Social interaction sequences form relationship trajectories along abstract dimensions of affiliation and power.

(A) An example of a power interaction. Participants read text that describes the narrative and on decision trials choose between two options. Based on their choice, the character moves −1 or +1 along the active dimension (affiliation or power). (B) The participant forms relationships with different characters through sequences of interactions in the narrative. (C) The decisions the participant makes in the affiliation and power interactions change the character’s location in social space, forming a relationship trajectory. Note that in the task, participants interact with 6 characters: 5 each with 6 affiliation and 6 power trials and 1 with 3 neutral trials.

Behavioral geometry is consistent with social mapping.

(A) Schematic of post-task subjective placements as compared behavioral maps, the affiliation and power coordinates calculated from decisions (shown faded). The behavioral locations were not shown to participants during the placements. The mapping error between the behavioral and subjective placements was calculated as the average character-wise Euclidean distance between the locations (shown as arrows). (B) Mapping error is smaller than permutation-based chance, suggesting the behavioral modeling captures elements of these subjective placements. (C) The amount of mapping error negatively correlates with task memory, suggesting the subjective maps depend on memory. 95% confidence intervals for regression line are indicated by the shaded region and p-value significance is indicated by asterisks: * < 0.05, *** < 0.005. Validation sample only (n = 32).

The left hippocampus represents affiliation and power interactions abstractly.

Dimensional abstraction: trials of the same dimension should have more similar patterns than trials of different dimensions. Representational similarity analysis searchlight results (pFWER < 0.05, with small volume correction in left hippocampus). HC = hippocampus, AG = angular gyrus, MTG = middle temporal gyrus. Combined sample (n = 50).

Trajectory location decoding analysis.

(A) Behavioral trajectories for each character were analyzed; one simulated behavioral trajectory is highlighted in tan. A leave one trial out approach was used; the held-out trial in this example is shown in red. (B) The hippocampal patterns were isolated in the same way. A simulated neural sequence associated with the behavioral trajectory is shown. We estimated the location of this held-out point on the -based approximation of the neural trajectory. Using a parameterization of the spline, we then decoded the held-out trial’s social location and calculated the distance of this estimate to the actual task location to estimate decoding errors. (C) The average error for the different trajectory models in the left hippocampus for both samples. The hypothesized model was compared to multiple null models: a linear embedding model to test the importance of nonlinear dimensionality reduction in preserving the neural sequence (“linear embedding”); a dummy model where only the trajectory midpoint was predicted (“trajectory midpoint”); a model where participant choices were shuffled within each relationship (“shuffled trajectories”); a model where participant choices were pseudo-randomly selected, across characters, to calculate locations (“pseudo trajectories”); a model where random choices were simulated to create trajectories that respected the task structure but did not have choice history (“random trajectories”); and a model that permuted brain-behavior relationships but preserved the temporal autocorrelation (“chance”). P-value significance are indicated by asterisks: ∼* < 0.1,* < 0.05, ** < 0.01, *** < 0.005, **** < 0.001.

The number of distinct left hippocampal clusters correlates with real-world social network size.

We estimated the number of distinct left hippocampal clusters that optimized the overlap with behavioral clusters, and then regressed these estimates onto estimates of social network size. None of the data points exceed an outlier threshold of 3 standard deviations from the mean. 95% confidence intervals are indicated by the shaded area and p-value significance is indicated by the asterisk: * < 0.05. Validation sample only (n = 32).

All trajectory models were evaluated on embedding dimensions 2 to 10. (A) Top plot is how much better the ‘real’ model does than the null. Larger values mean smaller errors for the real model compared to the null model. (B) Bottom plot is the comparisons for dimensions 3 to 10 against the previous dimension’s decoding, for each null. Larger values mean a larger improvement over the null from the previous dimension. For both plots, the mean error difference is shown along with 95% confidence intervals.