Is it possible to re-learn the meridian model?

[yhp2205]

Can we train Meridian with sample_prior and sample_posterior for a specific data for the first time and then re-learn the Meridian model by gradually adding data? I wonder if it is possible to upgrade the model based on the added period, not the entire period. The official document shows the model load and save, but there is no re-learning.

[cpulavarthi]

Hello @yhp2205,

Thank you for contacting us!

Refreshing the model in Meridian achieves what you described. Model refresh involves using the parameter posteriors from the previously trained model as a starting point (priors) for the new model. You can read more about it in our documentation linked here. Please also refer to previous discussions #424 and #248 on this topic to better understand it.

You may also check our documentation about the new_data argument in the methods of Analyzer module, which allows you to analyze the outcomes for new data using a model trained on some older data. Here you can, for example, pass only the media execution and the analyses would be done using new media and the rest of the old data. This could be an alternative you may be interested in but this doesn’t involve model retraining as you mentioned in your query.

Kindly go through the documentation shared and feel free to reach out for any further questions.

Thank you

Google Meridian Support Team