Given a dataset, I need to predict the amount of time it will take to fit a model to it. I plan to do this by creating a csv containing the logs of previously fit models, and passing that data itself into a regression algorithm.

For convention, I will refer to my regression model as the meta-model and all the models being used to train my meta-model as input-models.

For the meta-model, the dependent variable would be the time taken to fit an input-model while the independent variables would be features like the number of integer columns in the input-model, number of string columns, size of the data set, system capacity, etc.

I would like to have at least a few thousand rows of data for the meta-model. However, manually downloading and preparing* thousands of datasets is not feasible so I was thinking of generating datasets using random values (including strings, numerical, datetime etc). While this will be much faster, is there any merit to training on random data? I don't care about the accuracy of the input-models, but I wonder if passing in randomly generated input-models will hurt the accuracy of the meta-model.

*(by preparing I mean unzipping the files and passing in the name and label of the dataset to my code)

The reason I need to generate this fake data is because I don't know the kind of input-models the meta-model will have to deal with in reality. As a result I cannot train the meta-model on one specific kind of data. In such a case, is using random input-models justified?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.