Meta-learning Algorithms

Learning approaches that rely on another algorithm performing some enrichment, variation, or generalization.

---

OneVsAllLearning

Java class: OneVsAllLearning

Source code: OneVsAllLearning.java

Maven Project: kelp-core

JSON type: oneVsAll

Description: it is a meta algorithm that operates applying a One-Vs-All strategy over the base learning algorithm which is intended to be a binary learner. The One-Vs-All strategy will learn N different classifiers, where N is the number of classes involved in the dataset. In this strategy each classifier is learned by considering in turn the examples of a single class as positives, while all the other examples are considered as negative.

---

OneVsOneLearning

Java class: OneVsOneLearning

Source code: OneVsOneLearning.java

Maven Project: kelp-core

JSON type: oneVsOne

Description: it is a meta algorithm that operates by applying a One-Vs-One strategy over the base learning algorithm which is intended to be a binary learner. This meta-algorithms will learn classifiers, by comparing each class with all the others separately. The resulting classifier applies a voting strategy to perform the final decision. (N is the number of classes in the dataset)

Parameters:

---

MultiLabelClassificationLearning

Java class: MultiLabelClassificationLearning

Source code: MultiLabelClassificationLearning.java

Maven Project: kelp-core

JSON type: multiLabel

Description: it is a meta algorithm that operates applying a multi label learning strategy over the base learning algorithm which is intended to be a binary learner. A multi label classification task is a multiclass problem in which each instance can belongs to zero, one or multiple classes. The multi label strategy will learn N different classifiers, where N is the number of classes involved in the dataset. In this strategy each classifier is learned by considering in turn the examples of a single class as positives, while all the other examples are considered as negative.

Parameters:

• labels: the list of targeted classes
• baseAlgorithm: the base learning algorithm which is intended to be a binary learner

---

MultiEpochLearning

Java class: MultiEpochLearning

Source code: MultiEpochLearning.java

Maven Project: kelp-additional-algorithms

JSON type: multiEpoch

Description: it is a meta learning algorithms for online learning methods. It performs multiple iterations (or epochs) on the training data.

Parameters:

---

Stoptron

Java class: Stoptron

Source code: Stoptron.java

Maven Project: kelp-additional-algorithms

JSON type: stoptron

Description: it is a variation of the Stoptron proposed in [Orabona et al.(2008)]. Until the budget is not reached the online learning updating policy is the one of the baseAlgorithm that this meta-algorithm is exploiting. When the budget is full, the learning process ends.

Parameters:

---

RandomizedBudgetPerceptron

Java class: RandomizedBudgetPerceptron

Source code: RandomizedBudgetPerceptron.java

Maven Project: kelp-additional-algorithms

JSON type: randomizedPerceptron

Description: a variant of the Randomized Budget Perceptron proposed in [Cavallanti et al.(2007)]. Until the budget is not reached the online learning updating policy is the one of the baseAlgorithm that this meta-algorithm is exploiting. When the budget is full, a random support vector is deleted and the perceptron updating policy is adopted.

Parameters: