Statistical and psychometric modeling is my passion. Completing a complex model gives me a sense of accomplishment and satisfaction. But it’s not just about delivering a model — a big question for me is always how it will be used to achieve business results.
Research in this area can be disconnected from business process and impact, either because it is too focused on methodological issues in very narrow context or just too complex to implement. The big focus should be on connecting results from complex models to action.
Connected modeling is an analytical modeling process that uses both data and human knowledge sources, and can handle explanatory, causation, and prediction goals to drive business actions that move the needle. It should be an approach that is intuitive, transparent, and user-friendly, but it's difficult to achieve in conventional, sequence-based modeling approaches. There are modeling frameworks and tools that lend themselves to connected research. Bayesian network (or Bayes Net) modeling is a perfect example. But before getting into Bayes Net, let’s talk about more "conventional" modeling.
Conventional modeling involves connecting multiple results, often from different stages of the project. You may have tables to describe markets and products, segmentation solutions, sales forecasts, or interesting knowledge you’ve discovered through data mining. Each of these, in its own way, helps you understand both the data and the solution, but it’s up to you to connect these results in a cogent and meaningful way, and even when you do, it is difficult to explain and lay out for your audience.
This is where Bayes Net is useful as it shows the relationship among variables by connecting “nodes” (variables) together using directional “arcs.” For example, let’s look at predicting the risk of driving. The season affects the road condition and weather. The number of journeys is affected by both season and weather. The number of road fatalities is affected by the number of journeys and danger level, which is influenced by both road conditions and average driving speed. Even without this explanation, you can intuitively understand what this figure means and how the nodes are connected. The arcs show the causal relationships between variables, literally visualizing how variables link together and which ones, if acted on, will affect others.
But the most useful thing about this network? You can play with it by updating any part of it. This allows you to use this network to immediately see and answer what-if questions, providing "on-the-fly" simulation and personalization of results.
So how do you get there? Bayesian networks can be built in one of three ways:
- structured based on your own theory;
- structured solely by the data itself;
- or structured by combining the two, letting your own theory (e.g., by imposing some constraints and/or predefining relationships) and the algorithm work together.
Obviously, this third option is most appealing. Bayesian networks work best when they integrate expert knowledge with the natural patterns in the data.
Typical regression models ignore the relationships among predictors (e.g., a brand recognition to online and direct ad budgets). With Bayes Net you can specify elements you can control (e.g., online or direct ad budget) or cannot control (e.g., season) to optimize your ad budget between online and direct channels. All parts are connected and you can even use it as a market simulator to find the optimal solution to maximize market share.
To really get the to optimized connect results, Bayesian Net is a winner in performing connected research, by bringing data, knowledge, models, and client business objectives together all at once.
