People learn other people's preferences through inverse decision-making

Abstract

People are capable of learning other people's preferences by observing the choices they make. We propose that this learning relies on inverse decision-making-inverting a decision-making model to infer the preferences that led to an observed choice. In Experiment 1, participants observed 47 choices made by others and ranked them by how strongly each choice suggested that the decision maker had a preference for a specific item. An inverse decision-making model generated predictions that were in accordance with participants' inferences. Experiment 2 replicated and extended a previous study by Newtson (1974) in which participants observed pairs of choices and made judgments about which choice provided stronger evidence for a preference. Inverse decision-making again predicted the results, including a result that previous accounts could not explain. Experiment 3 used the same method as Experiment 2 and found that participants did not expect decision makers to be perfect utility-maximizers.

Keywords: Inverse decision-making; Preference learning; Social cognition; Trait inference.

Figure B1

Model parameter sensitivity analysis. The plots show a comparison between model predictions and data from Experiment 1 for various settings of the prior probability parameters. ρ is the Spearman rank correlation coefficients between the model predictions and data.

Figure C1

Comparison of absolute and relative utility model predictions for four clusters of participants in the negative-attributes condition of Experiment 1.

Figure 1

Two approaches to preference learning applied to Alice’s choice of boxed lunch. In both panels, the shaded nodes represent observed information and the unshaded nodes represent inferred information. (a) The inverse decision-making approach specifies a decision function that maps Alice’s preferences and choice options to her choice and then inverts this function to infer the preferences that led to her choice. (b) The feature-based approach maps a set of features directly to the preferences that led to Alice’s choice.

Figure 2

The set of 47 choices used in Experiment 1. In each case, a decision maker chose one of between 1–5 options. The columns represent different options; different letters represent different attributes. The chosen option is shaded. The choices are ordered by participants’ mean rankings from weakest evidence to strongest evidence of a preference for attribute X.

Figure 3

Experiment 1 results. The plots show mean human and model rankings of the choices in Figure 2 from weakest evidence to strongest evidence of a preference for X for (a) positive attributes and (b) negative attributes. Error bars indicate standard errors and the number labels refer to the choices in Figure 2. The diagonal lines indicate perfect correspondence between model rankings and mean human rankings. The ρs are Spearman rank correlation coefficients.

Figure 4

Residual plots for the inverse decision-making model predictions for (a) the positive-attributes and (b) negative-attributes conditions of Experiment 1.

Figure 5

The minimum number of features from Table 5 needed by the weighted feature model to match the predictive accuracy of the inverse decision-making model for (a) the positive-attributes and (b) the negative-attributes conditions of Experiment 1.

Figure 6

Experiment 2 results. The bars show mean human ratings and inverse decision-making model predictions for the pairs of observed choices in each row. The bars point toward the choice that provides stronger evidence of a preference for X. Error bars indicate 95% confidence intervals. The first six pairs of choices differ with respect to one feature, identified by the labels in the “Features Varied” column. The last two pairs differ with respect to two features.

Figure 7

Experiment 3 results. The bars show mean human ratings and inverse decision-making model predictions for the pairs of observed choices in each row. The bars point toward the choice that provides stronger evidence of a preference for X. Error bars indicate 95% confidence intervals. Predictions for the maximizing model (not shown) are 0 for every comparison.