Deep Learning on Tabular Data

Abstract

This thesis examines the capacity of deep learning models to handle tabular data and whether they can surpass traditional methods. Despite the development of deep learning, applying it to tabular data has been less explored. Also, tabular data generation has been investigated to try to improve deep learning performance on tabular data. A related work chapter has been written to review what has already been done for deep learning on tabular data and tabular data generation. Then, key concepts have been defined to clarify the theoretical framework. A special focus has been made on the limitations of deep learning on tabular data. The datasets in this thesis are all designed for regression tasks with numerical features. The methodologies include the explanation of each deep learning model chosen for this thesis, the interpretability tool SHAP (Shapley Additive Explanations), and the tabular data generation method using Large Language Models (LLMs). Models were ranked for each dataset, and an average rank across datasets was obtained. The results proved that the traditional methods outperformed the majority of deep learning models. However, some deep learning models were close to them proving their potential. The SHAP analysis provided insights into the performance of the models by highlighting which features contributed most to their predictions. Generating tabular data with LLMs has been tested. The results are dependent on the dataset used, meaning that performance can improve or deteriorate. To conclude, deep learning can be a viable alternative to traditional methods, but it has limitations, particularly computational.