# PyData London 2024
###### tags: `aeon-pydata`

## Title

Machine learning with time series data: classification and regression

## Abstract:

Time series machine learning is a rapidly evolving field that finds areas of application in all domains of machine learning and data science. In this talk we will walk attendees through the fields of time series classification (TSC) and time series extrinsic regression (TSER) and show why traditional machine learning approaches for tabular data are usually not the correct approach for temporal data. Using the `aeon` toolkit, a `scikit-learn` compatible open-source project for time series machine learning, we will highlight recent research advances in these fields and demonstrate how to use them on real word case studies.

## Description

This talk will provide an accessible overview to the fields of time series classification (TSC) and time series extrinsic regression (TSER). This is practical talk; alongside an introduction to the latest research into TSC and TSER we will provide demonstrations of how to train and apply these algorithms using the `aeon` toolkit. We will give real world examples and case studies. Participants will come away with an understanding of the basics for TSC and TSER, and an idea of when these algorithms may be useful in practice.

TSC involves fitting a model from a continuous, ordered sequence of real valued observations (a time series) to a discrete response variable. TSER uses a continuous response variable. Time series can be univariate (a single variable observed at each time point) or multivariate (multiple variables observed at each time point). TSC and TSER problems come from a variety of domains, including medical signals such as ECG and EEG; human activity recognition and other motion data; electricity usage; and many more with industrial and academic applications.

In the talk, we will highlight how we learn from time series data using the latest algorithms and why these techniques are preferable to more traditional machine learning approaches. Following this we will give an introduction to the classification and regression modules in `aeon` and demonstrate how these algorithms can be applied to time series data.

This talk is intended for those with basic knowledge of machine learning and should appeal to anyone interested in machine learning from time series. No prior knowledge about machine learning for temporal data is required, but familiarity with the `scikit-learn` package may be helpful. It will focus on practical application and high-level algorithm comprehension rather than maths or deep understandings of specific algorithms and will suggest further reading for those interested. We will provide easy to follow notobooks prior to the talk and all examples will be freely available.

## Organiser notes

N/A

THEY SAY

Talks
Talks have a length of 30 minutes, with additional time for Q&A. A talk proposal is a short description of a talk that is aiming to convince someone to part with 30 minutes of their time, in order to learn about something. A good proposal should disclose:
- The topic (the WHAT) and WHY it is interesting
- The audience to WHOM the talk is addressed
- The TYPE of talk (lots of maths, hands-on, etc) and possibly the tone (light-hearted, informative etc)
- The TAKEAWAY, a.k.a. what will I learn

There are two parts to a proposal:
1. Brief Summary - This informs attendees what the talk is about. Discloses the topic, domain and overall purpose. This is at most a few lines long, and will be printed in the conference programme.
2. Description - This is a self-contained statement that summarises the aspects of the talk. It should be structured and present the objective of the talk, its outline, central thesis and key takeaways. After reading the description, the audience should have an idea of the overall presentation and know what to expect. The description should also make clear what background knowledge is expected from the attendees. Both this and the summary will be included in the talk details online.

While there is no strict template for this, you should make sure that the audience can understand why your talk is relevant for them.

A proposal serves two purposes: 1) convince the Program Committee to accept your proposal, and 2) act as the description of the talk/tutorial on the schedule. If accepted, you will have the opportunity to further edit and clarify the proposal ahead of the schedule’s publication.

The guidelines for proposing a talk versus a tutorial are essentially the same.

Good proposals include:

The topic and why it’s interesting

The audience that will benefit from attending this talk/tutorial

The takeaway for the audience

Any background knowledge needed to understand the talk/tutorial

Approximate time breakdowns of how the material will be covered, e.g. minutes 0-10: X1, minutes 10-15: Y1, etc.

There are two parts to a proposal:

Abstract: The abstract tells attendees what the talk/tutorial is about. It should answer all of the above questions, but it should remain short and to-the-point. This is your chance to pitch attendees on watching your talk/tutorial. If you are submitting a tutorial proposal, the abstract should also specify the requirements for participating and how materials will be distributed (e.g. Github repo, links, etc). If your materials are already hosted publicly somewhere, you can include the link.

Description: The description expands on the abstract and provides background and additional details about your talk/tutorial. It is often helpful for talk descriptions to include an outline of the talk/tutorial. The Program Committee reads each description carefully to determine if a talk/tutorial is a good fit. This is your chance to pitch the Program Committee on accepting your talk/tutorial. The Program Committee does occasionally ask for more information after proposals have been submitted, but it’s better to include relevant details up front and try to anticipate what questions the Committee will have after reading your proposal. For example, if you think the Committee may think you’re trying to cover too much ground, an outline with time estimates could help alleviate that concern.

Additional Talk/Tutorial Proposal Suggestions

Who is your target audience? Think about your target audience in terms of job role (data scientist, engineer, researcher, etc.) and experience level. Being clear about who you are speaking to (and the background knowledge you can expect them to have) is helpful both to you as you prepare your presentation, as well as to the audience considering whether your presentation is a good fit for them to attend.

Clear title: A catchy title can be useful, but don’t overdo it. People should get a rough idea of what your presentation is about just by reading the title. Your proposal and your presentation should be consistent with your title.

Get feedback: Ask friends and colleagues to review your abstract; bonus points if they are your target audience. Take time to tweak your abstract if needed. Additionally, if you are a first-time speaker, please indicate this in your proposal and we will reach out with more information regarding mentorship opportunities.

Common Pitfalls

Here are some common pitfalls that could lead to the proposal not being understood or rejected by the Program Committee:

Overly long proposals: Keep it simple and clear. Good proposals typically provide all the important information within 200 words. This is not a strict limit, just a suggestion to help you stay focused.

Future work: While talking about future work is interesting and could be mentioned in your talk, the core content of the talk should already be shaped, and you should be able to describe it in your proposal. Don’t rely too heavily on future data collection or future prototyping, because things often don’t go as expected.

Sales pitches: We are a community of creators and users of open-source scientific computing tools. You can reference your closed-source product or platform, but the audience will find the talk more interesting if they can try your techniques with the open source PyData stack. Your problem definition, proposed techniques, and business domain are also interesting, but sales pitches are typically rejected.

Repeated talks: We have a strong preference for new talks, and new speakers. If your talk is already available online, it is unlikely to be accepted for the conference.