Knowledge Discovery: Comprehensibility of the Results

Knowledge Discovery in Databases (KDD) and Machine Learning researchers recognize that comprehensibility is an important condition for the use, and therefore usefulness, of knowledge discovery meethods. An investigation of the comprehensibility of the discovered patterns may benefit IS research as well. This paper attempts to identify some of the issues that such inquiry may face. Related findings and conclusions of four case studies focusing on applications of decision tree induction methods, are described. A discussion based on these studies suggests, among the rest, that: (1) The problem of comprehensibility is complicated by a complex context due to a diversity of problem-domain attributes, user and task characteristics, algorithmic methods, and concurrency of user goals. (2) Solutions to problems of discovered patterns that are not easy to interpret and validate may involve integration of available information technologies, and utilization of multiple information types and sources. (3) An investigation of comprehensibility issues may benefit from the adoption of multiple definitions in relation to this concept.