CodePudding user response:
1, at the time of training and the establishment of the model, starting from the relatively simple model, don't do characteristics very much from the start, jump model parameters is very complicated,2, increase samples, to cover all the data type, data model training again after cleaning, to prevent noise interference model data,
3, regularization, penalty function is added in the model algorithm to prevent a fitting, commonly with L1 and L2 regularization,
4, bagging integrated learning method (such as random forests) can effectively prevent fitting,
5, reduce the number of features (not recommended)
Note: the dimension reduction cannot solve fitting, dimension reduction is only reduced the feature dimension, does not reduce the characteristics of all the information,