Regression on feature projections

Abstract

This paper describes a machine learning method, called Regression on Feature Projections (RFP), for predicting a real-valued target feature, given the values of multiple predictive features. In RFP training is based on simply storing the projections of the training instances on each feature separately. Prediction of the target value for a query point is obtained through two averaging procedures executed sequentially. The first averaging process is to find the individual predictions of features by using the K-Nearest Neighbor (KNN) algorithm. The second averaging process combines the predictions of all features. During the first averaging step, each feature is associated with a weight in order to determine the prediction ability of the feature at the local query point. The weights, found for each local query point, are used in the second prediction step and enforce the method to have an adaptive or context-sensitive nature. We have compared RFP with KNN and the rule based-regression algorithms. Results on real data sets show that RFP achieves better or comparable accuracy and is faster than both KNN and Rule-based regression algorithms.