Some pipelines have huge resource requirements during training?

I've just started using TPOT and am training on a moderately sized dataset of 1000 observations with 5000 predictors (all continuous). During optimization, a few pipelines consume an enormous amount of memory and create huge temporary files (if I've set memory = 'auto') - on the order to 100 GB of memory. (!!)

I've tried to track down the offending type of model but have not been able to do so at this point. I have found that if I use config_dict = 'TPOT light', the problem does not occur. I'm not sure if this is a particular type of model or if it is a particular type of preprocessor (PolynomialFeatures?) or perhaps a combination of those two.

I can use config_dict = 'TPOT light', but I would rather have a larger palette of operators to choose from and merely excise the offending model or preprocessor.

Is there any information regarding what might cause this problem?