The opportunity
As a Senior Machine Learning Engineer on the Ads Response Prediction team, you will own and execute the development of ML models that power Instacart's ads ecosystem. This is a research-leaning role focused on theoretical problem formulation, training methodology, and model…
What you'll do
Own and execute research and development of pCTR and conversion prediction: models, with a focus on improving calibration, reducing training data biases (selection bias, position bias, optimizer's curse), and advancing model accuracy across Instacart's ads surfaces.
Design and implement debiasing techniques such as Mixed Negative Sampling: (MNS), Inverse Propensity Weighting (IPW), counterfactual risk minimization, and calibration methods (Platt scaling, isotonic regression) to address systematic prediction biases.
Contribute to the next-generation Multi-Domain Multi-Task (MDMT) model: architecture, incorporating innovations like Mixture-of-Experts (MoE), Transformer layers for sequential user behavior, and LoRA adapters for scalable domain fine-tuning.
Contribute to sequence modeling initiatives including the TIGER generative: retrieval system and Semantic ID representation learning, expanding their application across ads surfaces such as Product Details, Search and other placements.
Collaborate with the broader ML community in the company on the path toward: Foundation Models using autoregressive user behavior prediction.
Formulate and scope ambiguous modeling problems within your project scope: from first principles. Translate business observations (e.g., overcalibration patterns, cold-start underperformance) into well-defined ML research directions with clear evaluation criteria.
What they're looking for
- Experience in ads ranking or auction-based systems (pCTR, bid optimization,: ROAS feedback loops, marketplace dynamics).
- Hands-on experience with autoregressive sequence models for user behavior: prediction, generative retrieval, or transformer-based ranking architectures.
- Familiarity with learned representations such as Semantic IDs, product: embeddings, or other approaches to reducing feature cardinality and cold-start challenges.
- Experience with transfer learning or domain adaptation techniques (e.g.,: LoRA, adapter-based fine-tuning) applied to recommendation or ranking models.
- Publication record in top-tier venues (KDD, WWW, RecSys, NeurIPS, ICML, SIGIR, or similar).
- Familiarity with LLM-driven approaches to recommendation, including: prompt-based personalization and AI-assisted model development (AutoML).
