Abstract

There is a recent growing interest in applying Deep Learning techniques totabular data, in order to replicate the success of other ArtificialIntelligence areas in this structured domain. Specifically interesting is thecase in which tabular data have a time dependence, such as, for instancefinancial transactions. However, the heterogeneity of the tabular values, inwhich categorical elements are mixed with numerical items, makes thisadaptation difficult. In this paper we propose a Transformer architecture torepresent heterogeneous time-dependent tabular data, in which numericalfeatures are represented using a set of frequency functions and the wholenetwork is uniformly trained with a unique loss function.

 

Quick Read (beta)