What is it about?

In this training we provide all necessary knowledge to design a complete fully-functional recommender system from scratch. The training has three parts. In the first part, we focus on algorithms. We start from content-based filtering and neighbor-based vector similarity approaches. We then extend the similarity-based approaches to collaborative filtering and introduce the state-of-the-art matrix completion approaches based on matrix factorization. Further, we demonstrate how one can use social information and regularize matrix completion algorithms on the social graph. To make this training relevant for teams building modern "always-on" services and applications (e.g. location-aware geo services, personalzied chatbots and conversational UIs, style profiles), we will cover conxtextual recommendations and explain how one can use location, gender, real-time movement patterns and other personal data in recommendations.

In the second part, we talk about proper evaluation, A/B testing and experiment design methodologies specifically developed for recommender systems. For example, we will consider a very powerful explore-exploit conceptual framework and explain how it could be used to do dynamic model tuning and online experimentation. We will also cover the user interface aspects of the recommender systems, including explanations for recommendations, trustworthiness, and optimal rating scale design, among others.

Finally, we switch to the hands-on part of the training and help the participants design their own recommender systems using Spark (focus on algorithms) or Elasticsearch (focus on the working demo) by incorporating all lectured materials. We will use large-scale publicly accessible datasets of movies, academic publications, pins from Pinterest, Twitter posts, as well as data relevant for your company or organization.

Training Content