Domain Knowledge Integration in Hierarchical Learning Model for Accurate Decision Support System Michael Siek
Business Information Systems Program, Information Systems Department, Faculty of Computing and Media, Bina Nusantara University Jakarta, Indonesia 11480
Abstract
In many applications of business, the models produced by computational intelligence algorithms can be quite complex that can capture a set of business processes. The complexity here is because the processes must be treated as multi-stationary ones. The model accuracy on predicting the low or high extremes are crucial in many business applications, like in predicting stock prices or predicting machine maintenance. In such cases, the usage of a global model for a complex process is certainly inadequate. One solution of this issue is to employ several models, each responsible on characterizing a certain sub-process. In hierarchical learning models, an upfront option to include a domain expert is to allow the experts to construct the splitting rules and perform the hard splits of the input space of the training data set. This technique enables the experts to determine split attributes and values in the upper nodes, and then the machine learning algorithm takes care of the rest of the hierarchical model construction. The domain expert is typically interested in defining the split parameters for the nodes of upper two levels of the hierarchical model, which are very essential because they affect the splitting decisions in the lower nodes and impact the performance of the overall model. In the problems of business problems, a more accurate characterization for the detailed behavior of the underlying business system can be expected by incorporating domain expert knowledge into hierarchical-based machine learning algorithm compared to the other standard machine learning models using ANNs or others. This modelling technique can be more suitable and trusted than the purely machine learning predictors or models.
Keywords: model tree algorithms, prior knowledge, multiple local models, decision support system
