Detailed description of the EnsLearn unit (SDIA option, Centrale Lille)
In this unit, classifier / regressor ensembling is presented. For any prediction task, multiple algorithmic solutions are possible and some of them have comparable performances but complementary advantages. Ensembling consists in combining these algorithms so as to make the best use of them.
The unit is organized with respect to different classes of methods: linear aggregation of classifiers or regressors (vote mechanisms, weighting, exponential weights, bagging), iterated aggregation (preference fusion), non-linear aggregation (generalized averages, Bayesian approach, stacking), sequential aggregation (boosting and gradient boosting). We also touch related fields in connection with other units of the program (mixture models and Bayesian learning).