gtregoat/catboost-experiment

Testing catboost on the titanic dataset

Testing Catboost on the titanic dataset

I've heard that catboost was doing quite well on various benchmarks and that it came with built-in mechanisms to process categorical and text variables. I wanted to put it to the test.

Many of the other algorithms I've used (e.g. logistic regression, but even xgboost, my current go-to) require quite some preprocessing and a fair bit of data exploration. What I find very interesting here is that we could speed up the work greatly and get at least comparable results, if not same / better according to certain benchmarks. There can be a lot of value there, as time saved can be used to do other projects, or have a fast prototype / first version and trigger continuous improvement. Having a first working version has proved to be key in my experience, as it gives all stakeholders of a project an idea of the kind of output they will get, leading to a faster overall project integration as services interacting with the model can be built with real data, and the project can fail fast if it seems overall it won't work.

On Kaggle

https://www.kaggle.com/gtregoat/testing-catboost-on-the-titanic-dataset