I am looking to develop a prediction model using machine learning. The output is a scalar and it's an Integer. I have the flexibility to change the peer inputs oputpout for the learning model to find the solution. My question is that the range of outpouts has an impact on the effectiveness of the model and its accuracy ?. For example if my outpouts in learning peers vary between 0 and 100 in scenario 1, and vary between 0 and 1000 in scenario 2, would my learning model be more accurate in scenario 1 ?.
Hi Samir,
probably it depends on your learning stategy: if you are considering your task as a classification task, than the output range typically won't have impact because your output integers just represent discrete categories (categorical variables). If you consider it as regression task, than the range might have an impact (at least if you have more than two output integer values and you map them in a non-linear way to your range) .
Greetings, Sebastian
Hi Samir,
probably it depends on your learning stategy: if you are considering your task as a classification task, than the output range typically won't have impact because your output integers just represent discrete categories (categorical variables). If you consider it as regression task, than the range might have an impact (at least if you have more than two output integer values and you map them in a non-linear way to your range) .
Greetings, Sebastian
Dear Touzani,
Is scenario 1 for training and scenario 2 for testing?
Anyway, I recommend you to normalize the data before the learning stage by using z-score (subtract by mean and divide by standard deviation). In this way, your data will operate on the same scale.
Dear Samir,
Yes, it affects the predictive outcomes. Different machine learning algorithms may have different impact due to their learning procedure.
The answer of your question: It depends which machine learning algorithm do you apply. It's quite possible if an algorithm produces better results in scenario 1, its more likely it may not provide better accuracy in scenario 2.
Every machine learning algorithm has its strengths and limitations. you have to empirically evaluate it.
Dear Samir,
I do not think that the output range will directly affect the performance of your model. With the output, the most important criteria are separability for classification or uniqueness for classification. For classification you ideally want no overlap between output regions. For regression, you ideally want a single input to correspond to a single output.
Of course you may affect the outputs by the operation you are performing to change the domain in scenario 1 to scenario 2, and this may affect the separability or uniqueness of the problem, and indirectly change the performance of your model.
Thank you every body for the valuble reponses. the need is for regression purpose.
First of all, there is no rule of thumb. It depends on the distribution of dataset you are using.
Most common practice is to normalize your dataset before passing to the network. There are different types of normalization techniques.
So by doing analysis you can find out the co-relation.
Extracting more related domain information is also required sometimes. For some algorithms it will affect and for some it wont. It will be based on trail and test, I do believe.
I think the AD automatic differentiation is the key element of the response. AD algorithms are more precise in a small integer interval. So i think , the response is yes, it's will be more precise for the small integer interval.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training