data from the trenches

Detecting drift is critical in the monitoring of deployed models. So, how do we efficiently detect data drift ?

We review and experiment the various categorical encoding strategies of xgboost, lightgbm and catboost.

We present the inner workings of XGBoost, Lightgbm, HistGradBoosting and compare their performance. Does one stand out ?

Spoiler alert: on average, all algorithms except grid search produce similarly performant models.

When Machine Learning models are not relevant anymore, it might be due to underlying data drift. Here, we introduce and review data drift.

Studying Hyperparameter Importance to Speed Up Optimization

Rounding operations can lead Hyperband not to use 7% of the available budget. We propose a method that reduces unused budget to 3%.

An in-depth review of unfairness causes and their root in data.

How to make the most of your unlabeled data ? Can traditional semi-supervised techniques boost performance ?

Evaluation of the Learning Rate Finder

Lessons learned from reproducing “Diverse Mini-Batch Active Learning”, a strategy mixing uncertainty and diversity techniques.

Introduction to Active Learning through a quick benchmark of major Python packages: modAL, libact, and alipy.