Teaching / CSCI 6517
Summer 2026 · Dalhousie UniversityRecommender Systems
Recommendation is where machine learning meets changing human preference. This graduate course moves from neighborhood methods and matrix factorization to deep sequential models, production ranking pipelines, foundation models, and conversational systems.
What you will learn
Implement neighborhood methods, matrix factorization, and content-based recommendation.
Use neural, graph, transformer, bandit, and reinforcement-learning approaches.
Understand candidate generation, blending, multi-stage filtering, heavy ranking, and scalability.
Evaluate performance and revenue metrics while considering bias, privacy, explainability, and conversation.
From one score to a complete system
Weekly syllabus
Problem formulation and neighborhood-based collaborative filtering
Matrix factorization and preference embeddings
Model-based collaborative filtering; content representations
Neural architectures, training objectives, and bias control
Offline metrics, business metrics, dynamic preference, and representation propagation
Social networks and graph convolution for recommendation
Sessions, order, time, and transformer-based sequence modeling
Multi-objective and multi-stage systems; mixture of experts and heavy rankers
Personalized search and combining multiple candidate sources
Large language models and conversational recommendation
Bandits, preference elicitation, and reinforcement-learning-based conversation
Critiquing-based conversational systems and project development
Streaming recommendation and project presentations
What you will build
Turn similarity into personalized candidate rankings.
Implement weighted latent-factor learning efficiently.
Predict the next preference from ordered interaction histories.
Combine retrieval and a language model to support interactive recommendation.
The group design challenge asks teams of up to two students to develop and evaluate a recommender on a public dataset. The deliverable combines a conference-style report, configurable code, empirical results, presentation, and peer review.
Assessment at a glance
The course is research-facing and implementation-heavy. Students should arrive with machine-learning and programming experience and leave able to reason about both recommender algorithms and the systems that carry them into production.