changed 2 years ago  
Published Linked with GitHub ddojo/ddojo.github.io/pad_archive/21_datadojo.md

Data Dojo Würzburg 21

DataDojo@Lunch - live

April 2023

  • When: Wednesday, April 12th, 2023 at 11:00am until 12:30pm (90 minutes)
  • Where: CCTB or online (CCTB Seminar Zoom Link)
  • Info: DataDojo Website, Repo

Dataset

Machine Learning Series

We are doing a series of Data Dojos on machine learning. The task is to classify tree species by their traits (e.g. height, stem diameter, geographic location).

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →
Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →
Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →
We use a subset of the recently published database: Tallo

The full dataset contains measurements for almost 500k individual trees from more than 5k species.

In the first dojo of the series, we filtered the full set to 3 species with reasonable overlap (Fagus sylvatica, Pinus pinaster, Quercus ilex). Now we want to try different Machine Learning methods to classify tree species from traits.

In the second dojo we created our first models. A very simple "Majority Vote" model and some K-Nearest-Neighbor (KNN) models with scikit-learn.

In the third dojo we explored the effect of scaling on the performance of the KNN models.

In the fourth dojo we explored Decision Trees as models for classification

In the fifth dojo we used Support Vector Machines as models for classification

In the sixth dojo we used ensemble models, including Ada boosting and random forests.

In the seventh dojo we used imputation methods to also make predictions for cases with missing data.

Session 8 - Neural Networks

Ideas:

  • Hyperparameter tuning (cross validation and grid search)
  • Neural Networks
  • experiment tracking (e.g. mlflow)
  • auto ml (e.g. optuna)
  • inference/deployment (e.g. on hugging face spaces)
  • interpretability (e.g. feature importance, plots, top losses…)
  • try to estimate theortical limit for accuracy (e.g. find closest pairs of data points with different labels)

Collaborative Tools and Workflow

For Notebooks (R, python, julia, js, …) with real time collaboration CoCalc seems to be the best option right now. It worked great the last couple of times so we'll stick to it for now. You need to register an account there (it is free).

Future Suggestions

Add your suggestions to the list and

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →
to the end of a line you are interested in

Data Sets

Tools/Languages

Skills

  • interactive maps
  • dashboards
  • animations

Data Sources

all data types are welcome, including tables, images, videos, sounds, DNA, …

Select a repo