Knowledge acquisition of multi-granularity ordered information systems

Knowledge reduction is an essential issue in data mining. However, most of the existing studies consider the problem of knowledge reduction in a single-scale system. When the system expands to a multi-scale information system (MIS), the existing methods are not available. To overcome this challenge, we design a universal framework to reduce attributes and select optimal scales (OSs) based on the dominance relation in ordered information systems (OISs). Firstly, we analyze the effect of scale on rough approximations and correlation measures in multi-scale ordered information systems (MOISs). Secondly, the relation between attribute reduction and OS selection is explored in OISs. Moreover, the corresponding algorithms are designed around both tasks. Due to our universal framework considering the decisions at the finest, optimal and coarsest scale, our study is an extension of the three-way decision. Thirdly, a novel MOIS construction algorithm is explored to transform an OIS into an MIS. Thus, the proposed reduction algorithms can be operated in any OIS. Experiments substantiate that models trained by data at the OS can improve their stability and the attribute reduction task is also suitable for the proposed algorithm.