# Learning: the hows and whys of machine learning

Liam Wiltshire
https://liam-wiltshire.github.io/talks/?talk=machinelearning&conference=phpuk
https://joind.in/event/php-uk-conference-2019/learning-the-hows-and-whys-of-machine-learning

## Overivew

Charge backs

## Supervised learning
Training data
Learning functions
Categorisation / Classification
Regression - Where do we sit on a line

## Naive Bayes classifier
Standardise words
- Un pluralise
- Un gender
- Un tense
- etc

More data == better

## Tokenisation
https://en.wikipedia.org/wiki/Benford%27s_law
https://php-ml.readthedocs.io

Unique tokens for each unique context

## Imbalanced data
One category has more database
99% data not charge back
Just being accurate, not very helpful
    - Started by flagging 100% as fine.
    - Need to collect more data, change methods, resample data

## Understand data
- context
- Common data vs specific data
- Continuous vs discrete data

## KNN
K Nearest Number
https://en.wikipedia.org/wiki/K-nearest_neighbors_algorithm
    - Distances
    - less sensitive to imbalance
    - Keep K odd (no draws)

## Handling nominal data

Binary
- Increase amounts of dimensions
- normalisation required
- equal scales

## Contextless data is meaningless
Is it normal?

## Next to try
Weighting
Different dimensions
Change K value (was 3NN)
Remove outliers
Diff distance function
weighted distance




# Useful links
https://en.wikipedia.org/wiki/Benford%27s_law
https://php-ml.readthedocs.io
https://liam-wiltshire.github.io/talks/?talk=machinelearning&conference=phpuk
https://joind.in/event/php-uk-conference-2019/learning-the-hows-and-whys-of-machine-learning
https://en.wikipedia.org/wiki/K-nearest_neighbors_algorithm