Recently met a problem, when I was little training data, the trained model parameters of each is different, so the prediction accuracy is changing all the time, when I was training sets is larger, the output becomes stable, this is a normal phenomenon? My model no super parameters, training process does not introduce random variables, My understanding is that for certain model of LDA, for example, when the training set is small, we can get a lot of kinds of the possibility of a linear model, so cause the different model parameters, don't know whether this explanation is reasonable
CodePudding user response:
Data set is too small for some model: the first model easy to fitting, the model not only easily appeared on training set fitting, may also appeared on validation set fitting; The abnormal values, the second will outliers may appear in the feature, may also occur in response to the variable, outliers may also lead to inconsistent training sample and test sample data distribution; The third is difficult to optimize the model, all these reasons will reduce the stability of the model, but some model suitable for small sample,
