<p>Graph contrastive learning has recently been introduced into recommendation algorithms to address the issue of data sparsity by user–item representations learned from limited interaction signals. However, most existing approaches rely on heuristic graph augmentations that neglect the spectral characteristics of interaction graphs, which may distort low-frequency collaborative structures while failing to effectively model high-frequency variations. In this paper, we propose SSACCF, a Simplified spectral-aware contrastive learning framework for graph-based collaborative filtering. SSACCF introduces a learnable spectral perturbation mechanism that preserves low-frequency components representing stable and globally shared user–item interaction behaviors, while selectively perturbing high-frequency components associated with local variations and noise to construct informative contrastive views. To effectively exploit the resulting spectral variations, we further design a spectral-aware graph convolutional encoder with non-linear transformations, enabling adaptive modeling of frequency-dependent collaborative information without incurring costly spectral decomposition. The proposed framework is trained by jointly optimizing a Bayesian Personalized Ranking objective and a spectral contrastive learning, ensuring both recommendation accuracy and robustness to structural perturbations. Extensive experiments on multiple benchmark datasets demonstrate that SSACCF consistently outperforms state-of-the-art graph neural network and graph contrastive learning-based recommendation methods, highlighting the effectiveness of incorporating spectral structure into contrastive learning for recommendation tasks. Our experimental results are highly competitive, with performance improvements ranging from 6% to 10% across various evaluation metrics.</p>