ML Module Week 3 Topic 2: KNN Classifier

[EricThomson]

Intro to knn both what it is, and hands on with scikit-learn. Break data into test/train data with scikit-learn, and evaluate performance.

This is meant to be a fairly simple intro to KNN using iris dataset, and evaluation using evaluation metrics from the evaluation intro (which is in #79 )

[EricThomson]

This one is relatively easy on classifier explanation (knn),

Datasets we will be using are built-in with scikit-learn:
iris dataset only (no longer planning to use digits let's just keep it simple).

from sklearn.datasets import load_iris
iris = load_iris(as_frame=True)  #as frame if you want to load as df

Then, run and explain knn on the dataset. One good page on knn: https://www.ibm.com/think/topics/knn

Video that could be good for lesson:
https://www.youtube.com/watch?v=b6uHw7QW_n4

Note for KNN with iris dataset it is really important to scale the data (from sklearn.preprocessing import StandardScaler), otherwise, dimensions that are slightly different can dominate the KNN distance calculations: the different measurements live on very different ranges. I'll provide example that demos this...

To evaluate initially, we'll want to use scikit-learn confusion matrix, classifier report, (accuracy and F1).

[EricThomson]

For evaluating classifier for initial datasets (which are class balanced), accuracy will be fine, and showing how to visualize confusion matrix will be fine. But we will want to motivate precision and recall using highly imbalanced data, so we will want to create imbalanced data by doing 3 versus non-3 with the digits dataset, showing how you can get high accuracy just by guessing 'not 3' on every digit (this will give you 90% accuracy). This is why you need precision/recall, and lets us explain the different cells in confusion matrix. And explain the full output of classifier report from scikit learn (precision, recall, F1).

The main focus should be on intuition here, not the deep mathematical weeds.

This article is pretty good on this, but we should chat more to go over these concepts:
https://medium.com/@techwithpraisejames/how-to-calculate-confusion-matrix-in-python-using-scikit-learn-e51368716cbe

Update this is now a separate PR so don't worry about separately explaining classifier metrics it's now a separate lesson so no pressure there! #79