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

Highlights

• We propose that people learn others’ preferences by inverting a decision-making model.

• In three experiments, participants inferred people’s preferences from their choices.

• Inverse decision-making provided a strong account of participants’ inferences.

• Inverse decision-making is more parsimonious and principled than other accounts.

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.

Introduction

One way to learn what other people like is by observing the choices they make. For example, suppose that Alice orders a boxed lunch that includes an eggplant sandwich and you want to know how much Alice likes eggplant sandwiches. If Alice ordered the only box with an eggplant sandwich, you might infer that Alice has a strong preference for eggplant sandwiches. If the eggplant sandwich is part of the only box that contains a cookie, you might instead infer that Alice has no particular preference for eggplant sandwiches and she really wanted the cookie. Although people commonly make these sorts of inferences, this example illustrates that someone’s choice could have many different explanations, and deciding which of these explanations is best can be a challenging inductive problem.

Inferences like these have been studied in the literature on interpersonal attribution (Gilbert, 1998, Hamilton, 1998), and have been the target of developmental work with children (Diesendruck et al., 2015, Hu et al., 2015, Kushnir et al., 2010, Lucas et al., 2014, Luo et al., in press, Ma and Xu, 2011, Repacholi and Gopnik, 1997). Most of this literature, however, does not emphasize computational approaches (for some exceptions, see Kunda, 1998, Lucas et al., 2014, Medcof, 1990). Research in economics and marketing has produced multiple computational methods for inferring consumers’ preferences from their choices (Green and Srinivasan, 1990, Varian, 2006), but these methods have not been explored as psychological models. By contrast, there are multiple psychological models of how people make choices (Busemeyer and Johnson, 2008, Schneider et al., 2007, Shenoy and Yu, 2013, Train, 2009), but few attempts to apply models like these to the problem of inferring people’s preferences from observations of their choices. In this paper, we explore a computational approach to preference learning based on inverting a decision-making model and test it as a psychological account. We call this approach inverse decision-making.

The inverse decision-making approach is illustrated in Fig. 1a. The figure shows an example in which Alice chooses between three boxed lunch options: (1) eggplant sandwich and a cookie, (2) turkey sandwich and a slice of cake, and (3) tuna sandwich and an apple. The utility function in Fig. 1a (depicted by a bar chart) shows that Alice prefers the eggplant sandwich over the other sandwiches and the cookie over the other desserts. A decision-making model specifies a decision function that maps preferences to choices. Given Alice’s preferences, any standard model of decision-making will predict that Alice will choose the option with an eggplant sandwich and a cookie. The shading on the nodes of the graph in Fig. 1a indicates what information about Alice’s choice is visible to an observer. In this case, that includes the three boxed lunch options and Alice’s choice. The unshaded node indicates that Alice’s preferences are not visible to an observer. Even so, the observer can invert a decision-making model to make inferences about the unobserved preferences that led to the observed choice.

Fig. 1b shows an alternative feature-based approach that does not rely on a decision-making model. Instead, this approach characterizes Alice’s choice using a set of features. For example, the features in Fig. 1b indicate that Alice chose the only option with an eggplant sandwich and the only option with a cookie, that her choice had two attributes (eggplant sandwich and cookie), that she passed up four attributes (turkey, tuna, cake, apple), and that she passed up two options (the two boxes that she did not choose). These features carry information about Alice’s preferences, and the feature-based approach relies on an inference function that maps choice features to preferences. For example, the larger the number of chosen attributes, the less likely it is that she was specifically interested in the eggplant sandwich, and the larger the number of forgone options, the more likely it is that Alice has a strong preference for eggplant.

