A new machine learning algorithm: fixed partition averaging

We propose the Fixed Partition Averaging (FPA) method for reducing the storage requirement and classification time of memory based reasoning (MBR). This method extracts representative patterns from the training set using an instance averaging technique. First, the proposed algorithm partitions the pattern space into a fixed number of hyperrectangles, and then it averages patterns in each hyperrectangle to extract a representative. This algorithm then uses the mutual information between the features and class information as its weights to improve the classification accuracy. We present the FPA algorithm and verify its performance. We compare its classification accuracy, storage requirement and actual classification time with k-NN and the EACH system on 7 carefully chosen data sets from the UCI Machine Learning Database repository.