When we successfully implement a machine learning algorithm and use it to solve practical problems, we often find its performance
(classification, regression accuracy) does not reach our satisfactory state. In this case, we have the following six options
To improve the performance of the current algorithm
1 increasing the number of training sets this method is suitable for cases where the model has been fitted
2 reduce the number of feature (use less feature) This method fits the model to fit
3 Increase the number of feature (use more feature) This method is suitable for the model to occur under-fitting
4 Adding polynomial feature this method is suitable for the model of under-fitting
5 reducing λ fits under-fitted models
6 increasing λ fits over fitting model
Note The above points, you can avoid due to the direction of the wrong choice caused by the time wasted.
For neural networks, the performance of complex neural networks with multiple layers/layers of nodes is often better than that of a simple neural network.
If a neural network is not fit, consider increasing the number of layers, increasing the node count per layer, and reducing λ.
Conversely, if it is over-fitting, the best option is to increase λ.
How to determine how to further improve the performance of learning algorithms