There are a certain number of issues of type A that have occured for 12 different locations, each with a different number of users (the users report the issues, and this is where my data comes from).
There are some new locations opening at the end of July, and I am trying to be able to predict how many issues of type A they should expect to see at these new sites. Any advice?
Could you add a bit more explanation to your question?
What do you mean by a site? A website or a site of locality where the issue is raised or something else?
What do you mean by an issue? A security breach? Data loss? System failure of some sort?
It would be better to explain your question with some examples (or at least pseudo-examples) to make it clear.
The problem you are refering to is called regression and is a supervised machine learning problem. You need to train a model using some features and labels you get from previous examples. Your model when trained properly will be able to fit your data and make predictions.
You may need to do some research on regression algorithms in order to impement one that suits you. Look for gradient descent or conjugate gradients or some other advanced algorithm.
Maybe a keyword analysis of new sites will fit. And the link to type A.
X keywords : N query type A.
What is your take are the issues being reported because of the site characteristics or is it because of the user characteristics as well? If the second is true , you may have to model it @ user level and hopefully this will give you more data and as suggested by few , you may provide more description of the problem.
thanks for the answers!
The sites are locations where a new software is being put into effect. The issues are software malfunctions- each of these is also assigned a type, or a label indicating the section of the company that deals with the issue. does this clarify any?
I have tried regression, but am having trouble finding a model that fits the data.
In machine learning, you would have a set of features for each instance. A model is learned when you fit those features to predict the seen data well enough to be able to predict the unseen data in the future.
After your clarification, I don't see a scenario where features are applicable for you to be able to predict the next possible outcome. But, basic statistics can be applied to estimate it. This can be done as follows:
- Considering a location as an instance, each instance provides no. of issues raised of certain types (for example, type A)
- Find the class probability distribution for the given type. This would be simpler to first view it as a histogram similar to the one attached. The type of distribution for your problem is probably Poisson.
- Find the expectation for your distribution. For example, in Poisson, E[X] = lambda, which is the mean.
- The calculated expectation is the estimate of how many issues of the given type might be raised in the new site.
Note that 12 sites is a very small number for you to fit a probability distribution, but a general mean might help to make the most of what you have.
Hope this helps.
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining