Influence ranking in networks is a fundamental problem with applications in social network analysis, financial risk assessment, and information diffusion. Traditional methods, such as centrality-based algorithms and meta-path-based approaches, either assume homogeneous network structures or require extensive domain-specific knowledge, limiting their adaptability to real-world networks. Recent advances in Graph Neural Networks (GNNs) offer a promising alternative, but their reliance on labeled data restricts their generalizability and scalability. To address these challenges, we propose AdaRank, an unsupervised influence ranking framework. For homogeneous networks, AdaRank employs Laplacian-normalized adjacency coupled with node-specific PPR propagation, where each node automatically learns its optimal balance between retaining intrinsic influence and absorbing neighborhood information. In heterogeneous networks, it additionally incorporates automated multi-hop relationship extraction, eliminating meta-path dependency. The framework trains entirely through unsupervised objectives, requiring no labeled ranking data. Comprehensive evaluations on five real-world networks demonstrate more than 11% improvements over state-of-the-art baselines. Our findings highlight the effectiveness of unsupervised diffusion learning for influence ranking, demonstrating AdaRank’s ability to rank nodes influence. By combining structural modeling with unsupervised influence propagation, AdaRank provides a scalable, adaptive, and interpretable solution for ranking influential nodes in complex networks.