Algo­rith­mic mod­el­ing in Audi­ence­M­an­ager can help you ana­lyze a sub­set of your site vis­i­tors and tar­get oth­ers who are sim­i­lar.  If you’re just get­ting started with Audi­ence­M­an­ager or even if you just need to brush up on this fea­ture, here are some tips that should help you get the most out of the tool.

1. Build a gran­u­lar trait taxonomy

Imag­ine some­one gives you a task to com­pile a list of  char­ac­ter­is­tics that make Adobe Dig­i­tal Mar­ket­ing Sum­mit atten­dees unique from the gen­eral pop­u­la­tion.  At the top of the list you might add the really obvi­ous dis­tin­guish­ing char­ac­ter­is­tics, like:

  • Works in dig­i­tal advertising
  • Inter­ested in audi­ence targeting

A lit­tle fur­ther down you might list some looser defin­ing factors:

  • Has a smartphone
  • Has thick glasses frames
Adobe Summit attendees

We know that Sum­mit atten­dees are unique, but what is it exactly that makes them unique?

When you build a model from a base­line, you’re ask­ing Audi­ence­M­an­ager to per­form the same task:  com­pile a list of traits that make the base­line unique from the gen­eral pop­u­la­tion.  The model results will be more accu­rate when there are more traits, or defin­ing fac­tors, to ana­lyze.  Think of your traits as tools in your tool shed; the more tools you have, the bet­ter the fin­ished prod­uct will be.

Note:  the mod­el­ing method­ol­ogy above is sim­pli­fied for illustration.

2. Select a good baseline

  • The base­line should rep­re­sent a small sub­set of your vis­i­tors. Remem­ber, the goal is to take a group of vis­i­tors who exhib­ited a desir­able trait and deter­mine what makes them unique (so you can expand your tar­get­ing to users who look like the base­line). If your base­line already rep­re­sents a major­ity of users, then there is no way to deter­mine how they are unique from the gen­eral pop­u­la­tion since they are already the gen­eral population.
  • Make sure the base­line isn’t TOO small.  If the audi­ence size is insignif­i­cant there will be lit­tle data to look at when deter­min­ing the model results.  To illus­trate, if you have 10 users in your base­line and each user has a com­pletely dif­fer­ent trait pro­file (or only a few traits that dif­fer from the gen­eral pop­u­la­tion), the model will not pro­duce results because there is no real dif­fer­ence between the base­line and the gen­eral population.
  • The base­line trait or seg­ment shouldn’t be newly cre­ated.  A gen­eral guide­line is to allow for at least 1–2 weeks of data col­lec­tion on a base­line trait/segment before cre­at­ing a model.  30 days or more is ideal.
    • Note: Audi­ence­M­an­ager will peri­od­i­cally re-run the model so even if your base­line doesn’t have much his­tory to start, it will re-analyze user pro­files to update the model.

3. Select a short look back period

When you select a short look back period (say, 30 days), you’re telling Audi­ence­M­an­ager “I’m only inter­ested in what the base­line looked like over the last 30 days.”  Remem­ber that Audi­ence­M­an­ager will re-run the model peri­od­i­cally so in this case, each time we do, your model will always rep­re­sent the most recent 30 day snap­shot of the baseline.

Lookback period

Using a short look-back period ensures each time the model is run, it will use the most recent set of data

4. Cre­ate an “exclude” data source

As part of your trait tax­on­omy, you prob­a­bly already have “site vis­i­tor” traits which rep­re­sent a major­ity (or all) of your users.  If this trait is con­sid­ered when run­ning the model, you’ll get a result that looks like a flat line and you won’t be able to scale the algo­rith­mic trait up or down.  To avoid, you’ll want to clas­sify your site vis­i­tor traits as a sep­a­rate data source.  This also goes for very large or generic traits, like “Oper­at­ing Sys­tem” or “Language”.

flat line chart

Includ­ing large traits like “Site Vis­i­tor” traits in your model will pro­duce a flat line.

Ask your Audi­ence­M­an­ager con­sul­tant about set­ting up the appro­pri­ate “exclu­sion” data source(s) for your trait tax­on­omy.  Once set up, you can change a trait’s data source in the Trait Edit screen.

When you have your trait tax­on­omy grouped into your var­i­ous data sources, you can leave a “site vis­i­tor” (or sim­i­lar) data source des­e­lected when cre­at­ing a new model, under “Select Avail­able Data.”

algorithmic curve chart

When you exclude generic traits like “Site Vis­i­tor” traits, the model should pro­duce a nice algo­rith­mic curve.

5. Apply logic

Most of all, think through the busi­ness case and make sure your model makes sense.  Cre­at­ing a model based on vis­i­tors who have hit the home­page of your site prob­a­bly doesn’t sat­isfy any use case, while cre­at­ing a model based on spe­cific prod­uct pur­chasers prob­a­bly does.  Make sure your base­line answers the ques­tion “who are my most valu­able vis­i­tors” and you’ll be off to a good start.