A hyperparameter is a setting chosen by the model developer before or during training that influences how a machine-learning model learns. Unlike model parameters, which training learns from data, hyperparameters define the conditions of learning. Examples include learning rate, maximum tree depth, class weights, number of epochs, and optimizer weight decay. section 3.1 section 9.2.3
Use hyperparameters to configure the training workflow:
- Select settings appropriate to the model, such as tree depth for tree-based fraud models or learning rate and epochs for neural networks. section 3.1 section 9.2.3
- Pair the model with a loss function and optimizer; the MLP example uses cross entropy, Adam, and weight decay. section 9.2.3
- Run training, validate across epochs, inspect loss and evaluation metrics, and try alternative values to observe their effects. section 9.2.3
MLP for illicit Bitcoin transactions
An MLP configuration sets NUM_EPOCHS to 1000, the learning rate to 0.01, cross-entropy loss, and Adam with weight decay of 1e-4. Training updates learned parameters; validation accuracy and F1 are tracked across epochs before final test evaluation. section 9.2.3
Figure 9.10 Steps to train and evaluate the MLP model, starting with setting model hyperparameters, running the model training loop, and evaluating the model performance on the test set.
The workflow diagram connects hyperparameter selection with the training loop and test evaluation. section 9.2.3

























