Real-time machine learning is a challenging system’s domain, but one where huge value can be created, as shown by companies such as TikTok. The best personalized search and recommendation systems are based on real-time ML with a Feature Store, a Vector Database, and a Model Serving platform, serving recommendations on-demand taking into account user history and context (such as ‘trending’ content).

In this webinar, we will analyze the synergies of an integrated Feature Store and Model Serving platform for the operationalization of real-time ML-enabled services, including the key MLOps principles needed to ensure integrated version management for upgrading and downgrading models and the features that feed them. We will show an implementation of a real-time, personalized recommendation system using Hopsworks.

00:00 Introduction
—
00:20 Where business value is generated in AI
—
02:00 Where Feature Stores and Model Serving meet
—
03:00 From Raw Data to Online Predictions
—
07:06 Keeping Your Pipelines on Track
—
11:24 A Closer Look to Inference Pipelines
—
13:50 A Deeper Look to Real-time Inference Pipelines
—
18:15 Overview of Real-time Recommendation Systems
—
21:51 Candidate Retrieval with Two-Tower Model
—
24:04 Real-time Recommendation Systems
—
24:58 Real-time Recommendation Systems with Hopsworks
—
26:05 Demo
—
32:25 Conclusion

source