Measuring the consistency with prior knowledge of classification models

Growing popularity of the Internet and innovative storage technology have caused a true data explosion. The overall process of extracting knowledge from this vast amount of data is called data mining. Classification is a subtask of data mining and involves the assigning a data point to a predefined class or group according to its predictive characteristics. The classification problem and accompanying data mining techniques are relevant in a wide variety of domains such as financial engineering, medical diagnostic and marketing. The performance of a classification model is typically measured by its accuracy; however justifiability is also a major requirement in many data mining applications. Justifiability concerns the extent to which the model is in line with prior domain knowledge. The best known justifiability requirement is the monotonicity constraint (e.g. increasing income should yield an increasing probability of being granted a loan). Several adaptations to existing classification techniques have been proposed to cope with justifiability, yet a measure to identify the extent to which the models conform to the requirements is still lacking. A new measurement will be proposed, where with the use of decision tables we provide a crisp performance measure for the critical justifiability measure.