Total suggested work for this week

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 are intensive. You have both lab work to do and need to get going with the mini-project.

Before you start...

We prepared a self-assessment quiz with true-false statements from previous exams about important concepts of weeks 2 and 3. The quiz 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 of week 2 and 3.

3.1 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 at the drop-in sessions.

3.2 What are bias and variance?

We have already met some of the challenges of overfitting data in the context of all the models we have looked at so far. This video explains the basic challenge we are dealing with.

In short we have a tradeoff between:

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.

One method we will look at for creating more accurate models is k-fold validation. Here is an introduction to the idea.

3.3 Reading

This week you should set aside one hour for reading the The Supervised Machine Learning book. Read sections 4.1 and 4.2 in conjunction with the lecture.

3.4 Theory behind the methods

Due to unforeseen circumstances there will be no 'live' lecture on Friday 5th of February. A pre-recorded video is here. Please watch this before Monday the 8th of February.

Lecture notes are available here.

3.5 Pen and paper session

Spend about two hours attempting the questions before the tutorial on Monday the 8th of February

The exercise sheet is available here.

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

This will be covered at the tutorial on Monday the 8th of February 8:15-10:00.

Note: This tutorial will be more interactive and we might ask you to turn on your microphone at some point.

Also feel free to ask questions in the discussion section.

David will do a 15 minute summary of the lecture and answer questions at 10:15-10:45 on Monday to replace the missed live lecture on Friday. The drop-in session afterwards will continue to offer help.

3.6 Computer lab session

Spend about two hours attempting the questions before the tutorial on Tuesday the 9th Feb 13:15-15:00

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.

3.7 I love this and want to know more

The Netflix data challenge, started in 2006, offered a million dollars to the team who could best improve their 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.