The Qual-Quant Leap: Key Driver Analysis with Bayesian Networks and GenAI

Key Driver Analysis reimagined: machine-learned Bayesian networks now fuse survey data, customer narratives, and LLM knowledge. The Qual-Quant Leap: Key Driver Analysis with Bayesian Networks and Generative AI Abstract Key Driver Analysis answers a deceptively simple question: which changes matter most for the outcome you care about? This seminar traces how the answer has evolved across three generations of methodology and equips you to practice the newest one, in which numerical data, customer narratives, expert knowledge, and LLM-encoded knowledge all become admissible evidence in a single causal model. Three Generations of Key Driver Analysis How driver analysis evolved from expert-specified structural equation models, to probabilistic models machine-learned from survey data, to today's models that fuse data, text, and LLM-encoded knowledge. A simple orientation map, organized by modeling purpose (association vs. causation) and model source (theory vs. data), frames the journey and recurs throughout the day. Foundations, and Why Conventional Driver Analysis Fails A compact introduction to Bayesian networks: models of the joint probability distribution that reason in any direction, represent latent constructs, and measure importance in information-theoretic terms. Then an honest tour of the obstacles: causal language applied to observational data, the astronomically large space of possible causal structures, multicollinearity that produces significant but wrong-signed regression coefficients, ceiling effects in rating scales, and missing values that standard remedies quietly corrupt. Building a Probabilistic Structural Equation Model, with GenAI Assistance The core hands-on workflow, end to end, on a consumer survey dataset: unsupervised structure learning, validation through perturbation and arc confidence analysis, variable clustering, induction of latent factors, and assembly of the complete model. Generative AI assists throughout: proposing names and descriptions for latent factors as they emerge and cross-checking machine-learned clusters against semantic knowledge. The result is a compact, interpretable model of what shapes the outcome. What it does not yet provide is priorities. From Association to Causation What separates an observational effect from a causal one, and what it takes to move from the first to the second. A practical criterion for identifying confounders in minutes rather than weeks, effect estimation under explicitly stated assumptions, and a worked example in which the causal effect turns out to be half the observational one. LLMs enter as an additional source of causal knowledge, treated like an expert panel: elicited, inspected, and corrected, never taken as evidence. From Drivers to Decisions: Optimization Why a ranked table of effects must never be read as an action plan. Optimization under realistic constraints: the model's own joint probability as a built-in plausibility check, competitive benchmarks as achievability limits, and costs where they are known. Priority-sequence optimization for perceptions, which cannot be dialed in directly, versus point optimization for controllable levers. The output is a defensible, ordered plan, not just a ranking. The Qual-Quant Leap: From Narratives to Networks The centerpiece of the third generation. A corpus of open-ended customer reviews is transformed into a respondent-level dataset: thematic and emotional dimensions are extracted from the text, and every document is scored on every dimension. The machine-learning workflow from Module 3 then runs unchanged on this text-born data, yielding a full driver model without a survey instrument. Qualitative material has entered the quantitative pipeline. Synthesis and Q&A Triangulating across knowledge sources, maintaining an audit trail from raw input to reported effect, and the honest list of pitfalls with their safeguards. Open discussion.