Question : Why are regression tree ensemble predictions inconsistent with the ones from microsoft/LightGBM

Hello,

I inherited from a stored tree model trained using the python API from microsoft/LightGBM, which from my understanding uses the same underlying C++ code as this one.

Since the saved .txt file has the same format, I can load the model using the pyspark API provided in this repository. However the computed predictions are often slightly - sometimes completly - off when comparing to the ones I get using the original non-spark API.

This is a little odd to me as training aside, I would expect the evaluation of the same tree ensemble to give roughly identical results regardless of implementation. I'm guessing this stems from a different way of evaluating the model maybe due to distributed computing ?

While the interest of such a compatibility might be debatable, I would like to understand where lies the difference. Especially, if it is because of parallel computing, or if a certain set of parameters can help me achieve the same results.

Can you give me pointers so I can learn a little more about what is going on ?

Thanks !