Schedule

Topics:

• Recommendation tasks
• Explicit vs implicit feedback
• Popularity-based recommendation
• Content-based recommendation
• Collaborative filtering

Ethics thread: popularity bias, exposure inequality, and platform incentives.

Learning outcomes:

• Formulate a recommendation task using a user-item matrix, available signals, and either a prediction or top-k ranking objective.
• Compare popularity-based, content-based, collaborative-filtering, and knowledge-based recommendation in terms of their evidence, suitable use cases, and limitations.
• Explain cold-start and sparse-data challenges, and justify an appropriate response.
• Compute and interpret user-user or item-item similarity using appropriate neighborhood-based collaborative-filtering measures.
• Explain how neighborhood selection, similarity weighting, and ranking objectives affect prediction quality, exposure, and long-tail visibility.
• Critique recommendation choices in terms of privacy, fairness, robustness, novelty, and stakeholder impact.

Slides:

Lecture video: Watch the Week 01 lecture recording — requires NUSNet authentication.

Topics:

• Matrix factorization
• User-item embeddings
• Bias models
• Ranking objectives, including Bayesian Personalized Ranking

Ethics thread: bias encoded in historical interactions, representation, and interpretability.

Learning outcomes:

• Explain latent-factor models.
• Train embedding-based recommenders.
• Compare prediction and ranking objectives.
• Discuss risks of learning from historical behavior.

Slides:

Topics:

• Offline evaluation
• Precision@K
• Recall@K
• MAP
• NDCG
• Business metrics

Ethics thread: metrics as value choices, and accuracy versus user welfare.

Learning outcomes:

• Design evaluation protocols.
• Compute ranking metrics.
• Critique metric selection.
• Explain limitations of offline evaluation.

Slides:

Topics:

• Neural collaborative filtering
• Deep ranking models
• Representation learning

Ethics thread: explainability versus performance, and transparency concerns.

Learning outcomes:

• Build neural recommenders.
• Compare neural and latent-factor approaches.
• Analyze explainability challenges.
• Assess trade-offs between complexity and transparency.

Slides:

Topics:

• User sequences
• GRU4Rec
• SASRec
• Transformer recommenders

Ethics thread: engagement optimization and behavioral manipulation risks.

Learning outcomes:

• Model temporal preferences.
• Implement next-item prediction.
• Compare static and dynamic representations.
• Critically assess engagement-driven objectives.

Slides:

Topics:

• Candidate generation
• Retrieval
• Ranking
• Re-ranking
• Two-tower architectures

Ethics thread: visibility allocation and stakeholder impacts.

Learning outcomes:

• Explain industrial recommendation pipelines.
• Design retrieval-ranking architectures.
• Analyze scalability trade-offs.
• Evaluate how ranking affects different stakeholders.

Slides:

Student deliverables:

• Application domain
• Dataset
• User problem
• Recommendation objective
• Baselines
• Evaluation plan
• Ethical risks

Ethics thread: feasibility, bias, evaluation design, and project scope.

Peer critique themes:

• Is the recommendation task well-defined?
• Are the evaluation metrics appropriate?
• What biases may emerge?
• Is the problem feasible?

Learning outcomes:

• Defend recommendation-system designs.
• Critique evaluation strategies.
• Identify ethical risks early.
• Refine project scope based on feedback.

Suggested weight: 5-10% participation or milestone grade.

Slides:

Topics:

• Pointwise ranking
• Pairwise ranking
• Listwise ranking
• LambdaRank intuition

Ethics thread: position bias and fair ranking.

Learning outcomes:

• Formulate ranking objectives.
• Compare ranking approaches.
• Analyze position bias.
• Discuss fairness implications of ranking.

Slides:

Topics:

• User-item graphs
• Graph embeddings
• Graph neural networks
• LightGCN

Ethics thread: homophily, echo chambers, and community amplification.

Learning outcomes:

• Represent recommendations as graph problems.
• Explain graph propagation.
• Build graph-based recommenders.
• Analyze risks of graph-driven feedback loops.

Slides:

Topics:

• Diversity
• Novelty
• Serendipity
• Coverage
• Long-term satisfaction

Ethics thread: balancing stakeholder interests.

Learning outcomes:

• Define non-accuracy objectives.
• Measure diversity and novelty.
• Design multi-objective recommenders.
• Justify objective trade-offs.

Slides:

Topics:

• Multi-armed bandits
• Contextual bandits
• Exploration-exploitation
• Feedback loops

Ethics thread: online experimentation and fair exposure.

Learning outcomes:

• Explain exploration strategies.
• Design adaptive recommendation policies.
• Analyze recommendation feedback loops.
• Discuss ethical implications of experimentation.

Slides:

Topics:

• Conversational recommendation
• LLM-enhanced recommendation
• Retrieval-augmented recommendation
• Foundation models
• Causal recommendation
• Future directions

Ethics thread: trust, hallucination, persuasive AI, and governance.

Learning outcomes:

• Explain modern recommendation research directions.
• Evaluate LLM-based recommendation systems.
• Critique emerging approaches.
• Identify open research challenges.

Slides:

Teams selected to present at 29th STePS complete the presentation below and do not submit a project report. Teams not presenting at STePS submit a project report instead.

Required presentation components:

1. Problem formulation
2. Dataset and preprocessing
3. Baselines
4. Proposed recommender
5. Evaluation results
6. Error analysis
7. Ethical analysis
8. Future improvements

Ethics thread: technical, product, and societal considerations.

Learning outcomes:

• Present recommendation-system designs professionally.
• Defend technical decisions.
• Interpret evaluation results critically.
• Integrate technical, product, and societal considerations.

Slides: