e2.2 dash

Our proposal to Schneidel's challenge for Datathon 2025.

An easy to use dashboard to understand the viability of a potential sale based on a model trained on a set of parameters given by Schneidel Electric for FME's 2025 Datathon.

Started on Colab in order to find the best model and parameters, fit it and export it in order to use it to predict data points on the dashboard (the code for that is attached in colab_full_pipeline). We also started experimenting with SHAP explainability there. After that, we made the actual dashboard using streamlit, that runs the model in order to get a prediction and displays it in a nice way. We also built a page to compare what would happen if we changed some attributes in a record, which also calls our model and explainer, and then we print in a nice way what changed, and how the explanation changed. This is fully dynamic and done as the user is changing the values.

What we learned

We learned a lot about different models and how to optimize their parameters in order to get a nice F1 score, which we landed on 0.837 although we got to 0.9 with a very complex combination.

What's next for e2.2 dash

We didn't have time to add a page to interactively predict observations from completely new data, so that'd be a very simple and very powerful upgrade to the dashboard.

Try adding more features and do other kinds of preprocessing in order to improve the F1 score, as we've experimentally found that it's possible to achieve higher ones, as we were able to achieve 0.9, but in during the time of the datathon required a too slow and complex model.

Built With

• python
• streamlit