CODEWITHSUNDEEP
statisticsmachine-learningregressionsvmdecision-tree
avinash shah
avinash shah

Reputation: 1445

Prediction using machine learning

Say I have some data for past 5 years and I have trained my classifier (anything decision tree, svm etc.) based on that i.e. given the appropriate input feature data and correct output labeling.

Now for current year when I have to make prediction (predicting the output) I can supply the input feature data I am having for the current year and the classifier would predict the correct output labels.

So far so good.

However suppose If I dont have the current input feature data, how can I go about making predictions just based on the past data?

For an example election prediction, i.e. which party would win from each constituency. In this we have lots of past data but no current input feature data so how to go about this?

Upvotes: 0

Views: 788

Answers (4)

Robert Griesmeyer
Robert Griesmeyer

Reputation: 571

You can only make predictions based upon what is in your training set, that being old or current data. The best way to get an accurate prediction that is going to generalize for new samples is to ensure that you don't over fit your model. If you feel that your model doesn't accurately reflect what is in the most current data that you must predict then you might need to go into getting additional features.

Upvotes: 0

stefan
stefan

Reputation: 3759

i think he meant to do classification. on classification you usually split the existing data into two sets: test and training. you do all training on the training data and when you think you are done you verify with the test set. if the performance on the sets is very different, then you are either overtrained or the problem cannot be classified with this type at all. if you have very little data you can try the k-fold strategy.

Upvotes: 1

ogrisel
ogrisel

Reputation: 40169

What you are trying to do is probably called "time series forecasting" and your input can then be a window over the past predictions (or past data if still inside the current window position).

Upvotes: 0

Chris
Chris

Reputation: 4983

Bogatron's comment is the correct answer. If you aren't using an input feature and asking for a label in return, classifiers are the wrong approach. Modelling the future based on past data is often done through regression. The simplest approach for this is likely least squares, which will allow you to select a simple model (think curve fitting), from which you can select a data point and compute the predicted value.

Upvotes: 2

Related Questions