Representation Enhancement of Tabular Data by Embedding of Ordinal Features

Han Liu, Yinghui Pan, Hao Chen

Abstract: Motivated by the success that transforming symbolic words into numeric vectors through embedding models leads to a significant improvement of feature representation in text processing, in this paper, we propose an approach of numeric encoding to extend the idea of word embedding to general mapping of tabular data. Considering a supervised learning problem where features take symbolic values from ordered space, our method transforms ordinal features into numeric ones within the setting of representation learning. Specifically, our approach aims at transforming label-encoded features into ones represented in a dimensionality-varied space in the setting of multi-scale feature extraction driven by ensemble learning. Experimental results show that our framework indeed achieves enhancing the feature representation, which leads to a significant improvement of learning performance in comparison with widely used encoding methods such as label encoding and feature hashing.