Retrain Machine Learning Model On New Data

Hello,

I have trained an SVM model using fitcsvm and saved it to disk.

Now I have new data that were never used by the model before.

How can I retrain the saved model over the new data?

Please, note this is just a simple model not a real time streaming model update.

Thank you!

[Ryan Thomson]

Any solution to this question?

Answer 1

[the cyclist]

I don't really understand the question. There is no such as "re-training" an existing model. You can do one of two things:

1. Train the model on the new data
2. Make predictions from the old model on the new data

In the first case, just run fitcsvm on the new data, and you have a new model.

In the second case, use the the predict() method of the old model on the new data.

Or maybe I'm misunderstanding something.

[Ryan Thomson]

Guess what I am looking for is a way to do a version of transfer learning for deployed SVMs.

Say I have deployed a SVM as part of my product to an enduser, the enduser has the means to capture their own training data and access to the saved source SVM, and I want to allow the enduser to train (only on the new customer training data) the source SVM into a target SVM now customized for the enduser's system (without access to the original traning set and without losing previous knowlage). Is this possible with SVMs in Matlab? Maybe a version of incremental learning?