Inductive inference in an intelligent database system

Enhancing a database management system with inductive inferential capabilities makes it possible to extract new knowledge from a large amount of data stored in a database. There are special requirements for conceptual induction in a database environment: (1) the large amount of data and incremental insertion of new data require an efficient incremental induction method; (2) descriptions of complex entities and utilization of knowledge require an expressive representation language; and (3) the large variety of application domains requires a general method of induction. The main results of this dissertation are a conceptual inductive model called the Knowledge-Directed Model (KNOWD model) and a general methodology (called the KNOWD methodology) for conceptual induction in a database environment. There are three major innovations in this research work: First, the model supports an expressive representation language which allows efficient incremental processing. The model represents a concept description by a set of logical expressions, and this representation makes it possible to fulfill the above requirements. Second, the methodology has a unique control strategy which combines basic abstraction and advanced abstraction so that the conceptual induction can be either supervised or unsupervised and can address a broad range of applications. Third, as a general inductive method, the KNOWD methodology explicitly contains a bias management system. This provides comprehensive management of a variety of inductive biases including general heuristics and domain specific knowledge, so that this methodology can be applied efficiently to a wide spectrum of situations. The KNOWD model employs domain knowledge for directed searching for relationships among concept features instead of assuming that all features are independent or are related. Both the KNOWD model and the methodology have been implemented, providing an inductive database system (IDBl). The inductive database system has been tested on several applications, and the results demonstrate the high representation capability, inductive quality and efficiency of the methodology.