粗糙集和神经网络在心理测量中的应用

Application of Rough Set and Neural Networks in Psychological Measurement

探讨当因素分析和多元回归方法的使用条件未得到满足时，是否可采用粗糙集方法进行观察变量的精简，以及是否可采用神经网络方法进行预测效度检验。理论分析了粗糙集和神经网络在心理测量中应用的可能性，并运用粗糙集对于人事干部胜任力评估数据进行分析，比较了7种离散化方法和2种约简算法构成的14种组合，发现当采用Manual方法进行离散化、遗传算法进行约简时，能够很好地对观测变量进行精简；运用概率神经网络能够比等级回归方法更好地进行预测效度检验。研究结果表明对于处理心理测量中的非等距变量，粗糙集和神经网络是非常有用的方法。

Factor analysis has often been used to simplify and combine (reduce) observed variables in the construction of psychological measurement instruments， and multiple regression has frequently been used in validity studies. One major challenge， however， is that all these variables have to satisfy the requirements for interval scales. Otherwise，new statistical procedures have to be developed. In artificial intelligence， rough set can be used to reduce attributes and artificial neural networks can also be used to set up relations among the variables in a model. In this study， it was posited that these methods could also be similarly applied to psychological measurement. Under the rough set theory， there are different discretization methods and algorithms to reduce attributes， it is necessary to find the most suitable way for psychological measurement and to see whether the neural network methodology is better than that of multiple regression in validity studies. Responses data on the Competency Assessment Questionnaire of 718 civil servants who were not leaders working in a human resource section were obtained. The data were constructed into a rough set decision table with 21 condition attributes and 1 decision attribute. Seven methods， which included Boolean reasoning， Manual， Entropy/MDL， Equal frequency binning， Naive， Semi-naive and SOM， were used in the discretization. A total of 14 different reducts were obtained with the genetic algorithm and the Johnson algorithm. These were checked in validity studies with the probability neural networks method and the hierarchical regression method. The results showed that reduct D21 was the best among the 14 reducts， which used the Manual method in discretization and the genetic algorithm in reduction. Its correct classification rate was 89.30. In contrast， the average correct classification rate with the probability neural networks was 82.92 which was still higher than the 71.93 of the hierarchical regression method. The research showed the following findings. (i) In the construction of psychological measurement instruments， if the assumptions to apply factor analysis cannot be satisfied， the rough set methods can be used to reduce observed variables. (ii) The rough set provided several reducts. The best of them could be obtained through the neural networks in validity studies. (iii) In predictive validity studies， a higher classification rate could be obtained with the probability neural networks than with the hierarchical regression method. (iv) Theoretical analysis and empirical study showed that the rough set and neural networks could be applied in psychological measurement.