algo

algo

This directory contains implementations of machine learning algorithms organized into two main categories: regressors and classifiers. Additionally, it provides a neural network wrapper that offers a unified interface for deep learning models, simplifying the use of mlpack neural networks.

The main function signatures of all models follow those of mlpack. Classifiers use the Train/Classify interface, while regressors use Train/Predict. If you plan to use LCPP, it is recommended to adhere to this structure, since most of the metrics rely on this convention.

---

1. classifier

For classification tasks, the following models are available: Linear and Quadratic Discriminant classifiers (LDC, QDC), Nearest Mean and Nearest Neighbor classifiers (NMC, NNC), and the Non-linear SVM classifier (SVM). Additionally, a templated One-vs-All (OnevAll) framework is provided for multi-class classification, allowing you to train any binary classifier that follows the mlpack classifier signature. All classifiers are located in the algo::classification namespace. Example usage of OnevALL is

OnevAll<mlpack::LogisticRegression> model;
model.Train(...);
model.Classify(...);

2. regressor

For regression tasks, the following models are available: Kernel Ridge Regression (KernelRidge), Semi-Parametric Kernel Ridge (SemiParamKernelRidge), and Gaussian Process (GaussianProcess). These are implemented in the algo::regression namespace. Furthermore, a kernel smoothing function (kernelsmoothing) with internal bandwidth optimization is provided in the algo::functional namespace.

3. dimred

For dimensionality reduction, the available method is Univariate Functional Principal Component Analysis (ufpca), which can be found in the algo::functional namespace.

4. nn

For those seeking a simple training procedure with mlpack artificial neural networks, this wrapper provides task-invariant training and prediction/classification functionality. Users can easily define a neural network architecture and train it with built-in early stopping, where 2% of the training data is randomly separated for validation. This class is templated over the network type, optimizer, and loss function. It is located in the algo namespace. You can pass the created network to the model and optimizer parameters to the Train(...) function. Example usages of the ANN class for both classification and regression tasks can be found in the examples directory.

using OPT = ens::Adam;
using LOSS = mlpack::Accuracy;
using NetworkType = mlpack::FFN<mlpack::CrossEntropyError>;
using MODEL = algo::ANN<NetworkType,OPT,LOSS>;
... 
  // Define the architecture
  NetworkType network;
  network.Add<mlpack::Linear>(...);
  network.Add<mlpack::ReLU>();
  network.Add<mlpack::Linear>(...);
  network.Add<mlpack::ReLU>();
  network.Add<mlpack::Linear>(...);
  network.Add<mlpack::Softmax>();
...
  MODEL ann(network);
  ann.Train(...);
  ann.Classify(...);