Retail Analytics With Neural Product Embeddings

In today's rapidly evolving retail landscape, understanding market structure and consumers' perceptions of products is more critical than ever. Product2Vec is a robust representation learning approach inspired by advances in natural language processing (NLP) that captures product attributes and market structures in a way that is both scalable and easy to use.

What is Product2Vec?

Product2Vec is an approach inspired by techniques from natural language processing. It treats products like words in a sentence, analyzing how frequently they co-occur in shopping baskets. This generates latent product attributes, which are vector representations that reflect how consumers view products. Latent product attributes capture attributes such as product category, price, packaging, and size. Go to GitHub for a PyTorch implementation of Product2Vec.

Why Product2Vec?

Product embeddings are useful because they describe products without manually curating products attributes. This which Product2Vec an excellent foundation for recommender systems and targeting solutions. Embeddings also generate insights into cross-category relationships, such as complementary products often purchased together.

One application of Product2Vec is P2V-MAP, an approach for visualizing market structure for large retail assortments. P2V-MAP produces a two-dimensional product map which contains insights into products’ within-category competition and cross-category complementarity. The map is easy to understand and enables retailers and manufacturers to optimize assortments, store layouts, promotions, and product positioning.

We have applied Product2Vec to market basket data from over 50 retailers in online and offline settings. The results generated actionable insights for product strategy, target marketing, and assortment optimization.

References

Gabel S, Guhl D, and Klapper D (2019). P2V-MAP: Mapping Market Structure for Large Assortments. Journal of Marketing Research, 56(4):557–580.