If you’re working with machine learning models, you know how tough it can be to deal with drift. That’s when your model’s performance starts to degrade over time due to changes in your data. It’s like trying to hit a moving target. But what if I told you there’s a new kid on the block that’s showing some serious promise? Enter PKBoost, a gradient boosting algorithm that’s specifically designed to handle drift. I recently came across a Reddit post by Federal_Ad1812 that caught my eye. They’ve been working on PKBoost and have created a full drift benchmark report that pits it against LightGBM and XGBoost. The results are pretty impressive. PKBoost outperforms both LightGBM and XGBoost in 16 different drift scenarios, with some notable wins in heavy noise, sign flip (adversarial), temporal decay, and extreme covariate scenarios. In fact, PKBoost holds a PR-AUC of > 0.74 even under extreme distortion, while others degrade below 0.23. The takeaway here is that PKBoost is a strong contender when it comes to handling drift. If you’re interested in learning more, I’d recommend checking out the full drift benchmark report and the PKBoost GitHub repo. And if you’re feeling adventurous, you can even try out the documentation website. Who knows? You might just find your new go-to model for dealing with drift.