I assume you're doing something like minimizing MSE between inputs and reconstructions. Instead of calculating MSE for all 21 columns, you split it into two parts: do an MSE for the important columns, and an MSE for the unimportant columns. Then weight the important MSE higher than the unimportant MSE

​

So something like

loss = 0.9 * MSE_important + 0.1 * MSE_unimportant

I’m a CS student too. Curious what algorithm the pool inspired you with