Ranking Models & Feature Engineering

Microsoft via Coursera

Go to class Write review

Overview

Google, IBM & Meta Certificates – 40% Off

One Coursera Plus subscription covers most Professional Certificates on Coursera.

Unlock All Certificates

Progress from candidate retrieval to high-precision heavy ranking layers operating under strict online serving latency constraints. This course teaches you to build and deploy the deep learning ranking architectures used by enterprise platforms to maximize long-term user engagement. You will construct high-throughput feature schemas and real-time rolling aggregates using Databricks Unity Catalog Feature Store and Azure Cache for Redis. You will implement deep ranking architectures, including Wide and Deep, DeepFM, DLRM, and Residual DCN in PyTorch, optimizing for multi-objective outputs across high-cardinality feature matrices. You will also build multitask learning models using Multi-gate Mixture-of-Experts (MMoE) and Progressive Layered Extraction (PLE) to counter negative transfer and balance competing engagement signals, such as Saves, Dwell Time, and Private Shares — implementing the multi-task ranking network architecture and end-to-end training pipeline directly in PyTorch. Once trained, you will register and version your models into the Azure ML model registry using Azure ML SDK v2, integrating Azure Machine Learning into your deployment workflow alongside the Unity Catalog and Azure Databricks environment. This course is designed for Machine Learning Engineers, Recommendation Infrastructure Engineers, and Data Platform Architects responsible for designing multi-objective scoring layers that optimize long-term user value. You should be comfortable with Python, PyTorch, core ML evaluation metrics, and have foundational familiarity with Azure Machine Learning and Azure ML SDK v2.

Syllabus

Feature Engineering: Optimized Schema Design
- Learn how to engineer high-scale streaming transformations and schema definitions for heavy ranking models. You will design feature collections that unify static member attributes with rolling-window aggregates in the Databricks Unity Catalog Feature Store.
Online Feature Hydration & Caching
- Master the mechanics of real-time user state hydration. You will configure programmatic lookup interfaces that pull dense feature arrays from low-latency caches to feed live heavy ranking inference loops.
Deep Ranking: Implementing Core Architectures
- Learn how to translate industrial deep ranking paradigms into operational PyTorch models. You will build and connect specialized embedding tables, explicit factorization components, and multi-layer perceptrons (MLPs) to handle high-cardinality sparse fields.
Offline Evaluation & Optimization
- Learn how to run comprehensive offline evaluations and optimization loops for deep ranking networks. You will monitor validation curves via MLflow and interpret cross-entropy and tracking anomalies over high-cardinality datasets.
Scaling Feature Crossing with DCNv2
- Learn the mathematical foundations of bounded feature crossing. You will implement DCNv2 components in PyTorch and apply low-rank matrix approximations to control the exponential parameter explosion inherent in high-order interaction networks.
Residual DCN And LinkedIn's LiRank
- Elevate base cross networks to industrial production standards. You will implement attention scaling, skip connections, and residual pathways inspired by LinkedIn's LiRank framework to maximize heavy ranking precision.
Bias Mitigation and Inverse Propensity Scoring
- Learn to recognize and mathematically neutralize the bias injected by your platform's UI. You will implement IPS weights inside custom PyTorch loss functions to ensure high-quality items aren't penalized simply for being displayed lower in historical feeds.
Post-Hoc Prediction Score Calibration
- Ensure your model's outputs are trustworthy. You will evaluate miscalibration issues in deep neural networks and implement Isotonic Regression post-processing layers so that a raw score of "0.2" truly equates to a 20% empirical click probability.
Project Module: Multi-Task Ranking Engine
- Construct an end-to-end heavy ranking network that integrates Residual DCN cross-layers with multi-objective deep factorization modules. You will author a comprehensive PyTorch model script that implements parallel Residual DCN interaction blocks feeding into dynamic classification components, and trains the model to optimize long-term utility outputs such as Saves, Dwell Time, and Private Shares. Finally, you will programmatically register your final architecture and compile your validation metric summaries into a portfolio-ready engineering document.