Contributions from baseline/intercept, trend and extra features

[xijianlim]

Hi, i've noticed that the examples have 0% contribution from the baseline when running the functions:
plot_media_baseline_contribution_area_plot or create_media_baseline_contribution_df in your example notebooks; with media cannels having a 100% contribution.

In addition, i've the code base over a data set of mine and this was also the case. This would mean even the "extra features" contribution default to null or 0 contribution, which makes interpretation a little strange when suggesting media is 100% influencing the KPI target.

Is there a way to set or tune the baseline/intercept so that is has a ball park contribution? Are there any suggestions you would suggest in the mean time?

[pabloduque0]

hey @xijianlim !

As for what the examples concern, it is dummy data and the way it is generated it does have a very small baseline so I would not worry too much for that case. That said we can probably increase the baseline of the dummy data to be reflected by the model as well.

Okay if you run into this situation with actual data we can definitely look into it.

The easiest way to alter any of this is through custom priors (eg. decreasing media priors and increasing other priors such as intercept or extra_features). Im working on a better API for passing custom priors which should be ready very soon. I can notify you here when its released.

[pabloduque0]

You can do it in a more clean way directly when scaling the costs:

costs_scaler = preprocessing.CustomScaler(divide_operation=jnp.mean, multiply_by=0.1)
# then apply scaler

After all you are setting the scale of your priors, that will depend on your specific case, your data and what priors you want to set. You should have some ballpark idea of what media contribution you should have (same for other metrics/insights), you can test different scales of your priors that achieves both good metric models and also makes sense based on what you expect it to look like. Or you can also have a more heuristic approach.

[rafpinter]

Hi everyone!
I'm trying to get the contribution from the extra features as well.

You can modify the function yourself to calculate the decompositon values:

something like this from the mean extra_features_contrib=media_mix_model.trace['beta_extra_features'].mean(axis=0)*extra_features

I took the plot function and made my own modification to produce a dataframe. You just need to then use the trace function to get the intercept as well if needed.

@xijianlim where exactly did you implement this modification? Would you mind providing more details on the code? That would help me a lot.

Thanks!

[xijianlim]

Hi i implemented this mod by basically reverse engineering the plot function: "plot_media_baseline_contribution_area_plot".

You can use the trace function to get the model's multiplier in something like this:

`extra_features_contrib=media_mix_model.trace['extra_features'].mean(axis=0)*extra_features

[AruunnAnanda]

Hi @xijianlim, thanks for sharing the code to deduce the contribution by extra features.

I've two questions -

1. The column that you actually meant was this - mmm.trace['coef_extra_features'], right?
2. Shouldn't we inverse transform to get the real contribution from these extra features? The co-efficients of extra features refer to the scaled values, right?

I'm not a data scientist, but rather an analyst. Sorry for my lame questions. Thanks!!

[xijianlim]

sorry ive been away for too long. short answer is yes; you should inverse transform these variables.