<p>Predicting temporal links in dynamic graphs, especially in complex social networks, is a fundamental challenge in social computing, complicated by the non-stationary and multi-scale nature of interactions. Although recent frequency-domain approaches have demonstrated the potential of spectral encoding, they predominantly rely on the Fourier transform. A critical limitation of this paradigm is its inability to localize patterns simultaneously in the time and frequency domains, often failing to capture transient fluctuations in evolving networks. Furthermore, static frequency encoding often neglects the temporal decay effect, where historical interactions gradually lose relevance. To bridge these gaps, we introduce SpectraDyN, a unified framework integrating spectral-temporal modeling. Unlike conventional Fourier-based methods, SpectraDyN incorporates a wavelet transform layer to achieve multi-resolution decomposition, enabling the distinct capture of both long-term community evolution and short-term event-driven interactions. This aligns with the multi-scale nature of social dynamics, where stable circles and temporary gatherings coexist. Additionally, we design a time-aware decay mechanism to heavily prioritize recent structural changes while preserving essential historical patterns. Extensive experiments on six real-world datasets demonstrate that SpectraDyN consistently outperforms state-of-the-art baselines, particularly in inductive link prediction tasks, validating the robustness and effectiveness of our integrated approach.</p>