FAQ

Welcome to the OpenMeasurement Audience Modeling Toolbox FAQ

1. The modelling in the repo is based on a single campaign: how would this operate in practice?

This is just an example. In practice, an advisor demands specific dimensions to be modelled accurately and the rest using the previous techniques. This will assure that the reach/frequency surface is correctly modelled across those dimensions and will give correct aggregate answers for the rest.

2. Would the modelling be extended to take many campaigns as inputs, with the aim of minimising the frequency differences across all of these?

Yes, absolutely.

3. Or, as the ‘read me’ suggests, would there be distinct parameters for each campaign?

The suggestion to offer a virtual society for each advertiser is a strategic suggestion and for accuracy purposes to meeting individual advertiser's goal. It is not a limitation of the model.

4. If is the latter, how much time would be needed for data to build up?

Do you mean how much of the trager should be made or how long into the campaign can we start reporting? This depends on the source of data used, we can discuss more. However, we are not ruling out a generic VID model as well.

5. Is this approach suggested for TV, or websites (digital), or RPD primarily?

Our primary goal was to make sure the TV viewership data is included correctly. But there is nothing against using the same technique for the websites.

6. What is the intended input dataset: a normal panel or RPD or a mixture?

A mixture would be a better option. We don't recommend a specific input type at the moment.

7. Could the VID society approach for TV be implemented alongside a Dirac mixture approach for websites?

Yes, absolutely. We are thinking in that direction ourselves.

8. Significant emphasis is put on preserving observed frequency – is this based on a particular demand for this from broadcasters?

Yes, both explicitly and implicitly. Implicitly by asking for impressions to be valued differently on different media and in practice we realized a faithful approach is only possible if the frequency is respected. They have also just flat out asked us.

9. The allocation seems to be impression led, while other proposals have suggested that it should attempt to keep impressions together where possible. Together with the emphasis on frequency in the model fitting, is it likely that reach results are less well met?

Yes, in the examples the allocations are impression led but there is nothing against having the source randomness to be from the deterministic hash function. The main source of error is the choice of distribution, and the input data for training, if that is wrong a deterministic hash function doesn't save us.

10. The underlying distribution used to create the society is the same as the fitted distribution, as both are mixtures of 6 exponential ADFs. Does this go any way to explaining why this model fits the data better? Does this prove that using this model on real data is better than a Dirac mixture?

Correct. The reason we choose this distribution is that it matches the TV viewership frequency distributions. In fact almost the same exact issue is observed with real data.

11. The underlying distribution doesn’t seem to have any correlation between the two publishers (i.e. there is no correlation between the activities), how does the process cope with correlations?

It is a mixture of independent distributions, it can capture correlations. (In fact the mixture of deltas are also a mixture of independent distributions which can also capture correlations)

12. Does the model scale to having many publishers? The model is fitting pairwise frequencies, so grouping at 20+, there are 400 (or 441) targets for a single campaign. This increases exponentially with the publisher count. Grouping the frequencies lower reduces the number of targets, but also reduces the accuracy at higher frequencies, which is precisely the espoused benefit of the virtual society approach.

The 20+ is a suggestion, in fact one can do 5+ instead and still capture the structure of the frequency distribution. It does not reduce the precision a lot. I will add an example. If a good ansatz is used (for example a power law distribution) and we fit the frequencies 1,2,3,4,5+ (note the 5+) the structure of the distribution is learned with reasonable accuracy.

13. Can the model fitting have more than 6 groups? Or rather, is there any mechanism for this number to change itself?

Yes, of course, that 6 was just a simple example which we just threw in there. Just change it.

14. Why are the groups mixed, rather than having fractions of respondents in each? The latter approach would allow for faster allocation, possibly?

Good point. In practice you CAN assume each group is representing a portion of society. This would be just a simple tweak.

15. What are run times like for allocation, compared to the Dirac approach, with large VID counts?

Yes, the Dirac approach is faster - since allocation using CDFs are trivial but distribution with more structure requires numerical solutions to the CDF. But this is not a very large issue, as the VID activities can be designed to be sorted in the assignment node.

16. Could the VID society approach for TV be implemented alongside a Dirac mixture approach for websites?

Yes, absolutely. In fact we are thinking in that direction too.