In today's society, happiness has attracted more and more attentions from researchers. It is interesting to study happiness from the perspective of data mining. In psychology domain, the application of data mining gradually becomes widespread and popular, which works from a novel data-driven viewpoint. Current researches in machine learning, especially in deep learning provide new research methods for traditional psychology research and bring new ideas. This paper presents an empirical study of learning based happiness predicition approaches and their prediction quality. Conducted on the data provided by the "China Comprehensive Social Survey (CGSS)" project, we report the experimental results of happiness prediction and explore the influencing factors of happiness. According to the four stages of factor analysis, feature engineering, model establishment and evaluation, this paper analyzes the factors affecting happiness and studies the effect of different ensembles for happiness prediction. Through experimental results, it is found that social attitudes (fairness), family variables (family capital), and individual variables (mental health, socioeconomic status, and social rank) have greater impacts on happiness than others. Moreover, among the happiness prediction models established by these five features, boosting shows the most effective in model fusion.

Keywords：Happiness prediction; factor analysis; machine learning; model fusion