search

microbit-foundation / cctd-ml-machine

GitHub Location for Aarhus University's Center for Computational Thinking and Design's ML-Machine application.
https://ml-machine.org/
MIT License
5 stars 11 forks source link

Too few samples for effective KNN model #461

Closed r59q closed 6 months ago

r59q commented 8 months ago

Currently the number of neighbours used to decide the outcome of a KNN model is located in StaticConfiguration, and is set to 10 atm. But if there are only 10 samples, then the outcome will always be 50/50, since the K nearest samples are the entire dataset.

Should we compute an appropriate value of K at runtime? Or is this something we want to make appearent / configurable when training the model

r59q commented 6 months ago

We go with 3 for now, which is the minimum number of gestures. Therefore the issue will be closed, but probably should be opened up for configuration in later versions