  The default is to replace missing values with the mean if the attribute is numerical or mode for categorical fields.

Sparse Data in SVM Analysis

What is the difference between missing values and sparse data?  In the Irish wind data, there is data for every row and every attribute; the data is neither missing nor sparse.  But let’s say that you are analyzing patients who are hospitalized for reasons that may have been preventable, such as a hospital admission for complications arising from a chronic disease such as diabetes.  Fortunately such admissions are rare compared with most, so the target data is considered sparse in relation to the entire universe of hospitalizations.  Normally you won’t impose a missing value on sparse data, but you can if you want to by un-checking the box at the top of the Missing Values screen.

Normalization of SVM Data

SVM requires that all numerical data is normalized, which further reduces the variability in the raw data.  Min/Max is the default method for normalization, where all values are re-coded in the range of 0 to 1.  Z-score is a good choice for normalization if you have chosen to keep outliers in your dataset.  The default strategy is to use the min/max scheme.

Linear and Gaussian Normalization

In the Build options, the kernel used in the algorithm can be determined by ODMr, or you can specify linear or Gaussian.  If linear is used for the kernel, the coefficients for each attribute used to build the model will be rank-ordered and you can see which ones contribute the most in determining the target class.  Tolerance valuetells the algorithm to stop building the model; increasing this value to a higher number will build the model faster but may be less accurate.