| In today's big data age,semi-supervised learning plays an important role in many applications as a mainstream technology for learning with unlabeled data.However,the theoretical research and empirical analysis of traditional semi-supervised learning methods are mostly carried out under the assumption of random missing.That is,the labeled and unlabeled samples are required to be from the same distribution,but this is difficult to meet in real life.For example,in the issue of housing credit loans by commercial banks,the differences in culture and customs in different regions are likely to make a big difference in personal housing credit situation.For another example,the credit conditions of people at different periods may also show differences with the change of National policies or personal family income.In this case,if the traditional semi-supervised learning method is used for training,it may cause great errors.This paper first summarizes the traditional semi-supervised learning methods,analyzes the advantages and disadvantages of each method and Its application occasions,and then proposes a new semi-supervised learning method to address the above shortcomings of traditional semi-supervised learning methods-Semi-supervised learning method based on generalized additive model and Group Lasso penalty.As with the traditional semi-supervised learning method,this method uses both labeled and unlabeled samples and imposes Group Lasso penalties on them.Finally,the block coordinate descent algorithm is used to simultaneously estimate the parameters and select the variables.This paper proposed model description and detailed algorithm derivation for the new semi-supervise learning method,and on this basis,considering the various relationship among the data in life,we do simulations in a variety of situations for our semi-supervised learning method.The classification error,true positive rate and false discovery rate were used as indicators to compare our semi-supervised learning methods with the supervised Group Lasso model in predictive ability and variable selection ability.The simulation results show that on the whole,the new semi-supervised learning methods we proposed are significantly better than the supervised Group Lasso model in all aspects,and this advantage is most obvious especially in the case when the real functions of unlabeled samples are different with the labeled samples.Finally,this paper studies the application of this method in the field of housing credit risk assessment.The results show that our new semi-supervised learning method is superior in variable selection,result prediction and model robustness when there are irrelevant variables and the labeled and unlabeled samples come from different distributions,and they are better than the supervised semi-parametric Group Lasso and ordinary Group Lasso model. |
