|
|
 |
 |
 |
 |
 |
 |
 |
| Title | Compactness of rule-based learning systems for classification |
| Speaker | Mr. Helen Dam, ITEE, UNSW@ADFA |
| Date | Wednesday 6 June |
| Time | 11:00 - 12:00 |
| Venue | Building 15 - 152 |
| Abstract |
Classification is the process of identifying common and potentially
useful patterns in a data set sufficient for discriminating among a
finite number of groups. Classification accuracy is a key factor.
Other important factors include compactness and expressiveness.
Compactness refers to the length of the learned model, while
expressiveness relates to the understandability of the knowledge
represented by the learned model. There is usually a tradeoff among
accuracy, compactness and expressiveness. In this talk, I will
present on a novel integration of genetics based machine learning
systems and artificial neural networks to balance the three criteria.
UCS is a supervised genetics based machine learning system for classification in data mining tasks. In this system, a set of rules (called the population of classifiers) is evolved overtime to represent a complete solution to the learning tasks. The representation of a rule in UCS as a univariate classification rule is straightforward for a human to understand. The main drawback of UCS is that it normally requires a large population of specific rules (or a large model) to cover the whole input space. This makes the system less efficient in terms of computational time, memory usage, and comprehensibility. This motivated us to come up with a new representation, aiming for a population with fewer classifiers while still maintaining the predictive accuracy. Artificial neural networks, inspired by the human brain, are capable to learn nonlinearity, to adapt quickly to changes in the surrounding environment, and to obtain a compact model. These properties motivated us to investigate the use of neural networks in UCS. Their disadvantage is in expressiveness due to their complicated representation. In order to use neural networks in UCS without losing too much expressiveness, we decided to only modify the action part of classifiers. In our proposed representation, the conditions are unchanged but each action is replaced by a simple neural network. The conditions are used to decompose a complex problem into a number of relatively simple tasks. Each task is then learnt by a complete but simple neural network. We found that the neural-based representation is able to achieve equivalent or even better performance than the traditional representation. Also, the new representation in general requires a significantly smaller population size than the traditional one. |
