Network structure and the optimisation of proximity-based association criteria

Animal social network analysis (SNA) often uses proximity data obtained from automated tracking of individuals. Identifying associations based on proximity requires deciding on quantitative criteria such as the maximum distance or the longest time interval between visits of different individuals to still consider them associated. These quantitative criteria are not easily chosen based on a priori biological arguments alone.
Here we propose a procedure for optimising proximity-based association criteria in SNA, whereby different spatial and temporal criteria are screened to determine which combination detects more network structure. If we assume that biologically-relevant associations among individuals are non-random, and that proximity data are mostly influenced by those associations, then it is logical to select criteria that minimise random associations and show the underlying network structure more clearly.
We first used simulations to evaluate which of four simple descriptors of network structure remain unbiased (i.e., do not change directionally) when reducing the number of observations, since unbiased descriptors are necessary for comparing the structure of networks using different association criteria. Then, using two of those descriptors (coefficient of variation of the strength of associations, and network entropy), and empirical proximity data from automated tracking of common waxbills (Estrilda astrild) in a mesocosm environment, we found that the structure-based optimisation procedure selected the most biologically-relevant combination of spatial and temporal proximity criteria, in the sense that those criteria were also the best at distinguishing between previously known social sub-groups of individuals.
These results indicate that, provided that the assumptions for structure-based optimisation are met, this procedure can find the most biologically-relevant association criteria. Thus, under the condition that proximity data are shaped by non-random social associations, and if using adequate descriptors of network structure, structure-based optimisation may be a useful tool for SNA, particularly when a priori biological arguments are insufficient to inform the choice of proximity-based association criteria.

动物社交网络分析（Social Network Analysis，SNA）通常采用对个体进行自动追踪所获得的近距离接触数据开展研究。基于近距离接触识别个体间的社交关联时，需要设定量化判定标准，例如以最大距离或不同个体到访的最长时间间隔作为关联判定阈值。仅依靠先验生物学依据，往往难以确定这类量化标准。

本研究提出一种用于优化SNA中基于近距离接触的关联判定标准的流程，通过筛选不同空间与时间维度的判定标准，以确定哪种组合能更清晰地还原网络结构。我们假设：个体间具有生物学意义的关联并非随机产生，且近距离接触数据主要受此类关联影响；据此，选择能够最大程度弱化随机关联、更清晰地呈现潜在网络结构的标准，便具备逻辑合理性。

本研究首先通过模拟实验，评估在减少观测样本量时，四种网络结构简单描述指标中哪些能够保持无偏性（即不会出现方向性变化）——由于无偏描述指标是采用不同关联标准对比网络结构的必要前提。随后，我们选取其中两种描述指标（关联强度变异系数与网络熵），结合在中宇宙环境（mesocosm environment）中对普通梅花雀（Estrilda astrild）进行自动追踪所获得的实证近距离接触数据开展分析，结果显示：基于结构优化的流程筛选出的空间与时间维度近距离接触标准组合，最具生物学意义；具体而言，该组合同样最擅长区分此前已被证实的个体社交亚群。

上述结果表明，只要满足基于结构优化的前提假设，该流程即可筛选出最具生物学意义的关联判定标准。由此可见，当近距离接触数据由非随机的社交关联所塑造，且采用恰当的网络结构描述指标时，基于结构优化的方法可成为SNA的实用工具，尤其适用于仅依靠先验生物学依据不足以确定近距离接触关联标准的场景。