Facilitating the Explanation Consistency of Computational Models for Subjective Well-Being

Subjective well-being plays a crucial role in various psychological state analysis systems built on novel psychological computational models learned from online data. The explanation of these computational model results is beneficial to decision-making in our society. However, a series of confirmatory analyses indicate that there are previously inconsistent factor explanations generated from machine learning-based well-being computational models, even if they are trained on the same datasets and have achieved competitive performance. Simultaneously, we implement multiple computational models on two online questionnaire datasets. The results revealed that 1) there is explanation inconsistency among them, and 2) one model with high accuracy does not show high explanation consistency with other models, which causes hesitation in which explanation to accept. We delved into the underlying reasons behind these experimental results and put forward a simple solution to enhance the explanation consistency among different models. Firstly, the factor mapping decorrelates the factors, which is beneficial to reduce the spurious relations between factors and well-being. Secondly, we utilize a model-agnostic explanation method to compute the importance of factors for any computational models. Experimental results validate the 2-fold high explanation consistency between pairwise models produced by our solution. One is evidenced by models of each other, and another is additionally provided by the results of a systematic review of the research in social science. Therefore, our solution is an assistant for humans to analyze subjective well-being.