Abstract: ‍ ‍Channel state information is very vital to obtain high channel capacity and energy efficiency in massive Multi-Input Multi-Output （MIMO） systems. Because the uplink and downlink channels lack reciprocity， the user needs to feed back the downlink channel to the base station for processing such as precoding in frequency division duplexing system. The feedback overhead is huge cause the feedback information is proportional to the number of antennas in the massive MIMO systems. Many data-driven deep learning neural networks use encoders to compress channel state information and decoders to restore channel. However， due to the black box characteristics of data-driven neural networks， not only high complexity is required to recover channel state information， but better performance is difficult to obtain， especially improving performance by increasing complexity and computing resources at the user is impractical. This paper proposed an interpretable model-driven network that unfolds the Iterative Shrinkage-Thresholding Algorithm （ISTA）. Because channel state information does not strictly meet the sparsity requirements of compressed sensing， which leads to the degradation of recovery performance， a nonlinear sparse transform was designed to improve the reconstruction performance. In order to balance performance and complexity， a learnable compression matrix was ​​proposed at the UE to reserve more channel state information. Moreover， this paper expands the scope of application from single-antenna users to multi-antenna users without changing the existing network framework， which improves the compatibility of neural networks. Simulation results show that the proposed method had better channel state information recovery performance and lower complexity at the user than other methods.