sftekin
commented
1 year ago Hello Martinez,
Yes, you are correct; the predictions are 0 or 1. But, we specify the concrete hour by predicting the next time step depending on the frequency of the time intervals.
For example, if we select the time interval frequency as 24 hours, the time difference between two-time steps is 24 hours. The model always predicts the next time step given the time window. For example, let's give 20 days of observation to the model starting from the 1st of January at 00:00 AM until the end of the 21st of January at 11:59 PM. The model will predict the locations of events for the 22nd of January until 11:59 PM.
For the train-test split, I can quote from the paper, "we split the data into 10 months of training, 1 month of validation, and 1 month of testing in each experiment." This means if my model has 24 hours frequency, I can test its performance for 30 days since the test data contains 1 month of data. We do this for 4 times each year between 2015-2019.
I hope this helps, Best, Selim
On Thu, Jun 22, 2023 at 9:47 AM LMartinez @.***> wrote:
Hi,
I found your paper really interesting. I had a couple of doubts about the output though. When you say "predict locations of crime events ... in the partitioned regions for the next 24 hours" my understanding is that the prediction is just 0 or 1 depending on whether a crime is predicted to occur or not, without specifiying the concrete hour etc. Is that correct?
And my other doubt was, when you mention "for the next 24 hours" , the next 24 hours from which date? I suppose I don't understand how the dates are chosen for predicting the following 24 hours. Are they from the test split?
Thanks
— Reply to this email directly, view it on GitHub https://github.com/sftekin/high-res-crime-forecasting/issues/3, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKZ34CW4M7AU2I2LRDN2GEDXMREFLANCNFSM6AAAAAAZQHT6OU . You are receiving this because you are subscribed to this thread.Message ID: @.***>
-- Selim Furkan Tekin
Hi,
I found your paper really interesting. I had a couple of doubts about the output though. When you say "predict locations of crime events ... in the partitioned regions for the next 24 hours" my understanding is that the prediction is just 0 or 1 depending on whether a crime is predicted to occur or not, without specifiying the concrete hour etc. Is that correct?
And my other doubt was, when you mention "for the next 24 hours" , the next 24 hours from which date? I suppose I don't understand how the dates are chosen for predicting the following 24 hours. Are they from the test split?
Thanks