Алгоритм «обертки» для многообъектного обучения: первые результаты

Аннотация

Multi-instance learning is a generalization of supervised learning, where each example is represented by a labeled bag composed by a set of instances. Several multi-instance learning methods transform each bag into a single instance and then apply standard supervised learning methods. This paper presents a new multi-instance learning method that transforms the multi-instance data and is inspired by text mining. The proposed method transforms the multi-instance data into a traditional attribute- value representation by creating a corpus of documents formed by artificial words to reduce the loss of information during the transformation process. In addition, the proposed method was empirically evaluated using nine multi-instance datasets and two learning methods that transform the multi-instance data into a traditional attribute-value representation. The empirical study indicates that, in terms of classification accuracy, the proposed method is competitive with the learning methods used in the comparison. Keywords: Multi-instance learning, Bag-of-words, Wrapper method, data, algo- rithm, learning, learning methods АЛГОРИТМ «ОБЕРТКИ» ДЛЯ МНОГООБЪЕКТНОГО ОБУЧЕНИЯ: ПЕРВЫЕ РЕЗУЛЬТАТЫ Аннотация. Многофакторное обучение – это обобщение супервизорного обучения, в котором каждый пример представлен меченым мешком, состоящим из множества экземпляров. Некоторые методы многофакторного обучения преобразуют каждый мешок в один экземпляр и затем применяют стандартные 14 Digital technologies and law методы обучения с супервизией. В данной работе представлен новый метод многоэкземплярного обучения, который преобразует многоэкземплярные данные и вдохновлен разработкой текстов. Предлагаемый метод преобразует многоинстанционные данные в традиционное представление «атрибут-значение» путем создания корпуса документов, сформированного из искусственных слов, чтобы уменьшить потерю информации в процессе преобразования. Кроме того, была проведена эмпирическая оценка предложенного метода на девяти многофакторных наборах данных и двух методах обучения, преобразующих многофакторные данные в традиционное представление «атрибут-значение». Эмпирическое исследование показало, что по точности классификации предложенный метод конкурентоспособен с использованными в сравнении методами обучения. Ключевые слова: многофакторное обучение, метод «обертки», данные, алгоритм, обучение, методы обучения Introduction. Multi-instance learning is a generalization of standard propositional learning, also called attribute-value learning. While in standard learning an instance is represented by a fixed-size vector of attribute-value pairs that has a class label associated with it, in multi-instance learning an instance is represented by a bag of attribute-value vectors and the class label is associated with the whole bag. Multi-instance learning has attracted increasing interest primarily because of the wide variety of real-world problems that can be modeled quite naturally as multi-instance problems. These problems include text classification [1] image retrieval and classification [3, 4] prediction of pharmacological activity [3], index web page recommendation [5] and prediction of academic performance [6]. Since the introduction of multi-instance learning, the number of multi-instance classification methods has grown considerably. Many authors have proposed categories to try to capture the distinctive features of these methods [1]. Recently [5] proposed three main categories: Instance-based methods: these are algorithms where the learning process occurs at the instance level. Bag-based methods: includes sorters that work directly in the bag space. Mapping-based methods (wrappers): these are classifiers that apply a transforma- tion to the data of the multi-instance problem so that traditional supervised learning algorithms can be applied to obtain the solution. There are methods belonging to the wrapper category that transform multi-instance problems into traditional learning problems by replacing each bag with an attribute vec- tor consisting of a summary statistic derived from the instances in the bag. These meth- ods can lead to information loss when transforming the original multi-instance prob- lems, which affects the classification efficiency. Here we present a multi-instance learning method belonging to the category of mapping-based ones. The proposed method, called MIBoW, is inspired by text mining techniques and other fields where bag-of-words representation has been used [1–8]. MIBoW aims to achieve a reduction of information loss during the transformation of multi-instance data into a traditional attribute-value representation. MIBoW can be seen 15 Digital technologies and law Digital technologies and law as a transformation of the multi-instance dataset into a corpus of documents, where each bag becomes a document described by a set of artificial words that will be the attributes in the transformed dataset. This paper shows the initial experimenta