Variable Contribution/decomposition

[xijianlim]

Hi, is there a way to get variable contributions in particular: trend, extra features, base/intercept and media features?

[pabloduque0]

Hi there @xijianlim ! thanks for opening up this discussion :)

For media contributions you can rely con get_posterior_metrics. Which can also be plotted with plot_bars_media_metrics.

We are releasing another plot (hopefully next week) that does the contribution over time of each media variable and the baseline (anything non media: trend, extra features, base/intercept, ...).

However if you are interested in doing custom calculations yourself, everything can be found in the model object in the "trace" attribute. Being mu your "prediction" you can calculate the contribution of any variable by multiplying its beta against the data. Feel free to ask follow up questions if you have specific questions over certain calculations.

Also I would be interested in exactly what kind of contributions you are looking to obtain: over time, distributions, point estimates, ... so we can add more functionality in case is missing.

FYI @yukaabe

[pabloduque0]

My first guess will be that you did not pass the target scaler to create_media_baseline_contribution_df and therefore one is scaled and the other one isnt. Could that be the case?

If not let me know and we can try to see what is going on.

[selectabo]

I am using the target scaler in create_media_baseline_contribution_df.

"contribution = plot.create_media_baseline_contribution_df(media_mix_model=mmm,target_scaler=target_scaler)"

Should I be using extra features coefficients * scaled extra features inputs? If I do that then they're really small, like less than 1.

[pabloduque0]

Although I think you are doing it right, just making sure you are not summing across the samples dimensions, for example doing the mean across samples first:

mmm.trace['coef_extra_features'].mean(axis=0) * extra_features_train

Could you show the exact 2-3 lines that you are using that produce the exact output you are referring to? Thanks!

[selectabo]

I'm just using those two lines above, to get the contribution for everything and then the features one separately.

contribution = plot.create_media_baseline_contribution_df(media_mix_model=mmm,target_scaler=target_scaler)
features = mmm.trace['coef_extra_features'].mean(axis=0) * extra_features_train

[pabloduque0]

@selectabo sorry for the slow response!

The only case where that behaviour can happen is if the seasonality, trend and intercept sum are negative. Then extra features contribution will be higher than the "total baseline".