Recommender Systems for Merchants

2/25/2026

This post details the development of a foundational generative recommender system for Shopify. It covers architectural decisions for handling sequential buyer journeys using autoregressive models with causal masks, incorporating time as a first-class signal through RoPE-inspired rotary encoding and relative attention bias, and implementing optimized negative sampling strategies (shared negatives, positive-aware hard negatives) to improve retrieval quality. The system also employs a boosting-inspired approach to achieve incremental recall within an ensemble of recommender models. Significant focus was placed on training efficiency, achieving up to 7.3x speedup through optimized CUDA kernels and pipeline optimizations. Online A/B testing showed positive impacts on key e-commerce metrics.

7/4/2025

This post details the application of machine learning at Shopify, expanding on recommender systems for merchants. It introduces the use of finetuned Qwen multimodal models for product classification and enrichment, fraud detection models inspired by Feature Space, LLaMa and general models for the Sidekick merchant assistant, Nomic embeddings for product catalog vector representations, tabular transformer models for GMV forecasting, HSTU architecture for behavior modeling, and small language models for query rewriting. It also highlights investments in GPU infrastructure (GCP, Nebius), partnerships for GPU computation acceleration (CentML), data annotation services (Toloka), and early access to frontier models.

2/11/2022

This post details Shopify's playbook for scaling machine learning, focusing on a pragmatic, step-by-step guide applicable across any domain. It outlines five key steps for starting from zero: identifying a problem worth solving (e.g., order fraud detection), ensuring sufficient and accessible data, identifying downstream dependencies, understanding existing solutions to set benchmarks, and optimizing for product outcomes over pure model scores. It then details the 'Zero to One' phase, emphasizing the importance of well-defined pipelines for training and testing, and strategic model deployment decisions based on volume and user commitment. The 'One to One Hundred' phase focuses on building trust in models through input/output reconciliation, production backtesting, and continuous monitoring, and on scaling model building workflows by investing in data engineering practices and efficient deployment strategies. The post also touches upon the application of AI for complex routing problems and the development of AI assistants.

8/25/2021

This post outlines a five-step business-centric approach to building and productionizing machine learning models. It emphasizes asking if ML is the right solution, starting with simple models and heuristics, measuring impact before optimizing, and having a plan for iteration. Specific techniques discussed include using simple models like linear regression, composite indices, and keyword matching as baselines, and the importance of tying performance metrics to business impact. It also highlights the need to understand model tradeoffs through confusion matrices and residual distributions.

3/4/2020

This post details the evolution of Kit from a rule-based system to an AI-driven assistant. It describes the transition from heuristic rules for marketing campaign configuration to machine learning models for predicting optimal budget (regression) and likelihood of sales (classification). The post outlines the architecture for training and prediction flows, including feature engineering using Apache Spark, model training on Google Cloud's ML Engine, and real-time prediction using Apache Kafka. It also introduces a new use case for acquiring visitors to build lookalike audiences, requiring a shift to real-time prediction.

1/25/2019

This post details the design and implementation of Shopify's recommender system platform. It explains the use of collaborative filtering with user-item interactions, representing preferences as binary vectors. The core methodology involves treating recommendation as a binary classification problem and leveraging linear models (like LRec) to learn item-item similarity matrices. The post highlights the scalability through task parallelism, interpretability via model coefficients, extensibility by incorporating user features, and the importance of diverse performance metrics beyond precision for matching merchants with relevant content and experts.