Automated Construction of a Hierarchy of Self-Organized Neural Network Classifiers
MetadataShow full item record
This paper documents an effort to design and implement a neural network-based, automatic classification system which dynamically constructs and trains a decision tree. The system is a combination of neural network and decision tree technology. The decision tree is constructed to partition a large classification problem into smaller problems. The neural network modules then solve these smaller problems. We used a variant of the Fuzzy ARTMAP neural network which can be trained much more quickly than traditional neural networks. The research extends the concept of self-organization from within the neural network to the overall structure of the dynamically constructed decision hierarchy. The primary advantage is avoidance of manual tedium and subjective bias in constructing decision hierarchies. Additionally, removing the need for manual construction of the hierarchy opens up a large class of potential classification applications. When tested on data from real-world images, the automatically generated hierarchies performed slightly better than an intuitive (handbuilt) hierarchy. Because the neural networks at the nodes of the decision hierarchy are solving smaller problems, generalization performance can really be improved if the number of features used to solve these problems is reduced. Algorithms for automatically selecting which features to use for each individual classification module were also implemented. We were able to achieve the same level of performance as in previous manual efforts, but in an efficient, automatic manner. The technology developed has great potential in a number of commercial areas, including data mining, pattern recognition, and intelligent interfaces for personal computer applications. Sample applications include: fraud detection, bankruptcy prediction, data mining agent, scalable object recognition system, email agent, resource librarian agent, and a decision aid agent.