Dissimilarity based multiple instance learning using dictionary ensembles

Multiple Instance Learning (MIL) is one of important topics in the pattern recognition research field. It differs from many traditional machine learning problems in terms of real-world object representations. In MIL problems, samples are represented by multi-sets which are commonly named as bags where each bag include a set of feature vectors called instances. Many real world problems such as image, text or document classification, drug activity prediction and etc., can be formalized as MIL problems. Mostly, MIL problem, deals with supervised learning paradigm, which aims to learn models from the on hand samples and use it to solve regression or classification problems. Most of the cases, MIL problem is referred to binary classification where each bag has to be classified into one of the two categories; "positive" or "negative". Many MIL algorithms are developed considering the assumption that a bag is positive if at least it has one positive instance. Although this assumption has good results in some domains such as drug activity predictions, it may be restrictive for other domains of MIL problems such as computer vision MIL problems can be solved using instance based, bag based or embedded space algorithms. In this thesis, embedded-based strategy has been applied which converts the MIL problem to a standard supervised learning. In this approach, instead of using the mentioned assumption which relates the instance labels to the bags' labels, the dissimilarity of bags to the selected training instances which are called prototypes are taken into consideration. This mapping makes the bags be represented by standard fixed sized feature vectors. Most of the time, this feature mapping may include lots of redundant and irrelevant features. In this work we use Dictionary Learning as a classifier, which generates the sparse representation of each signal and classify them simultaneously. Using dictionaries would result in more efficient, fewer noises, simple and sparse manner, which make signals have more global look. As a third Approach, Ensemble Learning technique is implemented. This algorithm, apply two mentioned methods in its strategy and combine their strength where leads to higher and more reliable classification performances. Random Subspace and Bagging are two strategy implemented as Ensemble Learning approach. The proposed algorithm is evaluated on 11 different MIL datasets and compared with a recently proposed dissimilarity based ensemble MIL algorithm that uses Support Vector Machines (SVM). Experimental results show that the proposed algorithm outperforms the counterpart algorithm that uses SVM.