Improving feature interactions at Pinterest under industry constraints

TL;DR

This paper investigates how to improve feature interactions for Pinterest's Homefeed ranking under real-world industry constraints such as memory, latency, reproducibility, and stability. It systematically evaluates a range of interaction architectures, emphasizing high-order and parallelized patterns, and demonstrates that a configuration combining four DCNv2 layers with three parallel MaskNet blocks offers a favorable balance of predictive gains and operational costs. The study reports offline and online results, with a final deployment that preserves reproducibility while achieving measurable engagement improvements, and it provides practical guidance for architecting interaction learning under constraint-driven settings. Overall, the work highlights how constrained industrial settings can guide the selection and tuning of feature-interaction models to deliver robust, scalable improvements. It also emphasizes the need for collaboration between academia and industry to develop architectures that perform well under practical limitations.

Abstract

Adopting advances in recommendation systems is often challenging in industrial settings due to unique constraints. This paper aims to highlight these constraints through the lens of feature interactions. Feature interactions are critical for accurately predicting user behavior in recommendation systems and online advertising. Despite numerous novel techniques showing superior performance on benchmark datasets like Criteo, their direct application in industrial settings is hindered by constraints such as model latency, GPU memory limitations and model reproducibility. In this paper, we share our learnings from improving feature interactions in Pinterest's Homefeed ranking model under such constraints. We provide details about the specific challenges encountered, the strategies employed to address them, and the trade-offs made to balance performance with practical limitations. Additionally, we present a set of learning experiments that help guide the feature interaction architecture selection. We believe these insights will be useful for engineers who are interested in improving their model through better feature interaction learning.

Figures (2)

• Figure 1: Pinterest Homefeed Page
• Figure 2: Homefeed ranking model