Eigenbehaviors: identifying structure in routine
Eigenbehaviors: identifying structure in routine
复制标题
DOI：
10.1007/s00265-009-0739-0
复制
发表时间：
2009-05-01
影响因子：
2.3
通讯作者：
Pentland, Alex Sandy
中科院分区：
文献类型：
--
作者：
Eagle, Nathan;Pentland, Alex Sandy
纵向行为数据通常包含大量的结构。在这项工作中，我们确定了日常行为中固有的结构，这些模型可以准确地分析，预测和聚类来自人群社交网络中的个人和社区的多模态数据。我们用完整行为数据集的主成分来表示这种行为结构，这是一组我们称之为特征行为的特征向量。在我们的模型中，一个人在特定一天的行为可以近似为他或她的主要特征行为的加权和。当这些权重在一天的中途计算时，它们可以用来预测我们的测试对象当天剩余的行为，准确率为79%。此外，我们证明了这种降维技术的潜力，推断社区隶属关系内的主题的社会网络聚类个人到一个“行为空间”跨越一组他们的聚合eigenbehaviors。这些行为空间可以确定个人和群体之间的行为相似性，使人口级社交网络中社区隶属关系的分类准确率达到96%。此外，行为空间中个体之间的距离可以用作关系纽带（如友谊）的估计，表明受试者之间存在强烈的行为同质性。这种方法利用了之前在现实挖掘研究中捕获的大量丰富数据，这些数据来自移动的手机，这些手机在9个月的时间里连续记录麻省理工学院100名受试者的位置、附近的电话和通信。随着可穿戴传感器不断生成这些类型的丰富的纵向数据集，特征行为等降维技术将在行为研究中发挥越来越重要的作用。
Longitudinal behavioral data generally contains a significant amount of structure. In this work, we identify the structure inherent in daily behavior with models that can accurately analyze, predict, and cluster multimodal data from individuals and communities within the social network of a population. We represent this behavioral structure by the principal components of the complete behavioral dataset, a set of characteristic vectors we have termed eigenbehaviors. In our model, an individual's behavior over a specific day can be approximated by a weighted sum of his or her primary eigenbehaviors. When these weights are calculated halfway through a day, they can be used to predict the day's remaining behaviors with 79% accuracy for our test subjects. Additionally, we demonstrate the potential for this dimensionality reduction technique to infer community affiliations within the subjects' social network by clustering individuals into a "behavior space" spanned by a set of their aggregate eigenbehaviors. These behavior spaces make it possible to determine the behavioral similarity between both individuals and groups, enabling 96% classification accuracy of community affiliations within the population-level social network. Additionally, the distance between individuals in the behavior space can be used as an estimate for relational ties such as friendship, suggesting strong behavioral homophily amongst the subjects. This approach capitalizes on the large amount of rich data previously captured during the Reality Mining study from mobile phones continuously logging location, proximate phones, and communication of 100 subjects at MIT over the course of 9 months. As wearable sensors continue to generate these types of rich, longitudinal datasets, dimensionality reduction techniques such as eigenbehaviors will play an increasingly important role in behavioral research.