LocalGLMnet: interpretable deep learning for tabular data

Deep learning models have gained great popularity in statistical modeling because they lead to very competitive regression models, often outperforming classical such as generalized linear models. The disadvantage of deep is that their solutions are difficult interpret and explain, variable selection not easily possible solve feature engineering internally a nontransparent way. Inspired by the appealing structure we propose new network architecture shares similar features but provides superior predictive power benefiting from art representation learning. This allows for tabular data interpretation calibrated model, fact, our approach an additive decomposition can be related other model interpretability techniques.