Learning on graphs that exhibit both homophilic and heterophilic structures remains a fundamental challenge in graph representation learning, particularly for critical applications such as fraud detection. Existing studies typically model the underlying graph as either homophilic or heterophilic; however, real-world graphs often display varying degrees of homophily across different subgraphs. To address this limitation, we propose a novel model, the MultiScale Spectral Graph Neural Network (MSSGNN), which tackles this challenge by integrating multi-level spectral filtering with relation-aware subgraph decomposition. Our approach introduces a hierarchical spectral filtering framework employing Beta wavelets at multiple scales, enabling the model to effectively capture diverse heterophily patterns. Node clusters are dynamically extracted based on local edge homophily scores, computed by a lightweight Relation-Aware module. Customized wavelet filters with adaptive propagation depths are applied to each subgraph, and the outputs are fused through a learnable attention mechanism to adaptively integrate multi-level heterophily signals. Experimental results on benchmark fraud detection datasets demonstrate that MSSGNN outperforms existing methods, validating the effectiveness of hierarchical spectral processing and relation-aware subgraph modeling. This work provides a flexible and principled approach for learning robust node representations in highly irregular and adversarial graph environments.