Homework 3 ✍🏻

Back to the ML Course - 2026 Spring (AI5100/EC4224/AI4100/EC5402)

Due: Monday, May 25, at 11:59 PM
Language: English (spoken explanation, on-screen text, outline)
Question: Ed discussion

Objective

Record a 15–20 minute YouTube video where you talk freely over your slides about a topic in Survival Analysis (extending ISLP Ch.11 — optionally combined with Ch.10 Deep Learning). Go one step beyond the lecture: survey some related material as if you were preparing for a research project.

Topic Options

Pick one. Your topic should extend survival analysis beyond the Cox / Kaplan–Meier baseline covered in lecture — typically with an applied angle (churn prediction, subscription/retention, time-to-event in healthcare, reliability, or any everyday situation where survival analysis fits). Deep learning is one strong direction, but classical / parametric extensions and tree- or boosting-based survival methods are equally welcome.

• A. Methodology paper. Present one paper (or method) you would actually read. Walk us through what problem it solves, what the main idea is, how the method works, what the experiments showed, and what you think of it. Some directions to consider:
  • Deep survival: NSA, DeepHit, DRSA, DeepSurv, WTTE-RNN, RNN-based survival models, DL for churn.
  • Parametric / AFT: Weibull accelerated failure time models, or improvements to the Kaplan–Meier estimator itself.
  • Beyond standard Cox: time-varying survival regression when covariates change over follow-up.
  • Tree- / boosting-based: Gradient Boosting Survival Analysis (e.g., scikit-survival).
  • Main package for the course: lifelines — start here for fitting and comparing classical models; reach for scikit-survival or PyTorch-based libraries when your method needs them.
  • For orientation across the field: Machine Learning for Survival Analysis: A Survey. Hyperconnect’s three-part blog series is a friendly intro (Korean — read for understanding): part 1, part 2, part 3.
• B. Competition / dataset / real-world scenario. Take a look around at related competitions (e.g., KKBox Churn Prediction), public datasets, or a specific applied domain, and share what you found interesting. You don’t need to actually compete — just introduce a few that caught your eye and talk about what the tasks are like, what the data looks like, and what kinds of approaches seem to work. A domain deep-dive also counts: pick one real setting (e.g., Shopify merchant churn, telecom subscription, hospital readmission, hardware reliability), read an in-depth article on it, and share the business insights you took away.

Blending options (e.g., a paper + its dataset) is fine if it tells a better story.

Deliverables

Submit a short outline PDF to Gradescope under “HW3” containing: name & student ID, YouTube link, chosen option (A/B) + one-line topic, paper/dataset reference if any, video timestamps (copy from YouTube), citations, collaborators.

You don’t have to prepare slides. Show whatever fits your topic on screen — slides if you want, but it’s also fine to keep the paper open and walk through it, or to step through the data / a notebook as you talk. There is no separate problem-solving section; the video carries the talk.

Video Requirements

• Length: 15–20 minutes.
• Recording: Zoom (or equivalent) with screen share, your face visible throughout. Suggested guides:
  • PowerPoint as virtual background
  • Put your face in front of your slides in Zoom
  • 얼굴 넣고 녹화하는 방법 (Korean)
• YouTube upload:
  • Privacy: must be Public or Unlisted (anyone with the link can view) so the staffs can watch it.
    • Uploaded as Private, then your grade will be deducted + we’ll ask you to change it. If it’s still Private after we ask, it’s treated as no submission.
  • Timestamps: required in the video description (MM:SS Topic, e.g. 00:00 Intro, 02:30 DeepSurv setup, ...). Missing timestamps → grade deduction.
    • How to add timestamps
• Example videos (for a feel of the format — not a template to copy):
  • Example 1 (OK - Overlay with paper, without slides)
  • Example 2 (OK - With your slides)
  • Example 3 (OK - With Colab & Dataset)
  • Example 4 (BAD - Korean, no face on screen: your submission should be in English with face visible)
• On-screen material: whatever fits — slides, the paper itself, a notebook, the dataset page. If you do prepare slides, they should be your own work; cite figures you reuse from the textbook, lectures, or papers.

Grading

• Topic & depth — clear tie to survival analysis (deep, classical, or boosting/tree-based), goes beyond lecture, content is correct and not surface-level.
• Delivery & timing — well-structured and clear, within 15–20 min, video/audio/on-screen content legible, face visible. (Videos outside this range will lose points)
• Language & citations — everything in English (talk, on-screen text, outline), figures and sources properly cited.