Suppose I want to predict the profit of a company given a number of inputs (e.g. number of employees, industry and location of headquarters).
I know that profit = revenue - expenses.
Would it be possible to force the learning algorithm to predict the revenue feature, and the expenses feature such that "revenue - expenses" predicts the profit of the company?
Please note:
I only know the profit for each labelled company, I don't know their revenue or expenses.
I do not want to predict the revenue and expenses of each company, I only want to predict the profit. In other words, I do not care about the accuracy of the revenue and expenses, I only care about the profit.
The problem explained in this message is strictly for illustrative purposes.
More generally I want: Given an input vector X, I want to find the vector Y such that f(Y) predicts what I want to predict. The function f is designed by a human.
Is this doable? Has this been studied?
[link][2 comments]