Accurately capturing the behavioral factors of different types of customer groups and adopting targeted service strategies is the key to business competition in the hotel industry. In this paper, we combine the variance Boston matrix and PSO-based K-means algorithm to achieve hotel customer attribute segmentation based on customer behavior, customer value and word-of-mouth reliability, and then use deep learning algorithms to construct a hotel customer behavior prediction model. The feature fusion layer and SENet are incorporated into the residual network in order to utilize the feature expression ability of different layers and the spatial coding ability between different channels to enhance the hotel customer behavior predictive ability. Downloading the public dataset from the online wine travel platform for example analysis, it is found that the classification of this paper’s algorithm before customer segmentation has a correct rate of 83.75%, which is higher than the rest of the baseline models. After customer segmentation this paper’s algorithm achieves the highest recall rate in all customer categories, and the recall rate is as high as 84% on category 1 customer groups, and the superiority of the designed algorithm is verified. This study facilitates hotel management to target customer service and retention according to different customer groups.