Contrastive learning has emerged as a dominant paradigm for representation learning, while node homophily–the tendency for similar nodes to be connected–represents a fundamental principle in graph-based learning. This paper establishes a formal connection between these two concepts, revealing that contrastive learning directly optimizes homophilic graph structures through representation learning. Although intuitively related, these frameworks operate in fundamentally different domains: contrastive learning optimizes continuous vector representations, while node homophily is defined on discrete graph structures. To bridge this gap, we introduce a probabilistic framework based on influence matrices that translates between discrete graph adjacencies and continuous similarity relationships. This framework enables us to prove analytically that minimizing contrastive loss is mathematically equivalent to maximizing graph homophily, specifically: \(\mathcal {L}_{\text {contrastive}}^{(i)} = -\log L_i\) . Our theoretical analysis provides several key insights: (a) standard generalization practices naturally ensure that learned homophily reflects meaningful rather than spurious similarities, (b) an imperfectness parameter \(\varepsilon \) characterizes how real-world contrastive learning captures both explicit relationships and latent homophily potential, and (c) node selection strategies provide flexibility in applying the framework while maintaining theoretical guarantees. We address current limitations and outline promising directions for future empirical validation across diverse domains. This work establishes rigorous theoretical foundations connecting contrastive learning and graph homophily optimization, opening new directions for adaptive graph construction in graph neural networks.