Combining multiple predictions into one

I’m trying to predict if a player is cheating a FPS-game. The type of cheating I’m trying to catch is “aimbots” meaning a player that is using some external assistance to do the aiming for the player. My data consists of mouse movement information of shape (128,5), (128 timesteps and 5 features) of every kill the player got during the game.

So far, I’ve been using a GRU and feeding the individual kills into the model. The problem is that the model considers the kills independently of each other. If the data for one player, for the entire game, would be in the shape (number of kills, 128, 5), would it be preferable to concatenate the data into a big 2D array of shape 128,5*number of kills or keeping it and treating it like an image where the number of kills are the channels? On average players get around 50 kills per game.

Thank You!

Since the “number of kills” would be the variable dimension, I would probably not put it in the channel dimension which is usually fixed. I don’t know which approach would work the best, but I would probably try to treat each sequence as an individual sample and mix different players together first (as well as shuffling them) as this seems to be the easiest approach.
Depending on the outcome you could (and should) also try out your other approach.