Household Embeddings: Introducing continuous vector representations for car ownership on a household level

Abstract

This study presents a new method of representing travel-related categorical variables and observations using continuous vector representations, commonly known as embeddings. Specifically, we focus on generating embedding representations that aim to describe car ownership on a household level based on methods and approaches that are used in the field of Natural Language Processing (NLP) and Machine Learning (ML). We build on the previous work on traveling embeddings by Pereira [1] and extend it with a joint embedding space approach that allows us to leverage on the compositionality of the categorical vectors and introduce a method that uses embedding centroids to represent back individual observations, i.e. household embeddings. The efficiency of the centroid-based method is tested by comparing two binary logit models for car ownership: one that uses the embedding centroids encoding and one that uses the traditional dummy encoding. For our car-ownership modelling case, the results show that using embedding centroids encoding in utility specification performs better that direct categorical variables in out-of-sample prediction. We further demonstrate that the proposed method not only produces meaningful representations of the categorical space, but also allow us to define prototypical households for the behaviour at stake.