Models can handle missing values out-of-the-box more effectively than imputation methods. An empirical proof — Missing values are very common in real datasets. Over time, many methods have been proposed to deal with this issue. Usually, they consist either in removing data that contain missing values or in imputing them with some techniques. In this article, I will test a third alternative: Doing nothing.

How to calculate the Error Contribution of the features of a classifier, with the goal of understanding and improving the model — Feature importance is the most common tool for explaining a machine learning model. It is so popular that many data scientists end up believing that feature importance equals feature goodness. It is not so. When a feature is important, it simply means that the model found it useful in the…

“Feature Importance” is not enough. You also need to look at “Error Contribution” if you want to know which features are beneficial for your model. — “Important” and “Good” Are Not Synonyms The concept of “feature importance” is widely used in machine learning as the most basic type of model explainability. For example, it is used in Recursive Feature Elimination (RFE), to iteratively drop the least important feature of the model.

An in-depth explanation of “Thompson Sampling”, a more efficient alternative to A/B testing for online learning — Imagine you have two ads to choose from: the red one and the blue one. Of course, you would like to show your users the ad with the highest click rate.

What’s the cost of being wrong (and the gain of being right) according to different metrics — If you google “What metric is better, accuracy or F1-score?” you will probably find an answer like: “If your dataset is unbalanced, forget accuracy and go for F1-score.” If you look for explanations, you will probably find a vague reference to the “different costs” associated with false negatives and false…

How to find the best performance estimation approach for time-series forecasts among 12 strategies proposed in the literature. With Python code. — Data scientists are always focused on finding the best model for their dataset. However, they often overlook the importance of choosing the best performance estimation method. This is unfortunate because estimating a MAE of, say, 5.2 on your test set is useless if you are not reasonably sure that the…

The Elbow method is the most popular way to find the number of clusters for k-means. But there are much better alternatives — I asked ChatGPT to advise me on how to choose the right number of clusters for k-means. This was the answer:

Testing a generalist model on a ridiculously hard problem — In a previous article, I tried to debunk the somewhat diffuse idea that a bunch of models (each specialized on a subset of a dataset) should perform better than a single model. To do that, I took a portion of a dataset (e.g. only the American customers) and trained a…

Comparing the effectiveness of training several ML models specialized on different groups, versus training one unique model for all the data — I recently heard a company declaring: “We have 60 churn models in production.” I asked them why so many. They replied that they own 5 brands operating in 12 countries and, since they wanted to develop one model for each combination of brand and country, that amounts to 60 models.