From: Machine learning in TCM with natural products and molecules: current status and future perspectives
Machine learning algorithm | Advantages | Limitations | Applications | |
---|---|---|---|---|
Deep learning | CNN | High accuracy; Weight-sharing; Relieves the model overfitting problem; Shift-invariant feature enhances the robustness of the network | The pooling layer loses a lot of valuable information; Substantial hardware and dataset size requirements [41]; No memory function; Shift-invariant feature also prevents the neuron that recognizes the object from being activated when the object changes slightly | |
 | Elman RNN | Strong ability to extract time series features; Better generalization ability | Prone to show gradient exploding and gradient vanishing; Unable to solve the problem of long-term dependencies and parallel training [42] | Time series data, e.g., natural language processing [22] |
 | LMST | Achieve better analysis results in longer sequences; Solving vanishing gradient problem and stability problems in the time dimension of Elman RNN | When processing longer sequences data, LSTM is still difficult; Time-consuming [22] | For processing longer time series data than Elman RNN, such as in natural language processing [22] |
 | GAN | Generative model; Can still be used when the probability density is not calculated; Good at generalization | Unstable training process; Difficult to achieve Nash equilibrium; Not suitable for processing discrete forms of data | Data augmentation [46]; Text-to-image synthesis [26]; Image-to-image translation [26]; Computer vision [25] |
Traditional machine learning | MLP | Simple model; Easy implementation, and good generalization ability | architecture optimization; For training large datasets is very time-consuming [27] | Identification; Classification and prediction [32] |
 | SVM | Suitable for small-sample binary classification research; Good robustness and generalization ability | Sensitive to parameters and kernel function; Inappropriate for multi-classification research in non-optimized cases | Classification and Regression problems [32] |
 | DT | Simple to understand and to interpret; Requires little data preparation | Unstable for small variations in the data might result in the generation of a completely different tree | classification, and regression problems [32] |
 | RF | Simple structure; Easy to implement; Higher efficiency [37] | Unable to optimize its own parameters; Overfitting can easily occur when the amount of data is large [45] | Classification and regression problems [44] |