Interpretability Enhancement Strategy of Big Data Models based on Deep Learning

Current deep learning models have low interpretability in big data processing. This paper proposes an interpretability enhancement strategy that combines LIME and SHAP to solve the "black box" problem of deep learning models in big data scenarios. LIME is used to generate local pseudo-data to fit the linear model and analyze the prediction logic of specific inputs; SHAP is combined to quantify the contribution of global features and adjust model parameters to balance performance and interpretability. The experiment uses multi-dimensional evaluation of accuracy (0.93), interpretability score (0.83) and computing time (36.2 seconds), which shows that this strategy significantly improves decision transparency while maintaining high accuracy. Compared with traditional methods, this solution makes model behavior easier to understand and provides a reliable basis for big data analysis.