The basic intuition behind the perfect score in this competition is to use classification.

Why? If you are to check the label distribution, you will notice that they are all distinct values.

So, the part toward achieving a perfect score includes

1) Use the regression model but do classification-based postprocessing to move the predictions to the closest known values.

2)Set up a pure classification model.

Feature Engineering,

This competition involves interference in the communication channels to the best of my knowledge. So, you can try to get interaction between any ap-sta combination and any other available in the txop.

For instance, let's assume ap2-sta2, ap5_sta5, ap9-sta9 are available in a txop. One can get something like

Groups1:::

rssi_ap2_sta2 / rssi_ap2_sta5 , rssi_ap2_sta2 / rssi_ap2_sta9 etc...

Groups2:::

rssi_ap2_sta2 - rssi_ap2_sta5 , rssi_ap2_sta2 - rssi_ap2_sta9 etc...

All these will be for the ap2_sta2 within the txop of the specific deployment

First order statistics of the Groups ensure equal number of features per group.

Such features can be used to train lgbm (in my case)....

Enjoy

Apology on the typos.... smiles