Poor performance of LightGBM higher versions

[DSOTM-pf]

Description

Hi. I used LGBMClassifier in my program but I noticed that its runtime increased when adopting higher versions of lightgbm.

As shown in the following experiment results, this model performs best if we use lightgbm<=2.3.1.

My question is that why does such a library version perform better? Are there any issues in newer versions?

The detailed information is as follows:

Versions	Runtimes
3.2.1	179s
3.0.0	101s
2.3.1	86s
2.1.2	83s

Reproducible example

Here are the code, train dataset, and test dataset:
download code and datasets

Environment info

Command(s) you used to install LightGBM

pip install lightgbm== "versions"

[guolinke]

@wowhhh can you provide a reproductive example?