The inverse decision-making approach has received little attention in the social psychology literature, but the feature-based approach has served as the basis for several influential accounts of interpersonal attribution (e.g., Jones and Davis, 1965, Kelley, 1973, Newtson, 1974). One example of the feature-based approach is Jones’s and Davis’s (1965) correspondent inference theory (CIT). One choice feature identified by CIT is whether a chosen attribute is common to other options; if not, then the choice is especially informative about the decision maker’s preferences. For example, if Alice chose the only option that included an eggplant sandwich, then her choice provides strong evidence that she was interested in the eggplant sandwich. In CIT, this idea is called the principle of non-common attributes.

The choice features and inference principles identified by CIT and other feature-based approaches are intuitive. In addition, as Fig. 1 suggests, there are cases in which inverse decision-making and feature-based approaches make the same inferences about Alice’s preferences. However, feature-based approaches have two fundamental limitations. First, they assert that the inference function respects a set of principles, but they do not provide a complete set of principles or suggest a way to enumerate these principles. Second, CIT makes no clear predictions about how conflicts between different principles should be resolved (Newtson, 1974). Both limitations arise because it is difficult to characterize the inference function directly. The inverse decision-making approach overcomes these limitations by characterizing the inference function indirectly, letting it emerge from some simple assumptions about decision making.

The inverse decision-making approach is an instance of a class of modeling approaches that rely on what Jara-Ettinger, Gweon, Schulz, and Tenenbaum (2016) have called the naïve utility calculus. Naïve utility calculus refers to the expectation people have that others will generally make choices that produce greater utility. Combining naïve utility calculus with inverse reasoning has led to a number of useful accounts of social inference in recent years (Baker et al., 2017, Baker et al., 2009, Baker and Tenenbaum, 2014, Jern and Kemp, 2015, Tauber and Steyvers, 2011, Ullman et al., 2009, Wu et al., 2014). However, few studies in this literature have explored the basic question of how people infer what other people like and dislike by observing their choices. Studies that have explored this question have focused on how children learn simple preferences (Lucas et al., 2014), how adults predict other people’s choices (Bergen, Evans, & Tenenbaum, 2010), and how adults take into account deviations from optimal choice behavior when reasoning about other people’s choices (Evans, Stuhlmueller, & Goodman, 2016). However, previous tests of inverse decision-making as a psychological account of how people learn other people’s preferences have been limited. For example, using a model essentially identical to the one we present below, Lucas et al. (2014) tested predictions on children’s inferences for 11 different observed choices. In this paper, we test the inverse decision-making approach in much greater detail, comparing its predictions to adults’ inferences about many more choices: 47 different choices in Experiment 1, 6 different choices in Experiment 2, and 8 different choices in Experiment 3. Testing the model on a greater number of choices allows us to test the robustness of the inverse decision-making approach and more thoroughly compare it to the feature-based approach.

All of our experiments used a preference learning task in which a hypothetical person makes a choice between multiple discrete options, each with multiple attributes. Fig. 2 shows a set of such choices. Each choice in the figure has between one and four options, represented as columns. Each option has between one and five attributes, represented by letters, with identical attributes identified by the same letter. The attributes may be desirable, like different candies, or undesirable, like different electric shocks. In all cases, the decision maker chose the leftmost option, which includes attribute X. Suppose that the different attributes are different kinds of candy. Some of the choices in Fig. 2 provide strong evidence of a preference for candy X. For instance, consider Choice 47, in which the decision maker chose a single piece of candy X over one piece each of candies A, B, C, and D. Intuitively, this choice provides strong evidence that the decision maker has a preference for candy X. Other choices provide weak evidence of a preference for candy X. In Choice 14, for example, the decision maker chose candy X plus three other pieces of candy over only a single piece of candy. Intuitively, this choice provides little evidence of a preference for candy X because the decision maker may have wanted a piece of candy other than X, or may simply have wanted more candy.

In the next section, we describe a formal model that can capture these intuitions. We then discuss the results of three experiments that test the model by comparing its predictions to people’s inferences about choices like the ones in Fig. 2. Finally, we present an analysis of whether our results could be explained just as well by a feature-based model.