We test methods, based on cognitively-simple decision rules, that predict which products consumers select for their consideration sets. Drawing on qualitative research we propose disjunctions-of-conjunctions (DOC) decision rules that generalize well-studied decision models such as disjunctive, conjunctive, lexicographic, and subset conjunctive rules. We propose two machine-learning methods to estimate cognitively-simple DOC rules. We observe consumers' consideration sets for global positioning systems for both calibration and validation data. We compare the proposed methods to both machine-learning and hierarchical-Bayes methods each based on five extant compensatory and non-compensatory rules. On validation data the cognitively-simple DOC-based methods predict better than the ten benchmark methods on an information theoretic measure and on hit rates; significantly so in all but one test. An additive machinelearning model comes close on hit rate. Our results are robust with respect to format by which consideration is measured (four formats tested), sample (German representative vs. US student), and presentation of profiles (pictures vs. text). We close by illustrating how DOC-based rules can affect managerial decisions. Keywords:Consideration sets, non-compensatory decisions, consumer heuristics, statistical learning, machine learning, revealed preference, conjoint analysis, cognitive complexity, cognitive simplicity, environmental regularity, lexicography, logical analysis of data, decision trees, combinatorial optimization. Two-stage, consider-then-choose decision rules are particularly relevant in the automobile market, but modeling and forecasting such decision rules is of general interest. When consumers face a large number of alternative products, as is increasingly common in today's retail and web-based shopping environments, they typically screen the full set of products down to a smaller, more-manageable consideration set which they evaluate further (e.g., Bronnenberg and Vanhonacker 1996;DeSarbo et al., 1996; Hauser and Wernerfelt 1990; Jedidi, Kohli and DeSarbo, 1996;Mehta, Rajiv, and Srinivasan, 2003;Montgomery and Svenson 1976;Payne 1976;Roberts and Lattin, 1991;Shocker et al., 1991;Wu and Rangaswamy 2003). Consideration sets for packaged goods are typically 3-4 products rather than the 30-40 products on the market (Hauser and Wernerfelt 1990;Urban and Hauser 2004). Forecasting consideration sets can explain roughly 80% of the explainable uncertainty in consumer decision making (assuming equally likely choice within the consideration set, Hauser 1978). In complex product categories research suggests that at least some consumers use non-compensatory decision processes when evaluating 3 many products and/or products with many features (e.g., Payne, Johnson 1988, 1993). 2 CONSIDERATION SETS AND DECISION RULESIn this paper we explore machine-learning algorithms based on non-compensatory decision rules that model decisions by consumers in the consideration stage of a consider-thenchoo...

W e develop and test an active-machine-learning method to select questions adaptively when consumers use heuristic decision rules. The method tailors priors to each consumer based on a "configurator." Subsequent questions maximize information about the decision heuristics (minimize expected posterior entropy). To update posteriors after each question, we approximate the posterior with a variational distribution and use belief propagation (iterative loops of Bayes updating). The method runs sufficiently fast to select new queries in under a second and provides significantly and substantially more information per question than existing methods based on random, market-based, or orthogonal-design questions.Synthetic data experiments demonstrate that adaptive questions provide close-to-optimal information and outperform existing methods even when there are response errors or "bad" priors. The basic algorithm focuses on conjunctive or disjunctive rules, but we demonstrate generalizations to more complex heuristics and to the use of previous-respondent data to improve consumer-specific priors. We illustrate the algorithm empirically in a Web-based survey conducted by an American automotive manufacturer to study vehicle consideration (872 respondents, 53 feature levels). Adaptive questions outperform market-based questions when estimating heuristic decision rules. Heuristic decision rules predict validation decisions better than compensatory rules.

We investigate the feasibility of unstructured direct-elicitation (UDE) of decision rules consumers use to form consideration sets. With incentives to think hard and answer truthfully, tested formats ask respondents to state non-compensatory, compensatory, or mixed rules for agents who will select a product for the respondents. In a mobile-phone study two validation tasks (one delayed 3 weeks) ask respondents to indicate which of 32 mobile phones they would consider from a fractional 4 5 x2 2 design of features and levels. UDE predicts consideration sets better, across profiles and across respondents, than a structured direct-elicitation method (SDE).It predicts comparably to established incentive-aligned compensatory, non-compensatory, and mixed decompositional methods. In a more-complex (20x7x5 2 x4x3 4 x22 ) automobile study, noncompensatory decomposition is not feasible and additive-utility decomposition is strained, but UDE scales well. Incentives are aligned for all methods using prize indemnity insurance to award a chance at $40,000 for an automobile plus cash. UDE predicts consideration sets better than either additive decomposition or an established SDE method (Casemap). We discuss the strengths and weaknesses of UDE relative to established methods.

A new approach for measuring consumer brand perceptions from consumer-created brand imagery via deep learning.

Despite the growth of online retail, the majority of products are still sold offline, and the "touch-and-feel" aspect of physically examining a product before purchase remains important to many consumers. In this paper, we demonstrate that large discrepancies can exist between how consumers evaluate products when examining them "live" versus based on online descriptions, even for a relatively familiar product (messenger bags) and for utilitarian features. Therefore, the use of online evaluations in market research may result in inaccurate predictions and potentially suboptimal decisions by the firm. Because eliciting preferences by conducting large-scale offline market research is costly, we propose fusing data from a large online study with data from a smaller set of participants who complete both an online and an offline study. We demonstrate our approach using conjoint studies on two sets of participants. The group who completed both online and offline studies allows us to calibrate the relationship between online and offline partworths. To obtain reliable parameter estimates, we propose two statistical methods: a hierarchical Bayesian approach and a k-nearest-neighbors approach. We demonstrate that the proposed approach achieves better out-of-sample predictive performance on individual choices (up to 25% improvement), as well as aggregate market shares (up to 33% improvement). History: K. Sudhir served as the editor-in-chief and Scott Neslin served as associate editor for this article.