How to Interpret SVD Units in Predictive Models ?

Recent studies suggest that unstructured data such as customer comments or feedback can enhance the power of existing predictive models. SAS® Text Miner can generate SVD (Singular value decomposition) units from text documents which is a vectorial representation of terms in documents. These SVDs when used as additional inputs along with the existing structured input variables often prove to capture the response better. However, SVD units are sort of black box variables and are not easy to interpret or explain. This is a big hindrance to win over the decision makers in the organizations to incorporate these derived textual data components in the models. In this paper, we demonstrate a new and powerful feature in SAS® Text Miner 12.1 which helps in explaining the SVDs or the text cluster components. We discuss two important methods useful to interpret them. For this purpose, we used data from a television network company which has transcripts of its call center notes from 3 prior calls of each customer. We are able to extract the key terms from the call center notes in the form of Boolean rules which have contributed to the prediction of customer churn. These rules provide an intuitive sense of which set of terms when occurring in either the presence or absence of another set of terms in the call center notes may lead up to a churn. It also provides insights into which customers are at a bigger risk of churning from the company’s services and more importantly why.