This paper presents a novel data categorization approach designed specifically for Data Middle Platforms. It combines the robust global processing capabilities of the Transformer model with the nuanced insights of convolutional neural networks. The proposed method introduces an innovative associative attention mechanism that effectively integrates both global and local information, enabling precise classification of complex datasets within the Data Middle Platform framework. Through extensive experiments, we provide evidence of significantly improved classification accuracy and heightened sensitivity to subtle semantic differences in large-scale datasets. This approach marks a new paradigm for efficient and accurate data categorization within Data Middle Platforms, highlighting the transformative potential of deep learning in advanced data processing environments. The findings underscore the importance of leveraging a synergistic blend of global and local insights for enhanced performance in data categorization tasks within the evolving landscape of Data Middle Platforms.