29. Intelligent Customer Preference Engine with Real-time ML systems, Manoj & Praveen, Walmart Lab

We need to understand behaviour across the whole channels

Personalization plays a key role throughout the customer journey, and also after the purchase in case of recommendation

Dozen of recommenders in play in the background, to make baskets very quickly. Adding each item can be time-consuming, so personalisation is right from the home page.

Product discovery is another recommender use case.

Aside from product impressions, we also prioritize non-product impressions like banners, to show that we’re to help find items they’re looking for.

For last-call, we don’t want customers going back to the discovery again, but go ahead and checkout

ex of micro intent: Party supplies, then ice cream, then gift cards

detecting micro-intent with the inference engine

Variable time windows for online predictions

If a user already has an iPhone and search for AirPods, we’ll propose Apple ones first.

Model store allows to A/B test and to iterate quickly on different models

Multi-model, DS can bring their framework and model on it and spin up a microservice easily

Can tie different models together in an ensemble fashion

Inference graph is very high-level: expose human-readable API to our clients. Most of the time, embedding and transformations are only understandable by the fews who built the model. Given a feature, it can bundle features together using the feature store.

Post-processing is a filtering mechanism of the recall set, response to the client

Real-time platform is both async and real-time for different use-cases.