Wondering if I can aggregate individual data for binary prediction? I have an imbalanced dataset. I am looking at aggregating the individual data and thinking about using the aggregated data for binary prediction. The problem is I cannot make my mind whether aggregation on  individual data is fine or not. Is aggregating individual data still a part of data preprocessing?, is this ok to do for binary prediction, and what information do I lose?
 A: 
Is it ok?

If you only want to do prediction and only measure your performance you can do literaly anything as long as you have an independent test set. Only if you try to use your model for explanation of your independent variables you need to worry about statistical assumptions.

what information do I lose?

You will lose information that is specific to one of the variables. Anyways, it is always interesting to experiment with your ideas. I would compare all approaches with cross-validation. 
A: If I'm understanding you correctly, you're trying to create more meaningful features (i.e., denser or more informative variables than what you currently have). This is a great question, and there are many many ways to do feature curation and feature engineering, so a lot of it will deal with experimenting with your data. 
However, there are techniques (like this paper below) that offer a good starting point to aggregate data and features in a systematic fashion. This paper below specifically uses hiearchical clustering trees when your features are sparse, but there are plenty of other techniques. 
Yan, X., & Bien, J. (2018). Rare feature selection in high dimensions. arXiv preprint arXiv:1803.06675.
https://arxiv.org/abs/1803.06675
Since you mentioned information loss, this paper seems to explain how you can lose information when dichotomizing an discretizing continuous variables. They offer a solution to dos in such a way to minimize information loss.
Clarke, Ellis J., and Bruce A. Barton. "Entropy and MDL discretization of continuous variables for Bayesian belief networks." International Journal of Intelligent Systems 15, no. 1 (2000): 61-92.
Hope this helps!
