Deep learning approaches have been increasingly adopted for virtual analog modeling, which aims to replicate the sonic characteristics of analog audio devices. In the context of analog dynamic range compressor modeling, many existing methods operate directly on raw audio waveforms which are high-dimensional and contain fine-grained temporal features at high sampling rates. These representations are computationally demanding and limit model efficiency. We propose a feature extraction pipeline that leverages the magnitude component of the Short-Time Fourier Transform in combination with a spectral amplification mechanism which acts similarly to a spectral mask but can both attenuate and amplify selected frequency components. We employ multi-band Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures that split the magnitude spectrum into several frequency bands for independent processing, substantially reducing computational complexity while preserving high modeling accuracy. To evaluate our approach, we created two datasets consisting of recordings of the consumer-grade analog compressor Alesis 3630 and its digital counterpart, discoDSP NightShine. We conducted extensive experiments comparing our method against raw waveform baselines using four objective metrics, theoretical and empirical measurements of computational performance, and a subjective listening test. Results indicate that single-band models based on the proposed feature extraction pipeline outperform raw-audio baselines across all evaluation metrics. Multi-band configurations further improve the efficiency balance. In particular, four-band LSTM and GRU architectures achieve higher perceptual fidelity at substantially lower computational cost. Moreover, we conducted a subjective listening test that yielded results aligned with the objective metrics. All source code and pretrained models are provided for reproducibility.