Incremental supervised learning
This section introduces several techniques used to learn from stream data when the true label for each instance is available. In particular, we present linear, non-linear, and ensemble-based algorithms adapted to incremental learning, as well as methods required in the evaluation and validation of these models, keeping in mind that learning is constrained by limits on memory and CPU time.
Modeling techniques
The modeling techniques are divided into linear algorithms, non-linear algorithms, and ensemble methods.
Linear algorithms
The linear methods described here require little to no adaptation to handle stream data.
Online linear models with loss functions
Different loss functions such as hinge, logistic, and squared error can be used in this algorithm.
Inputs and outputs
Only numeric features are used in these methods. The choice of loss function l and learning rate λ at which to apply the weight updates are taken as input parameters. The output is typically updatable...
