🎥 Video: Spotify Recommendations

Recommendations Everywhere

CP Pilot Course

CP4285 is a pilot in the CP428x Emerging Topics in Computer Science series.

• We test and improve content, assessment, labs, tutorials, and delivery for future cohorts.
• Students help by self-learning, building labs and tutorials, slide notes for more comprehensive study.
• And documenting processes, gaps, and improvements.

🎰 Bandit Time: Objectives

📶 AI Voice Mode

We are using AI Voice Mode experimentally to investigate the ethics of recommenders: how an interactive system explains trade-offs, frames uncertainty, and responds when we question who receives attention.

It is a conversation partner for inquiry, not an authority or a production recommender. Listen for what it assumes, what it leaves out, and how its wording could shape our judgement.

👥 Find the Recommender

📶 Ethics Thread

Recommendation systems allocate attention, visibility, and opportunity.

Historical data and engagement objectives can reproduce or amplify bias.

Personalisation raises questions about privacy, transparency, exposure, fairness, and stakeholder impact.

🛣️ Roadmap: Course Learning Outcomes

By the end of CP4285, you should be able to:

Course Library

A course library for going deeper

Weekly Topics

Weekly Topics (cont.)

Class Meetings

• Time: Every Tue, 10:00–12:00 SGT
• Venue: Seminar Room 12, COM3-01-21

Lectures will be recorded where technology permits — recordings are for revision only, not a substitute for attendance

There are no tutorial sessions for this course — We’ll be building the materials for the next cohort.

Grading Overview

Class Participation (10%)

1. In-lecture activities
2. Pre-lecture activities
3. Pre-Flight Survey and Midterm Survey
   • Pre-Flight Survey (due Week 02; Mon, 17 Aug 2026 23:59 SGT)
   • Midterm Survey (due Week 08; Mon, 5 Oct 2026 23:59 SGT)
   Complete these on time — they are free marks

Essays (20%)

Two individual take-home essays, each worth 10%

• Essay 1: Due Week 03; Mon, 24 Aug 2026 23:59 SGT
• Essay 2: Due Week 08; Mon 5 Oct 2026 23:59 SGT

Short written analyses of recommender-system methods, evaluation choices, and social implications (8%). Randomised peer review counts towards the remaining 2%.

AI tools are permitted as a resource — but an AI declaration is mandatory for each submission

Project (30%)

• Groups have 2–4 students
• Define a clear problem anchored on a selected recommendation system dataset

Milestones:

• Week 03: Project Mini-team Declaration — Mon, 24 Aug 2026 23:59 SGT
  • Submit a self-formed team or indicate that you need instructor placement
  • Min has the final decision on group formation (may override student-formed groups)
• Week 07: Project Design Critique Workshop
• Week 13: Final Project Presentation at 29th STePS: Wed, 11 Nov
  • Teams presenting at 29th STePS do not submit a project report; other teams submit a report instead, due same Wed, 11 Nov 2026 23:59 SGT

Recommendation Dataset Catalogue

Click the screenshot to open the live dataset catalogue.

RUCAIBox. (n.d.). RecSysDatasets [GitHub repository]. github.com/RUCAIBox/RecSysDatasets

Quizzes (10%)

• Quizzes are jointly created by peers and instructors
• Peers contribute candidate questions, answers, and plausible distractors from their self-learning
  • Min will assign peers randomly to a week and section responsible.
  • Min will curate, verify, and finalise questions so that the assessments remains fair, accurate, and aligned with the course learning outcomes
• Combined worth 10% of total marks
• AI tools must not be used during in-class quizzes.

Authority Chain

Our slides will be the most authoritative source of truth.

While the we will try to get the slides synced with the website or Canvas, use information introduced in lecture as the definitive source.

There will likely be errors in the lecture, so check any doubts or questions with Min.

Unless a slide or Canvas announcement specifies otherwise, assignment deadlines are Mondays, 23:59 SGT.

The User–Item Matrix

From a Score to a Useful List

Signals and Usefulness

Four Classical Starting Points

When the Matrix Has No Answer

🎰 Bandit Time: What should this system optimise?

