Inductive learning, which tries to find rules from data, has been an important area of investigation. One major research theme of this area is the data representation language that the learning methods can use. The conventional rule learning methods use an attribute-value table as a data representation language, whereas inductive logic programming (ILP) uses the first-order logic. We propose colored directed graph as a data representation language for inductive learning methods. Graph-based induction (GBI) uses this data representation language. The expressiveness of graph stands between the attribute-value table and the first-order logic. Thus its learning potential is weaker than that of ILP, but stronger than that of the conventional attribute-value learning methods. The real advantage of GBI appears in the domain where the dependency between data bears the essential information. The behavior analysis of computer users is a typical example of such a domain. In this domain, the complex structure of dependency between the user tasks prevents us from using the conventional attribute-value learning methods, and ILP cannot meet the requirement for the efficiency. In this paper, we explain GBI method and give experimental results. We also discuss the relationship between this new method and conventional inductive inference methods such as conventional classification rule learning methods, constructive induction methods, inductive logic programming methods, macro rule learning methods, and concept learning methods. While this list of the methods covers the wide area of inductive inference, we find that most of them can use Stepwise Pair Expansion as their basic algorithm. The use of the pairs and the representation language define the function and characteristics of each method. We also discuss the use of the statistical measures such as gini index and information gain index to realize various inductive inference functions.