In order to improve the convergence and generalization of feasible solution sets for combinatorial optimization problems, a data-correlation-aware deep fusion pointer network model (DMAG-PN) is proposed according to the characteristics of different unsupervised learning strategies. The model integrates Mogrifier LSTM, multi-head attention mechanism, and graph convolutional neural networks through a pointer network framework. Firstly, the embedding layer in the encoder module encodes the input sequence, and the multi-head attention mechanism is introduced to obtain the feature informations in the coding matrix. Secondly, the data correlation model is constructed to explore the correlation between the sequence nodes, and the graph convolution neural network is used to obtain the multi-dimensional correlation feature information, so as to generate multiple embeddings to effectively capture the deep node and edge features of the sequence. Finally, the decoder module based on multi-head attention mechanism takes node embedding data and fusion graph embedding data as inputs to generate a global probability distribution for selecting the next unvisited node. The symmetric traveling salesman problem is used as the test problem, and compared with the current advanced algorithms, the experimental results show that the proposed DMAG-PN model has been greatly improved in terms of generalization and accuracy. The pre-trained DMAG-PN model can directly solve large-scale instances, avoiding the iterative search process of traditional algorithms, and has a high solution efficiency.