Classification vs Regression: Which to choose? Suppose I have some training data pairs $(x, y)$, where both $x$ and $y$ are real values between 0 and 1. I want to train a multi-layer neural network to predict the value of $y$ given a new $x$. Now, there are two ways to do this. One is to train the network for regression, mapping $x$ to a real-value for $y$ at the output node. The other is to discretise the data in $y$, and train the network for classification, mapping $x$ to several real-values for $y$ at output nodes.
My question is: Aside from the fact that classification will introduce discretisation errors, would there likely be difference in performance between the two methods? How can I determine which method would be the best for my particular data?
Thanks you!
 A: If I understand your problem correctly, the only solution is to use regression as you need to predict the value of y rather than the class it will fall into. Discretising is rarely a solution, unless you are certain it will bring out some important information.
I believe you are misunderstanding the two concepts and the ways they should be used.
The fundamental difference between regression and classification problems is the following: The regression wants to learn a continuous target variable while the classification wants to learn a discrete such. This essentially splits the problem in two, as you have pointed out. If you are willing to predict what is the value of your predicted y, then this is a regression problem. In contrast, if you are willing to predict which class y falls into, then you should use classification.
Also note that neural networks are not always the most suitable way to go.
In summary, it is not the type of the problem that implies the efficiency, its the way you will approach it, once you have established it as a classification or regression one.
A: Discretization will result in loss of information. There is no reason to do it, unless you are certain there will emerge some kind of strong overfitting in the continuous setting, stronger than the bias you create with discretization (lol).
A: I understand that your data points are both real. Assuming dataset does not have temporal dimension regression model is more suitable than classification. If  you use classification you -probably- will have to calculate continues values for y.
In summary:


*

*Use regression if you need exact values of y for all corresponding x values.

*Use classification if you need just general inferences for y values. 


For example:
Let x be the daylight duration and y be the average temperature.


*

*regression for predicting temperature in degrees.

*classification for predicting temperature as as low, medium or high.
