On The Combination of Feature and Instance Selection
Jerffeson Souza, Rafael Augusto Ferreira do Carmo, Gustavo Augusto Campos de Lim · Machine Learning · 2010
In this chapter we discussed two important problems in the pre-processing step of many supervised learning tasks. A list of well-known algorithms were presented and discussed. A new framework was proposed, extending the concept proposed by the authors in a previous work. This framework was validated by some simulations using the metaheuristic Simulated Annealing and NSGA-II. These simulations show that although the quality of solutions generated by this framework is quite similar to those obtained by sequential executions, this approach reaches the better solutions faster than the other approaches. The frameworks is based on what we called "power of influence", i.e. the quality of features in a given supervised learning task is intrinsically related to the quality of instances used in this task, and vice-versa. Based on this we created the framework that work with two separated wrappers for these two problems, jointing them in a single evaluation procedure. 5.1 Future Work - The Framework for Multi-Objective Feature and Instance Selection An important characteristic we want to add to this framework in the future is the possibility to handle the multi-objective versions of the two selection problems. The usage of multi objectives brings new power but also new problems to the search processes. In these formulations, the characteristic of total ordering is replaced by partial ordering, using the concept of Pareto optimality. The ideas of better and worse are replaced by dominance, nondominance. Given two solutions a, b and a set of functions F to be minimized (or maximized, but in this explanation we suppose they are to be minimized), we say that a weakly dominates b if and only if