Descriptor-Based Information Systems and Rule Learning from Different Types of Data Sets with Uncertainty

Abstract

We have coped with rule generation from tables (or information systems) and extend this framework to that from different types of data, like heterogeneous, uncertain, time series, clustered, etc. For this new framework, we use the term ‘rule learning’ instead of ‘rule generation.’ We newly specify descriptors, which take the role of words in data sets, and we define a DbIS (Descriptor-based Information System) $\varOmega$ as the unified format for different data types. We prove one DbIS $\varOmega$ is convertible to one NIS (Non-deterministic Information System), where we realized the NIS-Apriori system. Thus, we can obtain rules via DbIS and the NIS-Apriori system. Furthermore, we merge two DbISs $\varOmega _{1}$ and $\varOmega _{2}$ to one $\varOmega _{1,2}$ using missing values, and we can inductively merge any number of DbISs. We also estimate some missing values by the self-obtained certain rules and consider the application of DbIS to big data analysis.