Overview

Google, IBM & Meta Certificates – 40% Off

One Coursera Plus subscription covers most Professional Certificates on Coursera.

Ranking is not just about accurately scoring candidates. Production recommendation systems must balance competing engagement goals, enforce business constraints, and leverage generative AI to deliver relevant, fair, and diverse results. This course teaches you to fine-tune the final mile using multi-task architectures, post-processing re-ranking layers, and LLM-augmented pipelines. You will design MMoE and PLE architectures in PyTorch to isolate task conflicts across engagement objectives like Saves, Dwell Time, and Shares. You will apply MMR and Bayesian optimization for diversity, freshness, and creator equity constraints. You will also build generative recommendation proxy architectures referencing LinkedIn's 360Brew principles, using Azure OpenAI embeddings and zero-shot and few-shot LLM rankers. Designed for Senior ML Engineers and Recommendation Platform Engineers responsible for post-ranking optimization, multi-objective scoring, and LLM pipeline integration. You should be comfortable with Python, PyTorch, and core ranking evaluation metrics.

Syllabus

Re-Ranking: Business Rules & Diversity
- Learn how to apply post-processing business rules and diversity constraints to heavy-ranking outputs using score blending and Maximal Marginal Relevance (MMR).
Re-Ranking: Fairness & Exposure Bias
- Architect fairness-aware adjustments to ensure equitable exposure across new versus established platform members without severely degrading overall recommendation quality.
Multi-Task Arch: MMoE & PLE
- Learn how to isolate task conflicts across competing engagement objectives by designing advanced gating networks and specialized expert blocks in PyTorch.
Multi-Task Arch: Composite Scoring
- Transition from architectural design to objective optimization by implementing joint loss functions and tracking long-term utility metrics.
LLM-Augmented & Generative Recommendation
- Learn how to replace standard hand-engineered feature pipelines with deep semantic representations using Azure OpenAI, setting the stage for generative item scoring.
Evaluating Foundation Model Proxies
- Analyze the systemic trade-offs of deploying generative rankers, comparing semantic models against traditional deep interaction networks.
GenAI Module: AI-Assisted Generative Ranking
- Analyze the operational scalability, token economy, and deployment cost vectors of deep foundation model rankers. You will evaluate how foundation models act as centralized, surface-agnostic rankers compared to traditional feature-factory scoring layers leveraging Azure OpenAI and enterprise cloud vector architectures.
Project Module: Multi-Objective & Semantic Re-Ranking Engine
- Implement an integrated post-scoring pipeline that combines multi-task evaluation modules with post-processing re-ranking layers informed by explicit target constraints. You will analyze integrated model deployment metrics across holistic system dimensions by tracking precision (NDCG@10), diversity entropy, and exposure fairness parameter balances.