Identifying user-dependent information that can be automatically collected helps build a user model by which 1) to predict what the user wants to do next and 2) to do relevant preprocessing. Such information is often relational and is best represented by a set of directed graphs. A machine learning technique called graph-based induction (GBI) e ciently extracts regularities from such data, based on which a user-adaptive interface is built that can predict next command, generate scripts and prefetch les in a multi task environment. The heart of GBI is pairwise chunking. The paper shows how this simple mechanism applies to the top down induction of decision trees for nested attribute representation as well as nding frequently occurring patterns in a graph. The results clearly shows that the dependency analysis of computational processes activated by the user commands which is made possible by GBI is indeed useful to build a behavior model and increase prediction accuracy.