A dataset contains so many fields in which there is both relevant and irrelevant field. If we want to do a market campaigning using propensity scoring, which fields of the data set are relevant? How can we find which data field should be selected and can drive to the desired propensity score?
The problem you are examining is called feature selection. There are many different techniques, but they fall broadly into three categories:
- Filter approaches determine which features have high information content. A common approach is to score them based on their information gain.
- Wrapper approaches score subsets of the features, using similar measurements. This is slower (since there are many subsets), but may yield better performance.
- Embedded approaches use machine learning algorithms that can dynamically select features. For example, building a C4.5 decision tree learner on a dataset, and then selecting just the features that actually appear in the tree, yields a set of features similar to those found by information gain.