Your Data Teacher Podcast

In this episode, I'll talk about tuning the threshold in binary classification tasks. The usual value for the threshold is 0.5, but it's useful to optimize it in order to make the model fit our needs. I talk about optimizing according to the ROC curve and maximizing the balanced accuracy.  

In this episode, I'm going to talk about ensemble models, particularly bagging and boosting. Bagging is very useful for reducing variance, boosting is used for reducing bias. The most common bagging algorithm is Random Forest, the most common boosting algorithm is Gradient Boosting, whose most common implementations are XGBoost, LightGBM and CatBoost.

In this episode, I talk about accuracy, precision and recall. We're going to focus on what they are and when to use them in machine learning projects.

Today we're going to talk about how we can explain neural networks. Neural networks are like black boxes that hide the way they model and represent data. That's why explaining them is very difficult. A very powerful approach is called SHAP. Using this method, we can calculate the impact of a feature according to a given model independently of the type of model we're using. It's very useful for black boxes like neural networks.

Today we're going to talk about the standard error on proportions. In data science, it's very important to calculate the standard error on every estimate we calculate in order to see if finite-size effects are lowering the precision too much and in order to compare two different measurement results with each other.