Video - 'A Lightweight Transformer for Next-Item Product Recommendation'. In this talk, we present our latest model for next-item product recommendation using a sequential input of a customers' browsing history. We show that the standard log discount in nDCG is insufficiently strong for our real observed positional effect, suggest a way to determine a better discount empirically. We also show that our model is very plausibly learning to distinguish different item style/functionality/material and can even connect different types of furniture to provide cross-category recommendation. Lastly, we show that the ratio of customers viewing each distinct item in a catalog can be used to diagnose thin data issues and align model performance across different catalogs.

Presentation Video for "A Lightweight Transformer for Next-Item Product Recommendation". We apply a transformer using sequential browse history to generate next-item product recommendations. Interpreting the learned item embeddings, we show that the model is able to implicitly learn price, popularity, style and functionality attributes without being explicitly passed these features during training. We also show that these attributes transfer across different types of furniture. Our real-life test of this model on Wayfair?s different international stores show mixed results (but overall win). Diagnosing the cause, we identify a useful metric (average number of customers browsing each product) to ensure good model convergence. We also find limitations of using standard metrics like recall and nDCG, which do not correctly account for the positional effects of showing items on the Wayfair website, and empirically determine a more accurate discount factor.