Am exploring more with algorithmic lookalike models in AAM. Once (old) blog post I came across is the following: 5 Tips for AudienceManager Algorithmic Modeling | Adobe
I have a specific question around point number 4. in the blog post:
I understand the logic behind this. Which in summary suggests to set up a new data source containing these large traits as they are meaningless (such as all site visitors) as this is a common trait amongst your total population anyway.
However, I have NOT followed the recommendation and created the algorithmic model. Looking at the Influential Traits list in the model I see that the large common traits are weighted very LOW anyway with very little influence. Which begs the question... does the data source need to be created for to set exclude these large common traits or is the article old and TraitWeight takes care of that in it's evaluation process anyway?
Thanks for your question and I think, you got it right.
The algorithm takes care of the commonly occurring traits in the selected data source and these traits are ranked lower.
Please check this link for more information on the trait weight - https://marketing.adobe.com/resources/help/en_US/aam/traitweight.html
Marking a trait as common or rare is what happens behind the scene when the algorithm runs. What we get to see is the list of influential traits after the model has run.
Looking at the influential traits list will give you a good insight on which traits are influencing your model and perhaps bringing in more of the users as the look-alike users.
There isn't a direct relationship of accuracy with the influential traits. So, I think reviewing the influential traits list and accordingly selecting the accuracy vs reach is a good point to start.
Please let me know if you have any additional question.
A follow up question on the notion of common vs rare traits mentioned in the documentation - does one get to see a list of rare traits that separate the line between the common audiences (common traits) and the specific audiences (rare or very specific traits) ? The audiences from lets say a 3rd party data source who match the original set (from frequency standpoint) will get a better score for accuracy (and less reach) whereas who don't match will get an average or low accuracy score but a wider reach. If one's objective is to reach out to a wider audience, then he would definitely want to understand what are those differentiated traits. I feel these can aid in appropriate messaging that will resonate well to these new prospects. Or, do you think this is completely unnecessary to know those differentiated traits and instead just zero out on traits which are ~70-80% accuracy and get as much a decent audience base for targeting?
Were you able to find something relevant for your query here?