Basket2vec: Learning Retail Basket Representations From Transactional Data

Abstract

In this work, we present the novel basket2vec methodology to learn a rich retail basket representation in a latent space with an aim to preserve the co-occurrence relationship between items. We evaluate our methodology using real retail data consisting of millions of transactions; when compared with sparse approaches, our basket2vec methodology showed significant improved performance of at least an order of magnitude. We induced useful geometrical properties to structure the resulting latent space. Furthermore, we show how to use our basket2vec methodology for three applications: $\mathcal {K}$ -means clustering, item recommendation system, and generative modeling. Our basket2vec representation can be regarded as a general analytical tool, to be used in several retail applications, such as: next-item and -basket prediction, basket-level trend evolution analysis, and customer segmentation based on transactional data.

Keywords

• Market basket analysis
• representation learning
• transactional data analysis
• recommendation system