Interpreting the SVM Text Data

This is an excerpt from Dr. Ham's premier book "Oracle Data Mining: Mining Gold from your Warehouse".

The SVM classificationmining activity shows two new steps labeled Text and Text (Test).  The algorithm detected the CLOB field and correctly identified it as a text attribute.  In this step the algorithm applies context indexing or feature extraction to the text attribute, and uses the same settings for the test text fields. 

Clicking the ?Result? of the ?Build? step you can see that the algorithm used a linear kernel and the words that were used to classify the type target are listed along with their coefficients.

Not surprisingly, ?GOATS? and ?GOAT? were the top attributes for the goat target class, ?SHEEP? was the highest for sheep target class, ?INFORMATION? and ?MEDICAL? were the best for biomed, and ?IUMA? was the top for bands. 

Checking the Result of the Test Metrics step shows that the model was in the ?best? range at 82% predictive confidence. 

The confusion matrixshows that the model predicted bands with 100% accuracy, and the other web pages ranged from 81% to 83% correctly predicted.

Now we?ll repeat the SVM classification and choose ?rating? as the target.  Since there are only 11 web ratings = ?medium? we have recoded these to ?cold? using the Recode Transformation.   Repeat the steps as above and keep the default settings, picking ?text? and ?rating? as the input attributes.  The preferred target value will be ?hot?. 

The result of the test metrics show that predicting the users rating of Web pages is in the ?good? range at 35%, and 41% of the ?hot? ratings were correctly predicted. 

The Build results show the words that were used in the model to classify the test cases. 

If we repeat the SVM Activity Build and include ?type? along with ?rating? and ?text? in the model, we find that the predictive confidence of the model actually decreases to 28%.