Abstract: To optimize the ambiguity resolution in real-time kinematic (RTK) positioning, a GNSS ambiguity resolution method based on both data and model-driven approaches is proposed. This algorithm focuses on selecting the optimal subset of ambiguities to increase the success rate of ambiguity fixing. First, considering the influence of ambiguity subsets on the accuracy of baseline solutions, a subset method based on the ambiguity level is adopted to ensure a high success rate of ambiguity fixing and high accuracy of baseline solutions. Model- and data-driven approaches are combined during partial ambiguity resolution, incorporating long short-term memory (LSTM) neural networks to ensure the stability and reliability of ambiguity fixing solutions. Results show that, while ensuring high reliability of ambiguity fixing, compared to existing algorithms, the LSTM algorithm based on resolution rules improves the fix rate and fix success rate of the selected dataset from 73.59%, 72.83%, 93.07%, and 87.57% to 99.23%, 95.51%, 98.67%, and 92.74%, respectively. Both the accuracy and robustness of RTK positioning are thus enhanced.