Retraining a ML Model Incrementally

User 878 | 3/20/2015, 12:45:14 AM

Hi, I have a recommendation model that I got trained using GraphLab. However, I get some additional data everyday and I want to incrementally retrain my model with only the additional data. In a easy life, I could combine the old and new data and train the model afresh. However, the incremental volume of my data makes it impossible to do so. Is there a way, I can incrementally improve my model with the new data ? Any thoughts/suggestions would be highly appreciated.



User 19 | 3/20/2015, 1:07:20 AM

Hi Ashu,

At the moment we do not enable incremental training. One workaround is to keep a data set for each day, and train models on a rolling window of data from k of the past days. You can choose the value of k to balance training time with the how much past data you want to include.

Let me know if you have more questions, or if we can help in some way.

Cheers, Chris

User 878 | 3/22/2015, 8:22:43 PM

Thanks Chris for the response. Is there a plan to have such a feature included in the future releases?

User 19 | 3/22/2015, 8:26:42 PM

We are always collecting feedback from users and customers, and we prioritize new features accordingly, such as incremental training. It helps us to hear more about your particular use case. Can you tell me a bit more about your data and your task? Feel free to private message me directly, too.