Discovery Planning: Multistrategy Learning in Data Mining
Kenneth A. Kaufman, Ryszard S. Michalski · 1998
The process of applying machine learning to data mining may require many trials, backtracking, and multiple executions of different learning and inference procedures. Such a process can be time-consuming, laborious, and prone to errors. To overcome these problems, this paper proposes an integration of diverse learning and inference procedures into a system that can automatically pursue different data mining tasks according to a high-level plan developed by a user. The solution involves a meta-language, called KGL (Knowledge Generation Language), for specifying a data exploration process in terms of high-level operators and conditional statements that depend on the results of previous operators. Operators invoke corresponding machine learning and inference programs, and specify their parameters according to the current tasks and previous results. To assist in illustrating the outcomes of exploration they are organized into association graphs (AGs), which can indicate logical, statistical and equational relationships in the data. The methodology is exemplified by preliminary results in the areas of demographics and medicine.