Simplify pandas-iris starter example

[SajidAlamQB]

Motivation and Context

Currently, the pandas-iris starter and documentation contain four nodes split across two modular pipelines, and the hand-crafted linear regression data science algorithm is quite complicated. I think this example should be simpler than spaceflights, not more complicated.

Opted to convert the examples to linear regression to find the coefficient of determination between sepal length and other features.

Implemented a simple 1-nearest neighbour classifier and calculated its accuracy.

related to: kedro-org/kedro#1393

Development Notes

Remade the pandas-iris starter through kedro new on Kedro 0.18.0

How has this been tested?

Checklist

• Opened this PR as a 'Draft Pull Request' if it is work-in-progress
• Assigned myself to the PR
• Added tests to cover my changes

[antonymilne]

At a glance this looks great, huge improvement on what we had before! Would be good to get @noklam to give another independent opinion on it I think.

Given that there's only one pipeline, I wonder whether we should make this even simpler by removing the data_science pipeline folder altogether. So it would look like this:

{{ cookiecutter.repo_name }}/src
├── {{ cookiecutter.python_package }}
│   ├── hooks.py
│   ├── __init__.py
│   ├── __main__.py
│   ├── nodes.py
│   ├── pipeline.py
│   ├── pipeline_registry.py
│   ├── README.md
│   └── settings.py

and similarly for un-nesting conf and tests.

This is actually what the kedro template used to look like before modular pipelines. Advantage of it is that this is meant to be a very basic introductory example and it would make the project structure easier to understand. We already have the spaceflights tutorial which goes to the next step with multiple pipelines. Curious what others think of this idea.

More detailed comments to follow, but just a few initial things that sprang out at an initial scan through:

• did you generate this with kedro 0.18? Because I don't think hooks.py should be there any more
• I think you'll need to pull from main and resolve some conflicts - several of the changes in the diff I think are probably already in main after we released 0.18
• you'll need to make similar changes to the pyspark-iris starter. Probably easiest to leave that for when we've finalised pandas-iris though, since it will be almost exactly the same
• was deleting .gitignore intentional? See Missing .gitignore when using starters #56. Whatever .gitignore file is generated by kedro new should be in the starters too
• y_train and y_test will be pd.Series rather than pd.DataFrame. A series is what you get if you take a single column of a dataframe

[SajidAlamQB]

I remade the starter with kedro 0.18.0 hooks.py may have slipped in while I was copying some stuff over, I have now removed it. Also added .gitignore back in. Once we are happy with pandas-iris I will repeat it for pyspark-iris (may do this as a separate PR to prevent this one from getting too bloated). Also note we have to update the docs on Kedro for Iris dataset example project.

[merelcht]

I really like how much simpler the starter looks already! 🤩 There are still some mentions to data_science and ds, so don't forget to update those.

[antonymilne]

As per @noklam's comments that there's not really any model training going on, I wonder if we should go for the following terminology for function names, etc.:

• split_data
• make_predictions
• report_accuracy

[AhdraMeraliQB]

Nicely done!

[antonymilne]

This is awesome, great work! Lots of small suggestions to do before merging but nothing big.