社会与工程心理学研究室知识库

随着过去三年疫情影响和企业降本增效的大趋势下，企业开始思考业务场景 数字化转型的嵌入，这其中，人力资源部门特别是招聘业务受到环境波及和被迫 转型势头最为猛烈——以线下面试为主的招聘方式受到诸多限制，因此促使许多企业不得不思考将面试转移至电话面试或是远程线上面试等方式推进；同时以往传统的招聘方式如专家评价和量表分析不仅依赖于过往胜任力的经验总结，也容 易受主试主观评价和经验干扰，导致容易损失或用错人才。而新研究表明，一方面个体胜任力与其面部活动变化具有相关性；另一方面机器学习方法在面部特征识别方面拥有充分的实践基础，因此为我们使用机器学习来对面试者的面部特征 进行分析处理从而识别和预测胜任力方法提供了理论依据。

本课题正文主要分为三个部分，第一部分作者分析了个体的面部活动表现和当下的心理状态、胜任力之间有一定的相关性，同时通过大量的文献考证从情绪 心理学方面来证实个体面部活动和个体胜任力之间的确存在联系；第二部分作者主要研究的是基于被试在招聘过程中的面部活动单元(面部AU)与经由专家提供的被试胜任力评分等级之间的映射关系来建立胜任力等级预测模型；第三部分则通过被试在招聘过程中的面部活动单元(面部AU)与和被试在量表中的胜任力分值之间的映射关系来构建胜任力评分预测模型，从而为人才招聘提供更符合当下社会实情和机器学习技术辅助的人才识别模型。

本研究主要使用一系列数据挖掘方式和机器学习算法模型，并基于模型预测数据与真实数据进行多种效果评估，包括网格搜索、参数循环、XGB树算法、随机森林算法、支持向量回归机(SVR)、主成分分析法(PCA)、奇异值分解(SVD)、五折交叉验证、混淆矩阵等。最终实验结果表明，使用PCA降维和网格搜索调参方式的XGBoost三分类等级评分识别模型的准确率分数为0.61，预测性能较为可观；使用PCA降维和网格搜索调参方式的SVR胜任力评分预测模型胜任力10个维度的R2表现均超过0.6，其中成就导向得分、分析力得分、开放创新得分、推进执行得分和性格潜质得分预测模型的R2 分值最接近1，表示预测值与真实值的拟合程度表现最好；皮尔逊相关系数在0.7 - 0.9之间，达到强相关，整体模型预测评分效果很好，由此表明基于招聘中的面部活动单元构建胜任力预测模型具有一定的可行性。本课题借助先进的科学技术对人力资源面试的胜任力识别进行深入的探索，在基本心理理论框架的支撑下，具有较为可观的商业应用前景和延展空间。

In light of the prevailing trends of the past three years, influenced by the pandemic and the drive for cost reduction and efficiency enhancement in enterprises, there has been a growing consideration for the integration of digital transformation into business scenarios. Particularly, the HR department, especially the recruitment function, has been significantly impacted by the environmental changes and forced transformation. The traditional recruitment methods relying on offline interviews have faced limitations, leading many companies to shift towards telephone or remote online interviews. Moreover, conventional recruitment approaches such as expert evaluations and scale analyses not only depend on past job competency experiences but are also prone to subjective evaluations and experiential biases, resulting in potential talent loss or misplacement. Recent studies have shown a correlation between individual competencies and facial actions. Additionally, machine learning techniques have a solid foundation in facial feature recognition, providing a theoretical basis for utilizing machine learning to analyze and predict competencies based on interviewees' facial features.

The main body of this study is divided into three parts. Firstly, the author analyzes the correlation between individual facial actions, current psychological states, and competencies, supported by extensive literature review in emotional psychology to verify the connection between facial actions and competencies. Secondly, the study focuses on mapping the facial action units (AUs) of the subjects during the recruitment process to the competency rating provided by experts to establish a competency level prediction model. Thirdly, the mapping relationship between the facial action units (AUs) of the subjects during the recruitment process and the competency score on the scale is examined to construct a competency score prediction model, aiming to provide a talent identification model that aligns with current societal realities and leverages machine learning technology.

The research primarily employs a series of data mining techniques and machine learning algorithm models, evaluating the predictive models against real data using various methods including grid search, parameter tuning, XGBoost tree algorithm, random forest algorithm, support vector regression (SVR), principal component analysis (PCA), singular value decomposition (SVD), five-fold cross-validation, confusion matrix, among others. The experimental results demonstrate that the XGBoost three-class rating recognition model using PCA dimensionality reduction and grid search for parameter tuning achieves an accuracy score of 0.61, indicating a promising predictive performance. The SVR competency score prediction model using PCA dimensionality reduction and grid search for parameter tuning consistently achieves an R2 performance exceeding 0.6 across ten competency dimensions. Notably, the prediction models for achievement orientation score, analytical thinking score, openness to innovation score, driving execution score, and personality potential score show an R2 value closest to 1, indicating the best fit between predicted and actual values. The Pearson correlation coefficients fall within the range of 0.7 to 0.9, indicating a strong correlation, thus demonstrating a good overall predictive scoring effect of the models. Therefore, constructing competency prediction models based on facial action units in video interviews appears to be feasible. This study embarks on an in-depth exploration of competency recognition in human resource interviews using advanced scientific technologies. Supported by a fundamental psychological theoretical framework, it holds promising commercial application prospects and expansion opportunities.