Existing models of facial first impressions indicate between two and four factors that underpin all social trait judgements. Here, we submitted several large databases of these first impression ratings to unsupervised learning algorithms with the aim of clustering together faces, rather than traits, to examine the ways in which impressions may be grouped together. Experiment 1 revealed two clusters of faces that exist in both a full-dimensional, and two- or three-factor representations, of social impressions, while Experiment 2 indicated that these clusters also emerged in additional datasets. In Experiment 3, using Bayesian modelling approaches, we extracted the impression profile of each cluster and also derived a vector that maximally separated the clusters. The resulting vector related strongly to the valence and approachability components in existing models. In a further test of our model, we showed in Experiment 4 that mere facial appearance, rather than perceptions, is sufficient to separate these clusters, demonstrating probabilistically that facial cues like smiling may drive the perceptual profile that gives rise to the perceptual clusters. Finally, Experiment 5 showed that observer responses to faces in these two clusters mapped closely on to approach-avoidance behaviour, with observers responding rapidly and without instruction to approach faces from one cluster over the other. Taken together, our findings provide compelling evidence, drawing upon both computational and behavioural approaches, that existing models of social impressions are realised practically in terms of basic approach-avoidance mechanisms.

现有的面部第一印象模型表明，支持所有社会特征判断的因素有两到四个。在这里，我们将这些第一印象评级的几个大型数据库提交给无监督学习算法，目的是将面部而不是特征聚类在一起，以检查印象可以组合在一起的方式。实验 1 揭示了两个人脸集群，它们同时存在于社会印象的全维和两或三因素表示中，而实验 2 表明这些集群也出现在其他数据集中。在实验 3 中，使用贝叶斯建模方法，我们提取了每个集群的印象概况，并导出了一个向量，最大限度地分离了集群。得到的向量与现有模型中的效价和可接近性组件密切相关。在对我们模型的进一步测试中，我们在实验 4 中表明，仅仅面部外观而不是感知就足以分离这些集群，从概率上证明微笑等面部线索可能会驱动产生感知集群的感知轮廓。最后，实验 5 表明，观察者对这两个集群中的面孔的反应与回避行为密切相关，观察者反应迅速，并且没有指示从一个集群接近另一个集群的面孔。总之，我们的发现提供了令人信服的证据，同时借鉴了计算和行为方法，