Modeling behavior data is important for online platforms and their users. Machine learning on the behavior data aims at finding effective representations (for calculation) to support functions such as recommendation, fraud detection, planning, and decision making. The challenge of learning to model behaviors lies in the complex and evolving nature of the behavior contexts. We are witnessing interactions more than isolation and dynamics more than static state. Existing methods were designed to preserve pair-wise similarity of the components that interact with each other into their representations. For example, they assumed that in a paper co-authors should be similar, in a project co-investigators should be similar, in a course learning materials should be similar, and in a prescription drugs should be similar. However, decision makers such as project leaders, course instructors, and doctors are often looking for complementary (instead of the most similar) resources and ideas to create an effective solution. How to quantify and capture the complementarity remains an open problem. Moreover, real-world behaviors are constantly changing over time. So, the dynamics that drive the evolutionary process of the behaviors need to be properly modeled. In this dissertation I describe several studies in which I present computational approaches for modeling and learning complementarity and dynamics.