In machine learning, we intuitively adopt an Observation-Oriented principle where observational variables act as the bedrock for relationships. It may suffice for conventional models, but with AI's capacities incorporating big data, it accentuates the misalignment between purely observational models and our actual comprehension. In contrast, humans construct cognitive entities indexed through relationships, which are not confined by observations, allowing us to formulate knowledge across temporal and hyper-dimensional spaces. This study introduces a novel Relation-Oriented perspective, drawing intuitive examples from computer vision and health informatics, to redefine our context of modeling with a causal focus. Furthermore, we present an implementation method - the relation-defined representation modeling, the feasibility of which is substantiated through comprehensive experiments.