Design a data augmentation step that will introduce acceptable variance into your training data. For example, if your prediction task is to predict the selling price of a used car based on its features (make, model, year, mileage, condition, etc.), you could sample your training data and add +-1 years randomly for the model year. This accounts for potential misreporting or rounding errors in the year variable. Add this augmented data into your training dataset, and hopefully it improves generalization. The key is to understand the problem domain, and apply that knowledge into modelling.

When the size of the dataset is limited, one way to improve the training stage is to increase the iteration and the K-fold number of your cross-validation. The disadvantage will be a higher computing time. A rule of thumb is to use 10 folds, you can for example reiterate this cross-validation ten times to ensure that the learning is performed on all your data.