Suggested work for this block

2 hours lectures, 0.5 hour watching videos, 1 hour reading from book, 4 hours self-study, 2 hours computer lab session, 2 hours paper and pencil tutorial, 4 hours getting started with work for mini-project. Total 15 hours. The coming three weeks will now have to include some lab work and starting up the mini-project.

Before you start...

We prepared a self-assessment quiz with true-false statements from previous exams about important concepts of blocks 1 and 2. The quizz does not count towards the final grades and can be repeated multiple times. Note also that it does not cover everything that you should know about the topics so far.

Mini-project

The mini-project should start to be part of your working week! To get started we recommend you:

Familiarize yourself with and make use of the different tools in your group section on Studium
If you haven't done so, have an initial group meeting and start to divide up the work. You now have learnt quite a few of the methods needed.
Discuss how you want to work on the project and agree on some working rules.

If you have questions regarding the mini-project, write in the discussion forum. Also feel free to ask questions after the tutorials.

Lecture F5: Generalization and bias-variance decomposition

This video explains the basic challenge of overfitting.

When learning a model we may trade-off:

Bias: The inability of a method to describe the true patterns in the classification or regression problem. Low model complexity.

Variance: Sensitivity to random effects (noise) in the training data. High model complexity.

Lecture F5 on Thursday Nov 12th @ 10.15
Recommended reading: Ch. 4.1 to 4.4

Remark 1: For k-NN classifier with M classes, we need k odd if M is even (and vice versa).

Remark 2: The bias-variance decomposition is an important tool to understand the source of your errors, but you cannot readily estimate the bias because it requires the unknown function f0(x)

Slides for the lecture

This video introduces k-fold cross-validation

Tutorial L5 (pen & paper)

Spend about two hours attempting the questions before the tutorial.

The exercise sheet is available here.

Recommended problems: 5.1, 5.2, 5.3, 5.4
Additional problems: 5.5, 5.6

Feel free to ask questions in the discussion section.

Tutorial L6 (computer)

Spend about two hours attempting the questions before the tutorial.

The computer labs are available here:

Recommended problems: 6.1, 6.2, 6.3
Additional problems: 6.4

Suggested solutions are available here.

Also feel free to ask questions in the discussion section.

Want to learn more?

The Netflix data challenge

First issued in 2006 and offered a million dollars to the team who could best improve Netflix's recommendation system.

Screenshot 2020-10-30 at 09.47.23.png

The dataset is available on Kaggle. Netflix provided a training data set of 100,480,507 ratings that 480,189 users gave to 17,770 movies, divided in to various training, test and competition sets. Our Spotify challenge is like a mini-mini-mini-version of this challenge.

A scientific description of this competition, and how early improvments were made by different teams, can be found here.

The challenge here was to predict ratings. But it was Youtube that realised that actually ratings aren't as important as watch time, which reveals what you really want, not what you say by making a rating! Their algorithm, trained using a neural network, is described here.