社会与工程心理学研究室知识库

了解个体心理特征、主观感受对心理学研究、社会管理等具有重要意义。传统心理测量手段通常采用问卷、访谈等方式进行，受施测方法、环境等因素制约，在时效J比、施测规模等方面具有一定的局限性。

随着互联网的蓬勃发展，社会媒体在人们的日常生活中得到了广泛地使用。由于用户发布的这些信息是其日常生活行为的记录，具有内容丰富、自然发生、规模庞大、数据记录可回溯等特点，使得社会媒体平台成为了理想的“在线心理学实验室”。通过对用户的在线行为记录进行分析，可以实时、大规模地了解个体心理特征、主观感受，弥补了传统心理测量手段的不足。

本文将机器学习方法应用于个体心理特征测量这一领域，提出了基于社会媒体用户网络行为数据，对个体进行非侵扰式生态瞬时心理特征评估计算的方法。研究方法采用有监督的机器学习方法:首先收集了具有心理特征标注的微博用户数据，心理特征标注采用心理学量表获得;然后针对社会媒体的功能，从考察用户社会媒体使用基本行为、语言特征和时间序列特征的角度，建立了用户行为特征体系;进而基于用户行为的特征数据及与之对应的心理特征标注，利用有监督机器学习算法，建立心理特征分类、回归预测模型，并对模型的有效性进行了验证。

我们在研究中开展了基于微博的用户实验，利用上述特征体系和建模方法，建立并评估了基于社会媒体的用户人格、心理健康、主观幸福感、自杀风险的预测计算模型。模型在人格、心理健康等心理特征高低分组分类问题上准确率约为80%;在人格、心理健康、主观幸福感心理特征得分预测问题上，模型预测结果与心理问卷结果的皮尔逊积矩相关系数可达0.4~0.6;在高自杀风险用户分类预测问题上，相较于人工检测，模型可以协助减少约40%的人工工作量。

本文研究结果表明，运用机器学习方法建立的心理特征预测模型，可以利用用户在社会媒体上的行为数据，开展对个体心理特征的大规模、实时预测计算，预测计算结果的有效性与心理学量表测评或他评相当。本文探索了基于互联网行为数据进行个体心理特征预测的方法论，研究结果可以为后续相关领域研究提供方法上的借鉴，研究过程中建立的模型也可进行应用和推广，为心理学研究和社会管理提供服务。

Measuring individual psychological features and subjective feeling is of great significance to psychological research, social management and so on. Conventional psychologicaI testing approach usually employs survey methods like questionnaires or interviews. Restricted by testing methods and environmental factors，this approach has certain limitations in timeliness, sample size etc.

With the rapid development of the Internet, social media has been widely used in people's daily life. Records on social media posted or shared by users in daily life, are of rich content, naturally occurring, large scale and traceable. Therefore, the social media platform has become an ideal "online psychological laboratory". Through the analysis of users' online behavioral records, it is feasible to measure individual psychological features or subjective feelings in real-time and large-scale, which can make up for deficiencies of the conventional psychological measurement methods.

In this paper, we propose an approach employing machine learning methods to measure individual psychological features in a non-intrusive and ecological manner. The framework of this approach is to adopt supervised machine learning methods on social media users' psychological features labeled online behavioral records data.

First, ，We collected users' microblog data labeled with psychological features, in which the latter are measured by psychological scales. Second, based on social media features ，we designed users' online behavioral feature ontology, which describes users' online behavioral patterns in social media usage, linguistic expression and time sequence patterns. Then; we employed supervised machine learning algorithms on users' online behavioral features data extracted according to aforementioned feature ontology as input variables, and psychological features labels as prediction variables. Classification and regression prediction models for psychological feature were then established; and we also evaluated the validity of these prediction models.

this study we carried out experiments on Sina microblog. By extracting features from users' labeled microblog data and employ machine learning algorithms, we established prediction models to measure social media users' personality, mental health status, subjective well-being and suicide risk. Classification models for individual personality and mental health status achieve accuracy around 80%. Regression models to predict scores of users' personality, subjective well-being, mental health status; are correlated with scores measured by psychological scales with a Pearson's Correlation Coefficient about 0.4~0.6. Models to identify high suicide risk users, comparing to manual identifying, can help to reduce the manual workload of about 40%.

This study demonstrates that, based on web data analysis, it is possible to efficiently predict psychological features and to update the predicted outcomes in real time. The predicted results has the validity equivalent to psychological scales or judgmenu made by human. This paper explores the methodology for identifying individual psychological features based on online behavioral records. The research results can provide a. methodological reference for further research in related fields, and models established in this research can also be carried out in application to provide services for psychological research and social management.