Biomedical Data Science Lab. Instituto de Aplicaciones de las Tecnologías de la Información y de las Comunicaciones Avanzadas (ITACA), Universitat Politécnica de Valéncia, Camino de Vera s/n, Valencia 46022, Spain.
Instituto Universitario de Matemática Pura y Aplicada, Universitat Politécnica de Valéncia, Camino de Vera s/n, Valencia 46022, Spain; Departamento de Psicología Evolutiva y de la Educación, Universitat de Valéncia, Avenida Blasco Ibáñez, 21, Valencia 46010, Spain.
Comput Methods Programs Biomed. 2019 Jan;168:59-68. doi: 10.1016/j.cmpb.2017.11.004. Epub 2017 Nov 13.
Happiness is a universal fundamental human goal. Since the emergence of Positive Psychology, a major focus in psychological research has been to study the role of certain factors in the prediction of happiness. The conventional methodologies are based on linear relationships, such as the commonly used Multivariate Linear Regression (MLR), which may suffer from the lack of representative capacity to the varied psychological features. Using Deep Neural Networks (DNN), we define a Happiness Degree Predictor (H-DP) based on the answers to five psychometric standardized questionnaires.
A Data-Structure driven architecture for DNNs (D-SDNN) is proposed for defining a HDP in which the network architecture enables the conceptual interpretation of psychological factors associated to happiness. Four different neural network configurations have been tested, varying the number of neurons and the presence or absence of bias in the hidden layers. Two metrics for evaluating the influence of conceptual dimensions have been defined and computed: one quantifies the influence weight of the conceptual dimension in absolute terms and the other one pinpoints the direction (positive or negative) of the influence.
A cross-sectional survey targeting non-institutionalized adult population residing in Spain was completed by 823 cases. The total of 111 elements of the survey are grouped by socio-demographic data and by five psychometric scales (Brief COPE Inventory, EPQR-A, GHQ-28, MOS-SSS and SDHS) measuring several psychological factors acting one as the outcome (SDHS) and the four others as predictors.
Our D-SDNN approach provided a better outcome (MSE: 1.46·10) than MLR (MSE: 2.30·10), hence improving by 37% the predictive accuracy, and allowing to simulate the conceptual structure.
We observe a better performance of Deep Neural Networks (DNN) with respect to traditional methodologies. This demonstrates its capability to capture the conceptual structure for predicting happiness degree through psychological variables assessed by standardized questionnaires. It also permits to estimate the influence of each factor on the outcome without assuming a linear relationship.
幸福是人类普遍的基本目标。自积极心理学出现以来，心理学研究的一个主要焦点一直是研究某些因素在预测幸福感方面的作用。传统的方法是基于线性关系，如常用的多元线性回归（MLR），它可能由于缺乏对各种心理特征的代表性而受到限制。使用深度神经网络（DNN），我们基于五个心理测量标准化问卷的答案定义了一个幸福度预测器（H-DP）。
我们提出了一种基于数据结构的 DNN 架构（D-SDNN）来定义 HDP，其中网络架构能够对与幸福相关的心理因素进行概念解释。我们测试了四种不同的神经网络配置，改变了隐藏层中神经元的数量和有无偏差。我们定义并计算了两种用于评估概念维度影响的指标：一种以绝对值量化概念维度的影响权重，另一种则确定影响的方向（正或负）。
一项针对居住在西班牙的非机构化成年人口的横断面调查完成了 823 例。调查的总共 111 个元素按社会人口统计学数据和五个心理测量量表（Brief COPE 库存、EPQR-A、GHQ-28、MOS-SSS 和 SDHS）分组，这些量表测量了几种作为结果的心理因素（SDHS）和其他四个作为预测因素。
我们的 D-SDNN 方法提供了更好的结果（MSE：1.46·10），优于多元线性回归（MSE：2.30·10），因此提高了 37%的预测准确性，并允许模拟概念结构。
与传统方法相比，我们观察到深度神经网络（DNN）的性能更好。这证明了它通过标准化问卷评估的心理变量来捕捉预测幸福感的概念结构的能力。它还允许在不假设线性关系的情况下估计每个因素对结果的影响。
