In this thesis, learning refers to the intelligent computational extraction of knowledge from data. Supervised learning tasks require data to be annotated with labels, whereas for unsupervised learning, data is not labelled. Semi-supervised learning deals with data sets that are partially labelled. A major issue with supervised and semi-supervised learning of data streams is late-arriving or missing class labels. Assuming that correctly labelled data will always be available and timely is often unfeasible, and, as such, supervised methods are not directly applicable in the real world. Therefore, real-world problems usually require the use of semi-supervised or unsupervised learning techniques. For instance, when considering a spam detection task, it is not reasonable to assume that all spam will be identified (correctly labelled) prior to learning. Additionally, in semi-supervised learning, "the instances having the highest [predictive] confidence are not necessarily the most useful ones" [41]. We investigate how self-training performs without its selective heuristic in a streaming setting. This leads us to our contributions. We extend an existing concept drift detector to operate without any labelled data, by using a sliding window of our ensemble's prediction confidence, instead of a boolean indicating whether the ensemble's predictions are correct. We also extend selective self-training, a semi-supervised learning method, by using all predictions, and not only those with high predictive confidence. Finally, we introduce a novel windowing type for ensembles, as sliding windows are very time consuming and regular tumbling windows are not a suitable replacement. Our windowing technique can be considered a hybrid of the two: we train each sub-classifier in the ensemble with tumbling windows, but delay training in such a way that only one sub-classifier can update its model per iteration. We found, through statistical significance tests, that our framework is (roughly 160 times) faster than current state of the art techniques, and achieves comparable predictive accuracy. That being said, more research is needed to further reduce the quantity of labelled data used for training, while also increasing its predictive accuracy.

Concept drift refers to a problem that is caused by a change in the data distribution in data mining. This leads to reduction in the accuracy of the current model that is used to examine the underlying data distribution of the concept to be discovered. A number of techniques have been introduced to address this issue, in a supervised learning (or classification) setting. In a classification setting, the target concept (or class) to be learned is known. One of these techniques is called “Ensemble learning”, which refers to using multiple trained classifiers in order to get better predictions by using some voting scheme. In a traditional ensemble, the underlying base classifiers are all of the same type. Recent research extends the idea of ensemble learning to the idea of using committees, where a committee consists of diverse classifiers. This is the main difference between the regular ensemble classifiers and the committee learning algorithms. Committees are able to use diverse learning methods simultaneously and dynamically take advantage of the most accurate classifiers as the data change. In addition, some committees are able to replace their members when they perform poorly. This thesis presents two new algorithms that address concept drifts. The first algorithm has been designed to systematically introduce gradual and sudden concept drift scenarios into datasets. In order to save time and avoid memory consumption, the Concept Drift Introducer (CDI) algorithm divides the number of drift scenarios into phases. The main advantage of using phases is that it allows us to produce a highly scalable concept drift detector that evaluates each phase, instead of evaluating each individual drift scenario. We further designed a novel algorithm to handle concept drift. Our Dynamic Committee for Concept Drift (DCCD) algorithm uses a voted committee of hypotheses that vote on the best base classifier, based on its predictive accuracy. The novelty of DCCD lies in the fact that we employ diverse heterogeneous classifiers in one committee in an attempt to maximize diversity. DCCD detects concept drifts by using the accuracy and by weighing the committee members by adding one point to the most accurate member. The total loss in accuracy for each member is calculated at the end of each point of measurement, or phase. The performance of the committee members are evaluated to decide whether a member needs to be replaced or not. Moreover, DCCD detects the worst member in the committee and then eliminates this member by using a weighting mechanism. Our experimental evaluation centers on evaluating the performance of DCCD on various datasets of different sizes, with different levels of gradual and sudden concept drift. We further compare our algorithm to another state-of-the-art algorithm, namely the MultiScheme approach. The experiments indicate the effectiveness of our DCCD method under a number of diverse circumstances. The DCCD algorithm generally generates high performance results, especially when the number of concept drifts is large in a dataset. For the size of the datasets used, our results showed that DCCD produced a steady improvement in performance when applied to small datasets. Further, in large and medium datasets, our DCCD method has a comparable, and often slightly higher, performance than the MultiScheme technique. The experimental results also show that the DCCD algorithm limits the loss in accuracy over time, regardless of the size of the dataset.

