Intelligent news text classification and interpretability analysis

With the rapid development of social media and online news portals, the volume of news data has grown exponentially, posing new challenges for information filtering and classification. Deep learning models, such as TextCNN, have been widely adopted for news text classification; however, the issue of model interpretability remains a critical challenge. This study proposes an approach to enhance the interpretability of CNN-based classification models by leveraging logistic regression analysis. Specifically, the final convolutional layer’s output is first transformed into real-valued representations to facilitate regression-based modeling. Subsequently, a logistic regression model is trained and evaluated for accuracy. Finally, by analyzing the regression model’s weights and computing the importance scores of convolutional filters, we quantify the contribution of different filters to classification decisions, thereby improving the transparency and interpretability of the model.