Consumer Preference and
Decision Modeling

When consumers are asked how important a feature is, they generally give high scores to all features. Key Driver Analysis (KDA) ignores this misleading importance stated by the consumer (Stated Importance); it calculates the extent to which sub-feature performances trigger "overall satisfaction" or "purchasing intent" via statistical regression (Derived Importance). To resolve multicollinearity among variables, we use Shapley Value Regression, which is based on Game Theory. Thus, we decompose the model's total explanatory power (R²) according to the net and pure contribution of each feature to the final decision.

Especially in industrial purchases (B2B), technology infrastructure selections, or decisions requiring high-involvement such as real estate, the consumer or committee does not act based on a single criterion. The Analytic Hierarchy Process (AHP) breaks down these complex decisions into a tree structure, creates pairwise comparison matrices, and determines the mathematical weight of the criteria.

In traditional Likert-type surveys (scoring between 1-5), the tendency of participants to rate everything a "5" leads to "Scale Use Bias." MaxDiff modeling forces consumers to choose the "Best" and "Worst" (Best-Worst) each time by presenting them with algorithmically determined specific sets. This yields metric absolute distances (ratio-scaled scores) calibrated to an index of 0 to 100 among features, allowing us to state "Feature A is exactly 3.2 times more important than B."

Consumers do not evaluate a product in a vacuum (in isolation), but on a shelf (context) where competitors are also present. Discrete Choice Experiments (DCE), based on Daniel McFadden's Nobel-prize winning Random Utility Theory, present consumers with alternative competitive packages and ask them to make a rational "choice". This modeling mathematically predicts consumer utility via Multinomial Logit algorithms.

On a retail shelf or an e-commerce site, the consumer encounters hundreds of SKUs. However, the mental decision process is not linear. The consumer creates a mental branching by first looking at the "Format" (e.g., liquid vs. powder), then the "Brand", and then the "Size". Consumer Decision Trees reverse-engineer and decipher this mental hierarchy (node & leaf architecture) using agglomerative (bottom-up) clustering or Gini impurity algorithms.

Classical economics assumes that the consumer is fully rational towards price and value. However, Nobel laureate Daniel Kahneman's Behavioral Economics literature (Prospect Theory) has proven that consumers' perception of "gain" and perception of "loss" are entirely asymmetrical. People react much more intensely (on average 2.25 times more) to a loss (e.g., a price hike) of the same amount than they do to a gain (a discount).

Just because a consumer likes a product (Attitude) does not mean they will definitely buy it (Behavior). Purchasing is a combination of social circle pressure (Subjective Norms) and the individual's belief that they have the power/ability to buy that product (Perceived Behavioral Control). We apply Structural Equation Modeling (SEM) to model these abstract and latent constructs.

Standard Logit models assume that the consumer "compensates" for a deficiency in one feature (e.g., poor camera quality) with an excess in another (e.g., cheap price). However, in real life, consumers have non-compensatory absolute red lines (e.g., "Regardless of price, if it doesn't support 5G, I won't even include that phone in my consideration set"). Lexicographic and EBA algorithms model this non-compensatory process where consumers ruthlessly eliminate alternatives.