What is a target function?
The target function is the real, usually unknown, rule that maps inputs to the outcome a model is trying to predict. If there were a formula that took a distributor’s quote details and returned whether the customer would accept it, that formula would be the target function. Supervised learning assumes such a relationship exists and uses labeled examples to build an approximation of it.
The model is that approximation. Training adjusts the model’s parameters to make its outputs match the examples as closely as the loss function measures. How close the model can get depends on whether the chosen model architecture can represent the target function’s shape, whether the features contain the information the target function uses, and how much noise sits in the examples.
Why the target function matters
The concept is useful because it separates three reasons a model can be wrong. The model may be too simple to represent the function, which is underfitting. The model may have learned the training examples instead of the function, which is overfitting. Or the function itself may have changed, which is concept drift. A distributor’s quote acceptance rule shifts when a competitor opens nearby, and no amount of retraining on old quotes recovers the new rule.
It also sets expectations. Some outcomes have no stable target function at all. If the outcome is dominated by information the business never records, the best achievable model is weak, and knowing that early saves the cost of discovering it late.