Abstract:

Classification is one of the most practical techniques in data mining and analysis. Existing classification algorithms based on eager learning require a model assumption and do not address the correlations between individual instances, such that their performance can be affected. In this paper, we propose a new learning method based on the locally weighted regression, called MLWR. For a given test example, the MLWR firstly identifies the neighboring instances in the training set, and a locally weighted regression model is generated from the test instance and its neighboring instances.Then the test label is calculated by using the regression model and the neighboring labels. In the experiments, five classification methods are tested on 9 data sets of UCI. Experiment results show that the performance of the MLWR is superior to other methods and also suitable for big data.

Key words: classification;mapping relationship;locally weighted regression;kNN;lazy learning