Multi-objective approach based on grammar-guided genetic programming for solving multiple instance problems

Abstract

Multiple instance learning (MIL) is considered a generalization of traditional supervised learning which deals with uncertainty in the information. Together with the fact that, as in any other learning framework, the classifier performance evaluation maintains a trade-off relationship between different conflicting objectives, this makes the classification task less straightforward. This paper introduces a multi-objective proposal that works in a MIL scenario to obtain well-distributed Pareto solutions to multi-instance problems. The algorithm developed, MultiObjective Grammar Guided Genetic Programming for Multiple Instances (MOG3P-MI), is based on grammarguided genetic programming, which is a robust tool for classification. Thus, this proposal combines the advantages of the grammar-guided genetic programming with benefits provided by multi-objective approaches. First, a study of multi-objective optimization for MIL is carried out. To do this, three different extensions of MOG3P-MI are designed and implemented and their performance is compared. This study allows us on the one hand, to check the performance of multi-objective techniques in this learning paradigm and on the other hand, to determine the most appropriate evolutionary process for MOG3P-MI. Then, MOG3P-MI is compared with some of the most significant proposals developed throughout the years in MIL. Computational experiments show that MOG3P-MI often obtains consistently better results than the other algorithms, achieving the most accurate models. Moreover, the classifiers obtained are very comprehensible.

Speeding up multiple instance learning classification rules on GPUs

“…Similarly, there are many machine learning methods available to solve these problems, such as multiinstance lazy learning algorithms [43], multi-instance tree learners [8], multi-instance rule inducers [9], multi-instance bayesian approaches [27], multi-instance kernel methods [22,38], multi-instance ensembles [45,51], and evolutionary algorithms [48,50]. However, most of the MIL algorithms are very slow and cannot be applied to large data sets.…”