First of all - I'm new to all of this.

I'm trying to create a model to predict the release of ios 12 based on previous years. I've got an excel that has a format like this: ios version | release name | date

Each version consists of about 5 betas + general release. I've set up the import like this: File import

Secondly I've set up flow like this: Flow diagram

First step selects all the rows that have [date] param - those go to either kNN or Linear Regression, everything else (1 row with ios 12 release date) goes to prediction and then to a table.

Cool, now depending on the modeling I get different results (as expected). It's either 2015-09-17 (kNN) or 2018-09-10 (LR). I've used the Test & Score, which gives me:

Test & Score results

If I'm reading this properly - Linear Regression is more accurate (R2 column), meaning that iOS 12 will release on 10th of Sept 2018 (DUH! not 2015).

But if I look at release dates up to now, this might be wrong, because no other release happened so soon:

iOS release dates

So I also did an exercise where I've adjusted the dates, so that they are all in 2018 (so that the year is less important, and I'd care more about relation within each year, given the features) - here are predicted results:


and Test & Score for the offsets: enter image description here

Now... I would welcome all comments. Am I using the wrong modeling? Something other than kNN & LR? Should I be using offsets? Am I using the tool completely incorrectly?


2 Answers 2


So you have two independent variables. iOS version and release name, which is actually a type. Both are categorical and only one has any predictive power, because iOS version you are trying to predict does not appear in your training set. Thus, effectively, you are predicting based on the release type only and only the "release" category is actually meaningful because this is what you are trying to predict.

I can hardly think of a model that would perform ok with such limited information. Anyway, regression models will always give you some answer, even if your input is unsound.

I guess the better features would be:

  • month
  • day of month
  • days since last release
  • months since last release
  • years since last release
  • days since beta
  • months since beta
  • years since beta

This type of thinking. Then just by looking at the data you would clearly see that prod releases of iOS happen always in mid Sept each year. No need for model to figure it out :) Any method would suffice, especially kNN or least squares.


Not clear how to predict new column of data in Orange here nor provided in Orange doc and tutorials (prove me wrong). I have known dataset sorted in matrix form (named columns and rows). Now i want to predict new column based on past data pattern (what ever classifier used). I get only predictions for last already known column!? Why do i need predictions of already known data? Makes sense only to test performance of "trained model" and nothing else. The whole purpose of NN's is to learn from pattern data sets and then predict probability with confidence new ones! So my question here is how to mark new column in test file in order to get "New Predictions" ? Or this is entirely not possible in Orange ? it would be entirely more easy if in widget called "Prediction" would have option "Predict new" ! So why we have to struggle with this in 2023 ?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.