Relation-Oriented: Toward Causal Knowledge-Aligned AI

This study examines the inherent limitations of the prevailing Observation-Oriented learning paradigm by understanding relationship modeling from a unique dimensionality perspective. This paradigm necessitates the identification of modeling objects prior to defining relations, confining models to observational space, and limiting their access to dynamical temporal features. By relying on a singular, absolute timeline, it often neglects the multi-dimensional nature of the temporal feature space, compromising the robustness and generalizability of structural causal models and contributing significantly to the AI misalignment issue. Drawing from the relation-centric essence of human cognition, this study presents a new Relation-Oriented paradigm. Supported by extensive efficacy experiments, this paradigm, and its methodological counterpart, relation-defined representation learning, aim to construct interpretable AI grounded in established knowledge.