The use of Raman spectroscopy for real-time gas monitoring has the advantages of response speed, high sensitivity and low cost. However, due to the overlap of each peak position, the Raman spectral data usually exhibit high dimensionality, complex nonlinear relationships and significant noise interference, which makes it difficult to directly determine the composition of the mixture using traditional data processing methods. This work focuses on the optimization of a machine learning model, XGBoost, for predicting gas composition based on Raman spectral data, enhancing predictive accuracy through three different feature extraction and feature selection methods. The superior performance of the XGBoost model is demonstrated by comparison with other machine learning models, including decision trees, random forests, support vector machines and neural networks, using the Raman spectrum of a gas mixture of hydrogen, ethylene, propylene and butene. The results show that XGBoost has better accuracy and generalization ability for quantitative analysis of Raman spectra, making it suitable for complex chemical process monitoring.