Big Marketing with Small Data

Teaching kids to learn is tricky, I know – my wife’s a teacher and I hear about the challenges, daily. Teaching machines to learn is tricky too, particularly when you don’t have many examples of success to learn from.  It's kind of like teaching a toddler what dogs look like, with only a poodle.  This is a problem in marketing too, when you have a new marketing initiative but, very little success to teach from. But, one of our interns in data science, MIT student Mattie Wasiak, has just completed a set of experiments, guided by Dr. Luis Perez-Breva, our board advisor in artificial intelligence, that have revealed a way to materially improve Micronotes’ predictive performance; and it’s all about teaching methods.

Like kids, machines learn faster when the information is provided in a way that enables learning across multiple dimensions and contextualized. In the classroom, this is called differential instruction – where lessons are designed to enable multiple learning styles to process information. In machine learning, how the data is presented to the algorithm(s) is key to promoting learning.

Moreover, if you don’t provide kids or machines with enough information in the right form, you get no learning and very low test performance.  Data scientists use many methods to test predictive performance but, we like to keep it simple:

What fraction of our predictions were correct?

This measurement is called precision. For example, if we predict that 100 people are auto loan buyers and 83 of them actually took an auto loan, that’s an 83% (0.83) precision value.  

 So, let’s look at the predictive performance of one bank data set that had very sparse information on business lines of credit and mortgages using both old and new teaching methods (Table 1).

 Table 1 - Comparative Results

Table 1 - Comparative Results

Here you can see that using existing methods – we can’t even predict business line of credit or mortgage prospects due to sparse data; meaning, this bank doesn’t have enough sales in those categories to enable prediction.

However, using the new method, with the same bank’s data – we enable the prediction of business LOC and mortgages, and materially improve the precision of auto loan and CD scores as well.

This is important because it means that we can effectively support new business initiatives where sparse sales data is present, as well as improve existing predictions. For example, this bank has very few sales in business LOC and mortgages but, we are now able to predict the best audiences for these sales conversations and, because the machine is learning – the accuracy of those predictions will only improve as interview responses and sales train the model.

This new machine learning method will be available to all Micronotes Pro and Enterprise customers starting in the fall of 2017.  So, if you’re feeling a little short on teaching data for your new marketing initiative, no worries – we’ve got your back!

 If you’d like to learn more about Micronotes or care to try our software for free for 30 days – click the button below.