🃏 Key Users to Consider

We’ll use card suits as User Archetypes (treatment lenses) to reason about the outcomes. These are not demographic labels or permanent personalities. They are situational lenses for asking what a recommender should protect or provide.

🃏 Four Different Safeguards

“Good recommendation” changes when the risk, task, and user situation change.

Rating Types

The Long Tail

From a Sparse Matrix to a Neighborhood

Similarity I: Pearson Correlation

\[ \operatorname{Pearson}(u,v)= \frac{\sum_{\color{orange}{k\in I_u\cap I_v}}\!\bigl(\color{teal}{r_{uk}-\mu_u}\bigr)\bigl(\color{teal}{r_{vk}-\mu_v}\bigr)} {\color{navy}{\sqrt{\sum_{\color{orange}{k\in I_u\cap I_v}}\!\bigl(r_{uk}-\mu_u\bigr)^2}\;\sqrt{\sum_{\color{orange}{k\in I_u\cap I_v}}\!\bigl(r_{vk}-\mu_v\bigr)^2}}} \]

Worked Example I: Pearson Correlation

User–user example: users 1 and 6

Similarity II: Raw Cosine

\[ \operatorname{raw\text{-}cosine}(u,v)= \frac{\sum_{\color{orange}{k\in I_u\cap I_v}} \color{teal}{r_{uk}\,r_{vk}}} {\color{navy}{\sqrt{\sum_{\color{orange}{k\in I_u\cap I_v}}r_{uk}^{2}}\;\sqrt{\sum_{\color{orange}{k\in I_u\cap I_v}}r_{vk}^{2}}}} \]

Worked Example II: Raw Cosine

Users 1 and 6

Similarity III: Adjusted Cosine for Items

\[ \operatorname{adjusted\text{-}cosine}(i,j)= \frac{\sum_{\color{orange}{u\in U_i\cap U_j}} \color{teal}{s_{ui}\,s_{uj}}} {\color{navy}{\sqrt{\sum_{\color{orange}{u\in U_i\cap U_j}}s_{ui}^{2}}\;\sqrt{\sum_{\color{orange}{u\in U_i\cap U_j}}s_{uj}^{2}}}} \qquad \color{teal}{s_{ui}=r_{ui}-\mu_u} \]

Worked Example III: Item–Item

Items \(i_1\) and \(i_5\) after user mean-centring

From Item Neighbours to a Prediction

Similarity IV: Significance Weighting

\[ \operatorname{discounted\text{-}sim}(u,v)= \color{navy}{\operatorname{sim}(u,v)}\;\cdot\; \frac{\min\!\left\{\color{orange}{|I_u\cap I_v|},\,\color{teal}{\beta}\right\}}{\color{teal}{\beta}} \]

Worked Example IV: Significance Weighting

User–user example: users 1 and 6

Step 1 — start from the user-user Pearson score

\[\operatorname{Pearson}(u_1,u_6)=-0.189\]

User–User KNN in RecBole

from recbole.quick_start import run_recbole

run_recbole(
    model="ItemKNN",
    dataset="ml-100k",
    config_dict={
        "knn_method": "user",
        "k": 50,
        "shrink": 100,
    },
)

user = dataset.token2id["user_id"]["196"]
item = dataset.token2id["item_id"]["242"]
query = Interaction({
    "user_id": torch.tensor([user]),
    "item_id": torch.tensor([item]),
}).to(config["device"])
score = model.predict(query)

🔑 Key Point: Similarity Is a Modelling Choice

Neighbourhood Choices

From Neighbours to a Prediction

\[ \widehat{r}_{uj}=\color{teal}{\mu_u}+ \frac{\sum_{\color{orange}{v\in P_u(j)}} \color{navy}{\operatorname{sim}(u,v)}\,\bigl(\color{teal}{r_{vj}-\mu_v}\bigr)} {\sum_{\color{orange}{v\in P_u(j)}}\left|\color{navy}{\operatorname{sim}(u,v)}\right|} \]

Practicality: Making Neighbourhoods Work

Section 05 Summary: Neighbourhood CF

🔑 Summary