How to train a neural network with a data set that in which the target is a mix of 0-1 label and numeric real value label?

Question

I am running into an issue in which the the target (label collums) of my dataset contain a mixture of binary label (yes/no) and some numeric value label.

The value of these numeric value (resource 1 and resource 2 collumns) experience a large variation margin. Sometime these numeric value can be like 0.389 but sometimes they can be 0.389 x 10^-4 or something.

My goal is to predict the binary decision and the amount of resource allocated to a new user who have input feature 1 (numeric) and input feature 2 (numeric).

My initial though would be that the output neuron corresponding to the 0-1 decision would use logistic regression activation function. But for the neuron that corresponding to the resource I am not quite sure.

What would be the appropriate way to tackle such situation in term of network structure or data pre-processing strategy ?

Thank you for your enthusiasm !

Answer 1

Your question are missing some details and i will assume some scenarios.

• If you have a classification problem: you can try group the values in intervals that make sense (you should analyze and decide for this setup), if its possible. For example: 0.000-0.250 (0), 0.251-0.500 (1), 0.501-0.750 (2) and so on. Note that neural networks are sensible for distance between values (1 is closer to 0 than 2, so 1 is more similar to 0 than 2 and so on). If that is not your case, you should binarize the values in One Hot Encode manner.
• If you have a regression problem, you should be ok without anything else. You can try normalize your outputs and observe the results, but generally it's not necessary for regression problems.
• Be sure if your dataset are free of outliers and noisy data as much as possible.
• It's important choose activations functions that are adequate for the range of values in your attributes and output. This can depend on how do you treat and setup your dataset, the range of values, normalization etc.

Update after more details in question

Your neural network should have 3 neurons in the output layer, with linear activation. As said before, normalization usually is not necessary in regression problems, but if your values are too diferent (like the range in resource 1 and resource 2) maybe some kind of adjustment (normalization, standardization etc) can be helpful. But you need try and see the results.

Answer 2

In Neural Networks the function that provides the largest interval while activation is Tanh with a result between -1 and 1

You can use it to train your model , when the label has value of false it should be -1 , and when true it should be 1

In prediction , you'll see where the value is more close , for example if you get 0.4 more close to 1 so it'll be true