Join DMA

Solving Difficult Research Problems with Bayes Nets


Post Date: December 16, 2013
By: Stephanie Miller

Struhl Steven M Dr. portraitBayesian Networks – “Bayes Nets” to those in the know – can be a secret weapon in your marketing analytics approach.  Dr. Steven Struhl, founder of analytics consultancy Converge Analytics has some pretty astonishing ROI numbers to prove it.  He’s spent much of his 25-year career helping marketers and analytics professionals to use a broad analytical framework including more than regression-based analytics and to see how the more advanced and sophisticated forecasting methods like Bayes Nets are actually very approachable.

Dr. Struhl presented a compelling webinar as part of the DMA Analytics Council presents series on why using Bayes Nets techniques can help marketers solve very complicated data problems.

“Bayes Nets is an analytics technique that helps solve data problems that surpass the limitations of basic regression and correlation models.  It’s most useful when variables influence each other,” Steven said in the webinar.   “Each variable will have unique (and multiple) directions but the technique become ‘Bayesian’ when we understand how to connect the variables and their directions of influence to improve the estimates and predictions.”

In the webinar, Steven reviewed several examples where data variables have complex and multi-faceted relationships.  “We can use Bayesian techniques to match them up in various ways and forecast behavior.  Bayes Nets find those multi-dimensional linkages, but also find the variables that matter to the result we want to understand,” he said.   “Once you determine the importance of the variables, you will get answers even with sparse, missing or censored data (items that we skipped on purpose). “

The analyst does not have to know the connections or the importance of the variables in advance – the networks do the work.  “Bayesian Nets simply work better than other methods because the models can incorporate relationships and the connections between variables that are hard for the human brain to process.  Often, we find answers in addition to the hypotheses that we create – so these are findings that would never have been seen without these methods,” he said.

Steven gave an example from a reseller of services in predicting market share from questionnaire questions.  “This is a good example of how Bayes Nets can help solve thorny business questions,” he said.  A questionnaire result shows satisfaction, and we know that people who are satisfied spend more money, but we don’t really know why or how to predict what the impact on overall spending (market share) will be if satisfaction goes up or down.  The business question is, “How soon and how large of a return can I expect if I make investments in different areas related to satisfaction now?”

In this example, the model was built on the key variables related to satisfaction – the value that people found in the brand.  The Bayes Net first assembles the data in a format where you can show which values are tied to each other, and which are most  important – in this case, those most predictive of revenue (market share).  The initial model showed 85% accuracy in predicting revenue from certain variables.  When the marketer could see which factors were more predictive than others, then they could make decisions about which areas most repaid investment and also see what improvements in one area would require from improvements in others. Bayes Nets do not assume all variables remain the same when one changes, and in fact show what must happen in all areas that are measured for one to change. This helps ground decisions in a broader and more realistic framework, and also guides the implementation of any changes.

This Bayes Net goes beyond the value of a regression model not only because it performs better – in this case 85% accurate while the best regression model was only 11% accurate in predicting customer spending (revenue) – but because of how it shows the ways in which variables work together to lead to a change.

“This aspect is so powerful that it needs to be used with caution!” Steven says.  When you change one level of influence you must keep in mind that all other areas need to change as shown by a careful look at the whole analysis. This goes against the natural tendency to focus on change in one area alone. Here we can see the more complex – and realistic – picture before we act.

“Bayes Nets are the closest thing we have to a crystal ball that lets us play out various scenarios of how marketing and business/brand behavior will impact consumer behavior, without doing an actual controlled experiment.   You can draw a lot of fascinating connections from even messy survey or other unstructured data,” he said.

Why do Bayes Nets predict better?  It starts with the way two variables align.  A regression will draw a straight line through the joint results, and that is how it will estimate.   Regression is only going to offer high level of accuracy when the relationships are in that straight line. Unfortunately, much in life is not a straight line, where there is this type of a consistent relationship between variables.

The Bayes Nets approach will model the relationships accurately, considering the entire distribution of each variable, Steven said.  It’s more than a summary of two variables’ fit to straight lines.   The network tries thousands of arrangements and “lets the data speak.”

Some of the business challenges that Bayes Nets are particularly good at solving include:

  • Models of cause and effect (where supported by the data)
  • Incorporating expert judgment into models
  • Drilling into sparsely represented groups  (Far better than cross-tabs)
  • Induction by automatic learning
  • Data mining and Web analytics
  • Text analysis
  • Anywhere variables need to be tested or winnowed

Download Steven’s deck to see the taxicab example at the end, and see if you can make sense of why the Bayes Net reveals that a witness will correctly say that a white taxicab is actually white (and not yellow) only 41.4% of the time even though he is 80% accurate at identifying both white and yellow cabs. (Clue:  It’s about both how many cabs are yellow and how many white AND about how accurate the witness is.)

In order to use Bayes Net techniques successfully, you have to be able to set aside many preconceived notions about how one variable influences others, and so about the real drivers behind consumer behavior.  So many of these results are non-intuitive – until you see them.

“Yes, Bayes Nets think in ways that people don’t think,” Steven said.  “They will open up a lot of really powerful insights on predicting new market opportunity in areas that can completely defeat more traditional regression-based analysis. “

 

##  ##  ##

Editor’s note:  This webinar is one of six presented this year by the DMA Analytics Council.  Check out the full schedule and register now.   Many DMA memberships include Council participation – check with your account manager or email Stephanie Miller

Share Now: Tweet about this on TwitterShare on LinkedInShare on Facebook

0 Comments

Leave A Reply