Application of K-Means & Self-Organising Maps for Features Selection

Kannan (~kannan00)


Description:

Learning algorithms such K-Means and Self-Organizing Maps have existed for ages, however, the application of it is relatively limited when compared to supervised learning. As I work with these models, I realised how effective these learning algorithms can be in addressing the curse of dimensionality. Both these unsupervised learning algorithms are straightforward and can be effective in dimensionality reduction.

In this talk, I will demonstrate how unsupervised learning algorithms like K-Means and SOM can be fusion with supervised learning algorithms by considering 75+ commonly used technical features used in trading such as momentum, volatility, relative strength, and other factors and reduce it to a more meaningful level.

I will also brief about the underlying mathematical framework of these models and will employ these dimensionality reduction methods using K-Means and SOM separately and then compare these two results to arrive at consolidated features set that can be fed into the supervised learning models.

This talk will help you understand how to use K-Means and Self-Organising Maps (SOM) in feature selection and how to interpret the results.

Outline

  1. Overview of K-Means and SOM (5 min)

  2. Phases of the Algorithm (15 min)

    a. Preparing the input data

    b. Identification of features using K-Means Clustering

    c. Applying dynamic time wrapping for SOM input

    d. Identification of features using SOM

    e. Visualisation and graph

    f. Comparison and selection of features set

  3. Fusion of unsupervised and supervised learning algorithms (5 min)

  4. Q&A (5 min)

Who is this Talk for?

  • Data scientists/engineers who work with high dimensional datasets for financial time-series
  • People who want to gain an understanding of dimensionality reduction, its implementation, and execution
  • People who want to gain an understanding of applying unsupervised learning algorithms for feature selection
  • People interested to know how unsupervised learning algorithms can be used in supervised learning algorithms

Prerequisites:

  • A basic understanding of Python programming.
  • A basic understanding of common dimensionality reduction algorithms

Video URL:

https://drive.google.com/file/d/1uGArDLWU8MssfaUASckP4fdV8xks1NHQ/view?usp=sharing

Content URLs:

SOM Repository: https://github.com/JustGlowing/minisom

Speaker Info:

Hi, I'm Kannan Singaravelu. I am a data science specialist at the CQF Program at Fitch Learning. I am a quant professional and fond of scientific computing and machine learning. In my earlier avatar, I was an application specialist at Bloomberg and got early exposure to Bloomberg Quant and Data Science Platform.

Among other things, I have designed and developed comprehensive Python for Derivatives, Python for Quant Finance, and Python Lab modules that are now part of the financial derivatives curriculum in reputed global institutions.

I hold a Bachelor's degree in Mechanical Engineering and a Master's degree in Business Administration and am a CQF alumnus.

Speaker Links:

Find me on

GitHub

LinkedIn

Section: Data Science, Machine Learning and AI
Type: Talks
Target Audience: Intermediate
Last Updated: