Unlocking Data-Driven Insights with Bayesian Belief Networks

By Charlie Hammerslough

“Limited-Service Restaurants” (LSRs) is how the restaurant industry refers collectively to fast food and fast-casual dining establishments. Marketers who specialize in LSRs often employ marketing research to evaluate hypotheses about their brands or to detect segments within their markets. An important additional purpose of market research is to understand the total structure of a market, to find out what guests consider important about the LSR experience. Without understanding the way that LSR guests think, marketers fly blind about what innovations in menu or service will appeal to guests. Fundamental market research helps with brand positioning and allocating marketing resources (Marketing Mix Analysis), and also in generating unexpected directions for additional research.

Unsupervised Bayesian Belief Networks are a powerful innovation that allows the analyst to approach without preconceptions the underlying structure and dynamics of their marketing research data. The most important strategic questions that marketers ask and answer are nondeterministic. Such as, “What is the structure of our market?” or “What are the key facts to know about a dataset that characterize the underlying purchase process?”

Description and Exploration

Data analysis projects usually proceed in an orderly way through phases: Description, Exploration, Prediction, and Prescription. In practice, much of the glamour of data science is tied into Prediction (algorithm development such as Machine Learning) and deployment into Prescriptive systems (such as product recommenders or search engines). Such methods are fundamental to answering well-posed problems or evaluating specific hypotheses about how consumers behave.

Typical methods for Description and Exploration are univariate analysis of data variables, and an exhaustive search through all possible bivariate correlations. Analysts use multivariate methods such as cluster analysis or variable reduction such a principal components often to explore data.

Below we illustrate a newer tool for unstructured exploration, Bayesian Belief Networks (BBNs), which were developed in the 2000’s. BBNs use the paradigm of knowledge representation to simplify any data structure into its essential elements. This method can be applied to both unstructured and structured datasets, such as marketing surveys. BBNs generate an inference structure that represents each data variable a node in a network. Each node is probabilistically conditioned on the values of a “parent” node, and connected to the whole through arcs.  In addition to the network structure discovered by the algorithm, we can examine the overall influence of each node, called node force, which is roughly, the overall strength of correlation of a variable with the rest of the network.


The data underlying the network below were a set of attitudinal questions asked of 2,294 guests who had visited Limited-Service Restaurants (LSRs) in the previous four weeks. The questions are grouped by domain areas, such as Service, Food, Convenience, Future, Company, and Brand. The question groups are in the minds of the researchers — the BBN algorithm has no prior knowledge of how the questions are grouped. All data are attitudinal questions of rated agreement or importance across a 5-point Likert scale.

Two analyses of the modeled network are relevant: node interconnection and node strength. The first is the analysis of the overall structure and interconnections of the Bayesian Belief Network. The nodes which are highly interconnected are often “gateways” to other domains of attitudes. In probabilistic terms, the value of many attitudes is conditioned on the value of key nodes.

A second analysis is of the overall strength of each node or variable (an individual attitude) on the network as a whole. The nodes which are highly influential show which factors affect how LSR guests evaluate the available restaurant options. The influence analysis supplements the interconnection analysis in giving an overall picture of the influence on the network.

Analysis of Structure

[fig 1 — Unstructured CVA Colors.jpg]

Figure 1 shows the unsupervised network derived from the survey data on LSR attitudes[1].  Each question domain is colored differently for clarity. The first thing to notice is that, remarkably, the network algorithm discovered the similarity of the different question domains without any advance knowledge. For example, the network groups all of the questions about “Future” of the brand to the lower left of the graph, and most of the “Food” questions are grouped to the far right or upper-left. Running down the central “spine” of the graph are all of the “Brand” attitude nodes. Recall that the network algorithm has no prior information about the structure of the questions — it discovered it independently!

The nodes (or variables) that have a lot of interconnections are worth special note. The large number of connections implies that this node is the “gateway” to understanding through inference the dependencies within the network structure.

The structure of the gateway nodes gives insights, including the following:

  • Consumer answers to the Brand node “They are well-managed” is a gateway that influences perceptions of Service and Convenience.  This makes sense as an important question to the consumer because good management influences all aspects of a restaurant experience, but especially service and convenience, as well as perceptions of food safety.
  • “They are an innovative brand” conditions views of the Future of the brand, as well as perceptions about how innovative and satisfying the food is.  These food attributes are general attributes about the sense of how good the food is, rather than an evaluation of specific menu items.
  • “They are a company of strong values” is the gateway to the whole set of Company and menu-specific Food questions.  And Lunch is the key food to understanding attitudes to the rest of the menu.

Analysis of Node Strength

[fig 2 — Unstructured CVA Colors node force 2D.jpg]

Figure 2 displays the same network graph as before, with the node force shown as proportional to the radius of each node, represented as a sphere [2]. Only the top five in strength nodes are labeled. They are:

  • “Company: They are community-oriented.”  This is a surprising finding, but community relations appears to be the single strongest influence on market structure and brand perception.
  • “Company: They serve the common good.”  Again, the importance of doing good for the community is a strong influence on the entire market structure.
  • “Brand: They are well-managed.” and “Brand: They are an innovative brand.”  Understanding virtually everything about how consumers differentiate between LSR brands — brand perception, future orientation, service, convenience, and food satisfaction/value — runs through these two nodes.
  • “Service: It goes above & beyond.” This influential node conditions perceptions of food quality and other service attributes such as experience, accuracy, and operational efficiency.

Implications for Brand Marketing Strategists

The network analysis informs a strategy for allocation of marketing resources, and suggests some questions for future research. The results demonstrate that consumers are paying attention to how well a brand treats its community and whether is is perceived to be serving the common good. Some LSR brands are employers-of-choice in their communities and constructively engage in charitable and community activities. Brands that already engage with their communities should increase or redirect their brand message to highlight these activities. Other LSR brands might consider increasing their community involvement or seeking high-profile community activities and partners.

The other results highlight the importance of consumers’ perceptions of a restaurant being “well-managed” and innovative. Marketing often creates a deliberate image and story for the brand; it is up to the management of each unit to provide a consumer experience that is consistent with the brand marketing messages of cleanliness, innovation, convenience, and food quality.

Lastly, Bayesian Belief Networks can pose some questions that lead to fruitful additional research. For example, do consumers really think of restaurant food as providing three distinct domains of experience: the quality of the menu items, the health and cultural aspects, and the deliciousness/innovation? Why does brand perception of innovation condition all further discussion of the future of the brand? What exactly do consumers mean by “well-managed”? Are there unexplored areas of strength that my brand should better communicate, and what would be the effect on intention to purchase?


This post introduces a little-used, yet powerful, exploratory tool for understanding the underlying structure of marketing research data. Additional understanding generates unanticipated insights from data. In this case, a consumer survey of what differentiates LSR restaurants in the minds of consumers, in other words, the overall market structure in the mind of consumers. The tool used above is an unsupervised Bayesian Belief Network, which takes a knowledge representation approach to displaying the underlying core inference structure of the dataset.

The BBN had no prompting or prior information on the different categories of questions asked. It found them on its own, and constructed a network tree that neatly differentiates each major section of the survey.

The substantive results speak to how consumers differentiate LSR restaurant brands in their minds. The results show that main perceptual differentiators in the market are: community-orientation and common good perceptions, perceptions of management quality and overall innovation, and a perception that service at the restaurant goes “above & beyond”.

[1]  BayesiaLab Unsupervised Learning: Maximum Weighted Spanning Tree

[2] BayesiaLab Visualization Map of Node Force


Charlie Hammerslough is Discipline Lead for Data Science at VSA, leveraging marketing analytics to assure that advertising campaigns are optimally effective at targeting market segments, database marketing, and implementing relevant insights from marketing research. Charlie was formerly VP of Marketing Analytics at ServiceMaster, and VP of Research and Development at The Nielsen Company. He is a well-known expert in the field of market potential analysis, predictive analytics for marketing, and direct marketing. He teaches a course in Predictive Analytics at Northwestern University. Charlie holds a Ph.D. in Sociology and Demography from Princeton University. He can be contacted at chammerslough@vsapartners.com.