Applying Supervised Learning to Real-World Problems

The last years have seen machine learning methods applied to an increasing variety of application problems such as: language, handwriting and speech processing , document classification, knowledge discovery in databases, industrial process control and diagnosis, fraud and intrusion detection, image analysis and many others. Our work starts from the realization that most of these problems require significant reformulation before learning algorithms can be applied, and in many cases, existing algorithms require modifications before being applied to a problem. The problems mentioned above differ in many aspects but, if subdivided into smaller problems (and this is the approach commonly taken), the sub-problems can often be formulated and approached by employing similar , unified supervised learning techniques. However, this divide-and-conquer process creates dependencies in the data that violate the assumption that the data are independent and identically distributed (iid) and that all errors are of equal cost, issues that involve making changes to existing learning algorithms. The purpose of my thesis is to identify some procedural steps that are shared among the learning approaches to complex real-world problems, and to develop robust general purpose techniques and tools to replace ad-hoc decisions that are currently made during an application effort. The main topics my thesis will emphasize on are described as follows. Learning with misclassification costs. Most classification algorithms assume uniform class distribution and try to minimize the misclassification error. However, many applications require classifiers that minimize an asymmetric loss function rather than the raw misclassification rate. In general, the cost of a wrong prediction depends both on the actual class and on the predicted class. One way to incorporate loss information into classifiers is to alter the priors based on labels of the training examples. For 2-class problems, this can be easily accomplished for any loss matrix, but for k > 2 classes, it is not sufficient. We have studied methods for setting the priors to best approximate an arbitrary k x k loss matrix in decision tree learners. We are currently studying alternative methods to incorporate loss information into decision trees, neural networks and other supervised learning algorithms. Incorporating prior and common sense knowledge and learning with non-independent data examples. Generally, besides the data available, there is simple, common sense knowledge about the data that is not made explicit in the dataset or database. One question is whether this knowledge is useful, and, if so, how it can be represented and …