A prediction method for wine customer consumption behavior using data enhancement and attention-based CNN-LSTM

Purpose This study aims to address the challenge of low prediction accuracy in wine consumer behavior due to poor data quality and the limitations of traditional statistical models in capturing nonlinear relationships. It proposes a deep learning approach combining data augmentation and an Attention-based CNN-LSTM model to enhance the precision of consumer behavior prediction in China’s wine market. Design/methodology/approach The research integrates two key components. Data enhancement and dimensionality reduction, an Improved Gaussian Mixture Model (IGMM) was used to generate synthetic samples by learning from original data distributions, with K-means clustering for label assignment. Kernel Principal Component Analysis (KPCA) was applied to reduce dimensionality and mitigate noise. Prediction model construction, a CNN-LSTM-Attention architecture was designed, where CNN captures local features, LSTM handles long-term dependencies and Spatial Attention (SA) weights critical features. Findings Experimental results show that, IGMM generates reliable synthetic samples, boosting CNN prediction accuracy from 0.597 to 0.970 compared to traditional oversampling methods. The CNN-LSTM-Attention model achieves an accuracy of 0.987, outperforming state-of-the-art algorithms. Finally, the robustness of the two methods was verified based on the UCI datasets. Originality/value This study innovatively combines IGMM-KPCA for data enhancement with an Attention-driven hybrid neural network, providing a novel framework for consumer behavior prediction in the wine industry.