Skip to content

Dimlp algorithms

The Discretized Interpretable Multi-Layer Perceptron (DIMLP) is a specialized feed-forward neural network architecture, derived from the traditional MLP (Multi-Layer Perceptron). DIMLP performs predictive tasks and generates interpretable decision rules that explain the underlying reasoning behind the model's predictions. The DIMLP framework includes a set of algorithms that leverage this capability for model training, evaluation, and rule extraction. To get more details on the Dimlp algorithm, you can refer to this paper, and to this one for dimlpBT.

Architecture

The architecture is built as shown below:

graph TD;
    DimlpTrn[DimlpTrn] 
    DimlpBT[DimlpBT] 
    DimlpPred[DimlpPred] 
    DimlpCls[DimlpCls]
    DimlpRul[DimlpRul]
    DensCls[DensCls]

    d(Dimlp algorithms) --> DimlpTrn;
    d --> DimlpBT;
    DimlpTrn --> DimlpPred;
    DimlpTrn --> DimlpCls;
    DimlpTrn --> DimlpRul;
    DimlpBT --> DensCls;

Each algorithm has its purpose:

  • DimlpTrn: Trains the Dimlp model using a training dataset, obtains train/test/validation predictions and model weights, and can optionally extract global rules using the Dimlp algorithm.
  • DimlpPred: Generates predictions from the trained Dimlp model on a test dataset.
  • DimlpCls: Calculates accuracy, generates predictions, and retrieves the values of the first hidden layer from the trained Dimlp model on a test dataset.
  • DimlpRul: Generates global explanation rules using the Dimlp algorithm on the training dataset used to train the Dimlp model, and retrieves training, testing, and validation accuracy, if provided.
  • DimlpBT: Trains the Dimlp model using a training dataset with bagging, obtains train/test/validation predictions and model weights, and can optionally extract global rules using the Dimlp algorithm.
  • DensCls: Generates global explanation rules using the Dimlp algorithm and obtains train and test predictions and accuracy from a Dimlp model trained with bagging.