This talk makes a case for developing a new methodology, called knowledge mining, which will be able to derive useful knowledge both from data and from background knowledge, and present it in human-oriented forms, such as natural language-like rules, and various forms of graphical representation. Such a methodology can potentially generate knowledge both from very large data bases and from very sparse and indirectly relevant data. The methodology is currently being implemented using an inductive database system VINLEN. The presented ideas will be illustrated by applications to medicine and to building user models from datastreams characterizing user interactions with computers.