Understanding and longitudinally tracking the social context of people help in understanding their behavior and mental well-being better. Hence, instead of burdensome questionnaires, some studies used passive smartphone sensors to infer social context with machine learning models. However, the few studies that have been done up to date have focused on unique, situated contexts (i.e., when eating or drinking) in one or two countries, hence limiting the understanding of the inference in terms of generalization to (i) everyday life occasions and (ii) different countries. In this paper, we used a novel, large-scale, and multimodal smartphone sensing dataset with over 216K self-reports collected from over 580 participants in five countries (Mongolia, Italy, Denmark, UK, Paraguay), first to understand whether social context inference (i.e., alone or not) is feasible with sensor data, and then, to know how behavioral and country-level diversity affects the inference. We found that (i) sensor features from modalities such as activity, location, app usage, Bluetooth, and WiFi could be informative of social context; (ii) partially personalized multi-country models (trained and tested with data from all countries) and country-specific models (trained and tested within countries) achieved similar accuracies in the range of 80%-90%; and (iii) models do not generalize well to unseen countries regardless of geographic similarity.

翻译：理解和纵向追踪人们的社交情境有助于更好地把握其行为与心理健康状态。为此，部分研究摒弃繁琐问卷，转而利用被动式智能手机传感器结合机器学习模型推断社交情境。然而，现有研究多聚焦于特定场景（如用餐或饮水时）且局限于少数国家，导致对社交情境推断在（i）日常生活场合与（ii）不同国家间泛化能力的认知不足。本文利用涵盖五国（蒙古、意大利、丹麦、英国、巴拉圭）580余名参与者的21.6万份自报告构建的新型大规模多模态智能手机感知数据集，首先验证基于传感器数据的社交情境推断（如独处与否）的可行性，进而探究行为特征与国家层面多样性对推断效果的影响。研究发现：（i）来自活动、位置、应用使用、蓝牙及WiFi等模态的传感器特征可有效反映社交情境信息；（ii）部分个性化多国模型（基于所有国家数据训练与测试）与特定国家模型（在各国数据内训练与测试）在80%-90%准确率范围内表现相近；（iii）无论地理相似性如何，模型均难以泛化至未见过的国家。