Machine Learning

In my first machine learning project, I built predictors for the Iris Dataset in Python, combining decision trees, entropy and a neural network, with my ID3 decision tree frequently averaging over 80% accuracy.

RoleDeveloper · UTS

TeamSolo

TimelineAug – Nov 2022

I ownedAlgorithm Design, Coding & Testing

The sixty-second version

Problem

An intentionally vague brief: design and code at least one machine learning algorithm from scratch, to build one or more predictors using the Iris Dataset.

What I did

I began with an ID3 decision tree, later adding a general decision tree, entropy and an Extension Type B neural network on top of it, with a confusion matrix added in the final stages.

Result

The ID3 Decision Tree frequently averaged over 80% accuracy, whilst the Extension Type B Neural Network was performing around the 70% mark, and also included a confusion matrix.

>80%: accuracy, frequently averaged by the ID3 Decision Tree

~70%: accuracy from the Extension Type B Neural Network

↓ More details below if you want the reasoning.

§ 01

The Problem

The brief was intentionally vague, beyond naming the required dataset and the options for predictors and algorithms to use.

"The level of comprehension and tools used went above and beyond the requirements of this task."

What I couldn't change

01 The Iris Dataset, as named in the brief.

02 At least one machine learning algorithm had to be designed and coded from scratch, in Python.

03 My first time coding any sort of machine learning program and algorithms.

§ 02 — Decisions & trade-offs

From the beginning, my intention was to use multiple methods simultaneously rather than just one, as a single predictor would not be capable of achieving the level of accuracy I was seeking.

With that many elements involved, one decision shaped how the project was built.

Decision 01

Testing early and often

Chose

Testing frequently throughout the project, rather than just at the end.

Gave up

More effort up-front.

Why

With the size, scale and number of different elements involved, it was essential to catch errors before they compounded.

§ 03

Impact & what I'd change

A first machine learning project that integrated all of the available algorithm options.

Looking back

The predictors performed well, especially considering it was my first instance of making them. Testing early and often was a significant factor in the final success of the project. If I did it again, I would record the accuracy after each method was added, to show how much each one contributed.

— Noah

Code Individual Python Project Iris Dataset

Next case

Case № 01 →

KPMG CREATES - Consultant (Design & AI Specialised)

All projects →

Machine Learning · Scrapbook-styled Portfolio