Concept drift refers to a problem that is caused by a change in the data distribution in data mining. This leads to reduction in the accuracy of the current model that is used to examine the underlying data distribution of the concept to be discovered. A number of techniques have been introduced to address this issue, in a supervised learning (or classification) setting. In a classification setting, the target concept (or class) to be learned is known. One of these techniques is called “Ensemble learning”, which refers to using multiple trained classifiers in order to get better predictions by using some voting scheme. In a traditional ensemble, the underlying base classifiers are all of the same type. Recent research extends the idea of ensemble learning to the idea of using committees, where a committee consists of diverse classifiers. This is the main difference between the regular ensemble classifiers and the committee learning algorithms. Committees are able to use diverse learning methods simultaneously and dynamically take advantage of the most accurate classifiers as the data change. In addition, some committees are able to replace their members when they perform poorly. This thesis presents two new algorithms that address concept drifts. The first algorithm has been designed to systematically introduce gradual and sudden concept drift scenarios into datasets. In order to save time and avoid memory consumption, the Concept Drift Introducer (CDI) algorithm divides the number of drift scenarios into phases. The main advantage of using phases is that it allows us to produce a highly scalable concept drift detector that evaluates each phase, instead of evaluating each individual drift scenario. We further designed a novel algorithm to handle concept drift. Our Dynamic Committee for Concept Drift (DCCD) algorithm uses a voted committee of hypotheses that vote on the best base classifier, based on its predictive accuracy. The novelty of DCCD lies in the fact that we employ diverse heterogeneous classifiers in one committee in an attempt to maximize diversity. DCCD detects concept drifts by using the accuracy and by weighing the committee members by adding one point to the most accurate member. The total loss in accuracy for each member is calculated at the end of each point of measurement, or phase. The performance of the committee members are evaluated to decide whether a member needs to be replaced or not. Moreover, DCCD detects the worst member in the committee and then eliminates this member by using a weighting mechanism. Our experimental evaluation centers on evaluating the performance of DCCD on various datasets of different sizes, with different levels of gradual and sudden concept drift. We further compare our algorithm to another state-of-the-art algorithm, namely the MultiScheme approach. The experiments indicate the effectiveness of our DCCD method under a number of diverse circumstances. The DCCD algorithm generally generates high performance results, especially when the number of concept drifts is large in a dataset. For the size of the datasets used, our results showed that DCCD produced a steady improvement in performance when applied to small datasets. Further, in large and medium datasets, our DCCD method has a comparable, and often slightly higher, performance than the MultiScheme technique. The experimental results also show that the DCCD algorithm limits the loss in accuracy over time, regardless of the size of the dataset.

Concept drift refers to a problem that is caused by a change in the data distribution in data mining. This leads to reduction in the accuracy of the current model that is used to examine the underlying data distribution of the concept to be discovered. A number of techniques have been introduced to address this issue, in a supervised learning (or classification) setting. In a classification setting, the target concept (or class) to be learned is known. One of these techniques is called “Ensemble learning”, which refers to using multiple trained classifiers in order to get better predictions by using some voting scheme. In a traditional ensemble, the underlying base classifiers are all of the same type. Recent research extends the idea of ensemble learning to the idea of using committees, where a committee consists of diverse classifiers. This is the main difference between the regular ensemble classifiers and the committee learning algorithms. Committees are able to use diverse learning methods simultaneously and dynamically take advantage of the most accurate classifiers as the data change. In addition, some committees are able to replace their members when they perform poorly. This thesis presents two new algorithms that address concept drifts. The first algorithm has been designed to systematically introduce gradual and sudden concept drift scenarios into datasets. In order to save time and avoid memory consumption, the Concept Drift Introducer (CDI) algorithm divides the number of drift scenarios into phases. The main advantage of using phases is that it allows us to produce a highly scalable concept drift detector that evaluates each phase, instead of evaluating each individual drift scenario. We further designed a novel algorithm to handle concept drift. Our Dynamic Committee for Concept Drift (DCCD) algorithm uses a voted committee of hypotheses that vote on the best base classifier, based on its predictive accuracy. The novelty of DCCD lies in the fact that we employ diverse heterogeneous classifiers in one committee in an attempt to maximize diversity. DCCD detects concept drifts by using the accuracy and by weighing the committee members by adding one point to the most accurate member. The total loss in accuracy for each member is calculated at the end of each point of measurement, or phase. The performance of the committee members are evaluated to decide whether a member needs to be replaced or not. Moreover, DCCD detects the worst member in the committee and then eliminates this member by using a weighting mechanism. Our experimental evaluation centers on evaluating the performance of DCCD on various datasets of different sizes, with different levels of gradual and sudden concept drift. We further compare our algorithm to another state-of-the-art algorithm, namely the MultiScheme approach. The experiments indicate the effectiveness of our DCCD method under a number of diverse circumstances. The DCCD algorithm generally generates high performance results, especially when the number of concept drifts is large in a dataset. For the size of the datasets used, our results showed that DCCD produced a steady improvement in performance when applied to small datasets. Further, in large and medium datasets, our DCCD method has a comparable, and often slightly higher, performance than the MultiScheme technique. The experimental results also show that the DCCD algorithm limits the loss in accuracy over time, regardless of the size of the dataset.

