microsoft
Custom Heirarchical Time series forecasting with Azure Automl
Description
So what was really happening with hierarchical time series forecasting is that I wanted forecast for a very niche query and had to wait for almost 3-4 hours of the model to train over a dataset that was barely even 20MB Plus creating additional compute instances and all was quite a headache and not to forget getting predictions too was very time-consuming. It was like I wanted answers for a very small sample combination (like forecast for store 3,dept 6,state California ) but had to wait for training on other combinations as well. Hence I created a class-based solution that makes use of the classic time-series forecasting service of azure and rather trains it only on the Hierarchical combination the user wants predictions for I feel this whole thing could be converted into a proper Azure service with a frontend to render graphs and an SDK as well.
Kindly let me know what you feel about this solution.I am open to discussing the implications of the same and would like to know your feedback as well.
** There might be a lot of scope for correction and improvement in the solution that I've created. Let me know your feedback about this as well, where I could have made this more efficient or could have made the process even more optimal.
Related Issues
Checklist:
- [x] My code follows the code style of this project, as detailed in our contribution guidelines.
- [ ] I have added tests.
- [] I have updated the documentation accordingly.
So yhank you kuch graduate know my code isn't really perfect,but I would love to know your feedback on where I could improve,or perhpas make this whole process even more optmized. Incase if you'll do go ahead with this idea for the service or the sdk(I would love to contribute towards it's development).I just hope my solution happens to serve as a good starting point.
@chenhuims , The problem with evaluation is that in the test dataset, we don't have the actual values with us
@Jash271 Thank your for updating the dataset. Please take a look at my additional comments. Besides, I think we should move this example to the /contrib folder based on the current content of the notebook.
@Jash271 Thank your for updating the dataset. Please take a look at my additional comments. Besides, I think we should move this example to the /contrib folder based on the current content of the notebook.
@chenhuims ,Yes sir ,I'll look to incorporate all the changes as you have suggested.But with respect to the evaluation metric is that we don't have the actual values for the test dataset with which we can compare our results
@chenhuims , Sir besides should this could we connect over a teams call, I guess I haven't been able convincingly put across my point
@chenhuims , The problem with evaluation is that in the test dataset, we don't have the actual values with us
Can you partition the dataset into a training set and a test set so that we can evaluate the results on the test set?
@chenhuims , The problem with evaluation is that in the test dataset, we don't have the actual values with us
Can you partition the dataset into a training set and a test set so that we can evaluate the results on the test set?
@chenhuims ,I was just typing this thing out only,was thinking to do this only .
So I'll create another notebook for time being for the results thing
@chenhuims , The problem with evaluation is that in the test dataset, we don't have the actual values with us
Can you partition the dataset into a training set and a test set so that we can evaluate the results on the test set?
@chenhuims Azure ML takes care of that and a log with respect to the performance of various models like Arima have been listed. The best model then is chosen for predicting
@chenhuims , Sir please have a look now. As suggested by you, I have corrected the Typos and have added the evaluation metrics as well. I am making use of normalized_root_mean_squared_error.
Incase, If there's a problem viewing the changes, you could additionally visit this to view the changes here :
https://github.com/Jash271/forecasting/tree/master/contrib/AzureML%20Custom%20Hierarchical%20Time%20Series%20Forecasting
@chenhuims , Sir please have a look now. As suggested by you, I have corrected the Typos and have added the evaluation metrics as well. I am making use of normalized_root_mean_squared_error.
Incase, If there's a problem viewing the changes, you could additionally visit this to view the changes here :
https://github.com/Jash271/forecasting/tree/master/contrib/AzureML%20Custom%20Hierarchical%20Time%20Series%20Forecasting
Thank you for making the changes!
@chenhuims , How about creating a no code application for generating forecasts for a niche query set over a data that has multiple levels of heirarchy.
This could perhaps be a service much like Azure's other services like ocr,custom vision etc. Azure currently doesn't have something of this sort to generate time series forecasts .