Learning Qualitative Models through Partial Derivatives by Padé

Padé is a new method for learning qualitative models from observation data by computing partial derivatives from the data. Padé estimates partial derivatives of a target function from the learning data by splitting the attribute space into triangles or stars from Delaunay triangulation, or into tubes, and computing the linear interpolation or regression within these regions. Generalization is then accomplished by any attributevalue learning method. The methods for estimating partial derivatives differ regarding their resistance to noise, ability to handle noisy and missing values, computation speed and other properties. The experiments show these methods to be quite accurate, fast and robust. Being well integrated into our general machine learning and data mining suite Orange, Padé should also prove useful in practice.