Abstract
Inexact fielding learning (IFL) (Ciesieski & Dai, 1994; Dai & Ciesieski, 1994a, 1994b, 1995, 2004; Dai & Li, 2001) is a rough-set, theory-based (Pawlak, 1982) machine learning approach that derives inexact rules from fields of each attribute. In contrast to a point-learning algorithm (Quinlan, 1986, 1993), which derives rules by examining individual values of each attribute, a field learning approach (Dai, 1996) derives rules by examining the fields of each attribute. In contrast to exact rule, an inexact rule is a rule with uncertainty. The advantage of the IFL method is the capability to discover high-quality rules from low-quality data, its property of low-quality data tolerant (Dai & Ciesieski, 1994a, 2004), high efficiency in discovery, and high accuracy of the discovered rules.
TopBackground
Achieving high prediction accuracy rates is crucial for all learning algorithms, particularly in real applications. In the area of machine learning, a well-recognized problem is that the derived rules can fit the training data very well, but they fail to achieve a high accuracy rate on new unseen cases. This is particularly true when the learning is performed on low-quality databases. Such a problem is referred as the Low Prediction Accuracy (LPA) problem (Dai & Ciesieski, 1994b, 2004; Dai & Li, 2001), which could be caused by several factors. In particular, overfitting low-quality data and being misled by them seem to be the significant problems that can hamper a learning algorithm from achieving high accuracy. Traditional learning methods derive rules by examining individual values of instances (Quinlan, 1986, 1993). To generate classification rules, these methods always try to find cut-off points, such as in well-known decision tree algorithms (Quinlan, 1986, 1993).
What we present here is an approach to derive rough classification rules from large low-quality numerical databases that appear to be able to overcome these two problems. The algorithm works on the fields of continuous numeric variables; that is, the intervals of possible values of each attribute in the training set, rather than on individual point values. The discovered rule is in a form called β−rule and is somewhat analogous to a decision tree found by an induction algorithm. The algorithm is linear in both the number of attributes and the number of instances (Dai & Ciesieski, 1994a, 2004).