Learning from Missing Data: A Reflex Fuzzy Min-Max Neural Network Approach

摘要

A Reflex Fuzzy Min-Max Neural Network (RFMN) capable of learning from missing data is presented. Many real world problems involve machine leaning with missing values or attributes. Thus, learning with missing or incomplete data is an important issue for the practical implementation of pattern recognition systems. The conventional approach to this problem is to impute missing values by substituting their estimates. The proposed RFMN uses an evidence based learning approach. It extracts the underlying structure in the missing data from the available features or attributes only. The learned classes are represented by an aggregation of hyperbox fuzzy sets. A novel concept of reflex mechanism, inspired from human brain is used to handle class overlaps. The main advantage of RFMN is that it learns the data on-line and in a single pass though. Experimental results on real datasets show a better performance of RFMN.