NEW CLASSIFICATION METHODS FOR GATHERING PATTERNS IN THE CONTEXT OF GENETIC PROGRAMMING
Alma Lilia, Garcia Almanza · 2008
Machine learning techniques extend the past experiences into the future. However, when the number of examples in the minority class (positive cases) is very small in comparison with the remaining classes, it poses a serious challenge to the machine learning [63],[119],[5],[81]. In this kind of problems, the prediction of the majority class is favoured because it has a high chance of being correct. This characteristic is present in many real-world problems, whose objective is to classify the minority class in imbalanced data sets. However, a prediction that detects more positive cases may be paid for with more false alarms. It is important to determine a balance between the detection of positive cases and false alarms. A range of classifications would give users the option to choose the best tradeoff between detecting positive cases and false alarms according to their requirements. On the other hand, we consider it is important to provide a comprehensive solution, which shows the real variables and conditions in the prediction. Thus, the users could combine their knowledge in order to make a more informed decision. In this thesis, we present three novel approaches: Repository Method (RM), Evolving Decision Rules (EDR) and Scenario Method (SM). We use Genetic Programming (GP) and supervised learning to build the methods proposed in this thesis. The main objectives of RM and EDR are: to predict the minority class in imbalanced environments, to generate a range of solutions to suit different users’ preferences and to provide an comprehensible solution for the user. On the other hand, SM has been designed to improve the precision and accuracy of the prediction. However, such improvement is paid for with a decrease in the recall. But, the users have to make the decision of which of these parameters is more adequate to satisfy their needs. This work is illustrated predicting future opportunities in financial stock markets. Experiments of our methods were carried out, and these showed promising results for achieving our goals. RM and EDR were compared to a standard Genetic Programming, EDDIE-Arb and C5.0. The methods presented in this thesis can also be used in other fields of knowledge, these should not be limited to financial forecasting problems.