The study firstly tries to segment the value of telecommunication customers through data mining methods, and introduces variable convolution on the basis of depth-weighted K-mean algorithm for improvement. Meanwhile, grid search is introduced on the basis of extreme gradient boosting tree for optimization, and finally a telecom customer recommendation marketing model is proposed by combining the two optimization algorithms. The experiments use a publicly available dataset from Kaggle that contains telecom customer behavior data, call records, billing records, and service usage from China in 2019, totaling about 2 million pieces of information. The experimental results show that the highest value of classification accuracy of the improved depth-weighted K-mean algorithm is 95.5%, and the highest separation degree is 96.3%. In summary, the proposed model can effectively categorize telecom customers and rationally implement telecom product recommendation. The study aims to provide telecom companies with more accurate marketing decision support to improve customer satisfaction and market competitiveness.