Next Generation of Click Through Rate Prediction — Episode 1
DCN v3 Paper Review
In the past few years, I have been mostly reading about Causal Inference and Experimentation. There is a Chinese saying that goes like: “Never forget the first fire.” I’m rekindling my first fire by delving deeper into state-of-the-art algorithms in Recommendation systems.
In this article, I will try to share my understanding of the DCNv3 architecture.
Introduction
In e-commerce and ads, click is a signal of interest (and sometimes $ for the supplier), so predicting it and getting it right is beneficial for the consumer who can receive relevant topics/items and for the advertisement ecosystem.
What is DCN?
Deep & Cross Networks (DCN) is a model architecture that was proposed to improve CTR prediction by modeling feature interactions (ways different features combine and influence the prediction outcome) more effectively.
DCNs construct both Explicit and Implicit feature interactions. Explicit feature interactions are directly modeled and can be interpreted. DCN models use a Cross-Network component that systematically combines features in a way that explicitly captures their interactions. It applies cross terms between features across multiple layers…
