The use of bistatic reflected global navigation satellite system (GNSS) signals as a means of sensing the Earth’s surface is attracting widespread interest. It has the advantages of non-contact, large coverage area, and real-time which have attracted much attention during recent years. These reflected signals contain the information of the reflecting surface and therefore were applied to investigate the properties of the observed object, such as soil moisture (SM). Machine learning (ML) methods are featured with flexibility and are good at handling non-linear problems, modeling complex interactions between inputs and outputs, and have been rise attention for the GNSS-R SM retrieval field. The contribution of different input variables to SM is quite significant for optimizing the ML-based SM retrieval. In this paper, the typical random forest (RF) algorithm was adopted to evaluate the weight of input variables for ML-based SM retrieval. A simulation data set was built for training RF models, since the simulated data provide sufficient samples and show a more accurate relationship between the inputs and outputs. The SM predictions made by the RF methods are evaluated and compared with the simulation data set. The results show the contribution of a single variable to soil moisture retrieval, which can help with the ML-based GNSS-R SM retrieval to overcome the complex auxiliary variable problem.