Contrasting Computational Models of Mate Preference Integration Across 45 Countries

Humans express a wide array of ideal mate preferences. Around the world, people desire romantic partners who are intelligent, healthy, kind, physically attractive, wealthy, and more. In order for these ideal preferences to guide the choice of actual romantic partners, human mating psychology must possess a means to integrate information across these many preference dimensions into summaries of the overall mate value of their potential mates. Here we explore the computational design of this mate preference integration process using a large sample of n = 14,487 people from 45 countries around the world. We combine this large cross-cultural sample with agent-based models to compare eight hypothesized models of human mating markets. Across cultures, people higher in mate value appear to experience greater power of choice on the mating market in that they set higher ideal standards, better fulfill their preferences in choice, and pair with higher mate value partners. Furthermore, we find that this cross-culturally universal pattern of mate choice is most consistent with a Euclidean model of mate preference integration.

IUL), CIS-IUL, Lisboa, 1649-026, Portugal. 29  Agents in six of the primary agent-based models paired based on a mutual attraction 102 model of mate choice. In these models, the attraction matrices for each sex were multiplied 103 together elementwise and pairing began with the most mutually attracted pair. This market 104 structure produced strong correspondence between the agent-based models and the human data. 105 However, to test the robustness of these results, we also ran a separate set of models with a 106 different mating market structure.

107
In this alternative model, agents paired based on minimum, rather than mutual attraction.

108
In these models, agents computed their attraction to all opposite-sex agents using their preference 109 integration algorithms just as in the primary models. However, the model then identified the least 110 attracted member of each possible couple rather than the mutual attraction of all possible 111 couples. The model next paired the agents with the highest minimum in-pair attraction value, 112 iterating this pairing until all possible couples were formed. These minimum attraction models 113 were identical to the mutual attraction models in all other respects. 114 We compared the results of the minimum attraction models to the human cross-cultural 115 sample using the same model training and testing procedure as in the primary agent-based 116 models. Supplementary Fig. S2 shows the results of this model comparison process. Just as in the 117 primary models, the model in which agents integrate their preferences according to a Euclidean 118 algorithm provides the strongest fit to the cross-cultural human sample among the six alternative 119 models of mate preference integration. The results of the primary agent-based models, in which 120 the Euclidean algorithm produces the best approximation of the cross-cultural human data, are 121 therefore not limited to the mutual attraction model of mate choice.

123
Agents in each of the primary models conduct a complete search of their local mating 124 market: each agent has information on and ultimately selects from the total set of 100 potential 125 mates that exist in their population. While this number is within estimated limits on human social 126 group sizes, this simplified population structure constitutes a potentially unrealistic assumption 127 for at least three reasons. First, this implicitly supposes that all individuals in the population have 128 perfectly overlapping social networks. Second, this population structure assumes that there is no 129 randomness within and no limitations on the mate search process. Third, this is a large set of 130 potential mates to consider, which may be computationally implausible. To assess whether the 131 results reported in the primary agent-based models are dependent on this assumption, we created 132 an alternative model in which mate search is incomplete. networks in that each agent functionally "knows" just a random subset of the total population. 140 We compared the populations produced by these incomplete search models to the human 141 cross-cultural sample using the same training and testing procedure as used for the primary The primary agent-based models showed a strong correspondence between the Euclidean 153 model and the human cross-cultural data. However, a limitation of the human sample is that all 154 data is self-report: participants reported both their own traits and preferences as well as the traits 155 of their partners, if applicable. It is possible that this led to biased reports of mates, yielding 156 biased results. 157 We addressed this problem in two ways. First, the preference-updating model allowed us 3). This suggests that rating bias alone cannot account for the correspondence between the 164 Euclidean agent-based model and the human cross-cultural sample.

165
Second and furthermore, we were able to leverage the design of this study to extract 166 partner ratings for a subset of the sample, allowing us to conduct the same tests on data that did 167 not rely exclusively on self-report. Although data collection in the cross-cultural sample was entirely self-report, and although participants were not specifically recruited in dyads, in some 169 cases participants did complete the study along with their actual romantic partner. These dyadic 170 participations were not recorded; however, we can, through participant responses, infer which 171 participants were members of dyads rather than participating alone. We used two sets of criteria 172 for inferring dyads from the cross-cultural human sample: a "strict" criterion and a "less-strict" 173 criterion. For the strict criterion, we classified two participants as belonging to a dyad if they had 174 complimentary answers on the following questions: city of residence, own age and partner age, dyads by chance, we ran the same dyad inference procedure on a sample in which we first 185 randomly scrambled the responses used to pair participants into dyads within city. On this 186 scrambled data, both dyad inference procedures produced zero inferred dyads.

187
With dyads, we can compare agent-based models to the human samples using responses 188 beyond self-report. Rather than relying on self-reports for self and partner traits, we calculated 189 composite trait scores for all participants by averaging self-and partner-reports. We then