Cette thèse s’intéresse à la problématique de la classification dynamique en environnements stationnaires et non stationnaires, tolérante aux variations de quantités des données d’apprentissage et capable d’ajuster ses modèles selon la variabilité des données entrantes. Pour cela, nous proposons une solution faisant cohabiter des classificateurs one-class SVM indépendants ayant chacun leur propre procédure d’apprentissage incrémentale et par conséquent, ne subissant pas d’influences croisées pouvant émaner de la configuration des modèles des autres classificateurs. L’originalité de notre proposition repose sur l’exploitation des anciennes connaissances conservées dans les modèles de SVM (historique propre à chaque SVM représenté par l’ensemble des vecteurs supports trouvés) et leur combinaison avec les connaissances apportées par les nouvelles données au moment de leur arrivée. Le modèle de classification proposé (mOC-iSVM) sera exploité à travers trois variations exploitant chacune différemment l’historique des modèles. Notre contribution s’inscrit dans un état de l’art ne proposant pas à ce jour de solutions permettant de traiter à la fois la dérive de concepts, l’ajout ou la suppression de concepts, la fusion ou division de concepts, tout en offrant un cadre privilégié d’interactions avec l’utilisateur. Dans le cadre du projet ANR DIGIDOC, notre approche a été appliquée sur plusieurs scénarios de classification de flux d’images pouvant survenir dans des cas réels lors de campagnes de numérisation. Ces scénarios ont permis de valider une exploitation interactive de notre solution de classification incrémentale pour classifier des images arrivant en flux afin d’améliorer la qualité des images numérisées. / This research contributes to the field of dynamic learning and classification in case of stationary and non-stationary environments. The goal of this PhD is to define a new classification framework able to deal with very small learning dataset at the beginning of the process and with abilities to adjust itself according to the variability of the incoming data inside a stream. For that purpose, we propose a solution based on a combination of independent one-class SVM classifiers having each one their own incremental learning procedure. Consequently, each classifier is not sensitive to crossed influences which can emanate from the configuration of the models of the other classifiers. The originality of our proposal comes from the use of the former knowledge kept in the SVM models (represented by all the found support vectors) and its combination with the new data coming incrementally from the stream. The proposed classification model (mOC-iSVM) is exploited through three variations in the way of using the existing models at each step of time. Our contribution states in a state of the art where no solution is proposed today to handle at the same time, the concept drift, the addition or the deletion of concepts, the fusion or division of concepts while offering a privileged solution for interaction with the user. Inside the DIGIDOC project, our approach was applied to several scenarios of classification of images streams which can correspond to real cases in digitalization projects. These different scenarios allow validating an interactive exploitation of our solution of incremental classification to classify images coming in a stream in order to improve the quality of the digitized images.

[pt] Os conceitos do mundo real muitas vezes não são estáveis: eles mudam com o tempo. Assim como os conceitos, a distribuição de dados também pode se alterar. Este problema de mudança de conceitos ou distribuição de dados é conhecido como concept drift e é um desafio para um modelo na tarefa de aprender a partir de dados. Este trabalho apresenta um novo modelo neuroevolutivo com inspiração quântica, baseado em um comitê de redes neurais do tipo Multi-Layer Perceptron (MLP), para a aprendizagem em ambientes não estacionários, denominado NEVE (Neuro-EVolutionary Ensemble). Também apresenta um novo mecanismo de detecção de concept drift, denominado DetectA (Detect Abrupt) com a capacidade de detectar mudanças tanto de forma proativa quanto de forma reativa. O algoritmo evolutivo com inspiração quântica binário-real AEIQ-BR é utilizado no NEVE para gerar automaticamente novos classificadores para o comitê, determinando a topologia mais adequada para a nova rede, selecionando as variáveis de entrada mais apropriadas e determinando todos os pesos da rede neural MLP. O algoritmo AEIQ-R determina os pesos de votação de cada rede neural membro do comitê, sendo possível utilizar votação por combinação linear, votação majoritária ponderada e simples. São implementadas quatro diferentes abordagens do NEVE, que se diferem uma da outra pela forma de detectar e tratar os drifts ocorridos. O trabalho também apresenta resultados de experimentos realizados com o método DetectA e com o modelo NEVE em bases de dados reais e artificiais. Os resultados mostram que o detector se mostrou robusto e eficiente para bases de dados de alta dimensionalidade, blocos de tamanho intermediário, bases de dados com qualquer proporção de drift e com qualquer balanceamento de classes e que, em geral, os melhores resultados obtidos foram usando algum tipo de detecção. Comparando a acurácia do NEVE com outros modelos consolidados da literatura, verifica-se que o NEVE teve acurácia superior na maioria dos casos. Isto reforça que a abordagem por comitê neuroevolutivo é uma escolha robusta para situações em que as bases de dados estão sujeitas a mudanças repentinas de comportamento. / [en] Real world concepts are often not stable: they change with time. Just as the concepts, data distribution may change as well. This problem of change in concepts or distribution of data is known as concept drift and is a challenge for a model in the task of learning from data. This work presents a new neuroevolutive model with quantum inspiration called NEVE (Neuro- EVolutionary Ensemble), based on an ensemble of Multi-Layer Perceptron (MLP) neural networks for learning in non-stationary environments. It also presents a new concept drift detection mechanism, called DetectA (DETECT Abrupt) with the ability to detect changes both proactively as reactively. The evolutionary algorithm with binary-real quantum inspiration AEIQ-BR is used in NEVE to automatically generate new classifiers for the ensemble, determining the most appropriate topology for the new network and by selecting the most appropriate input variables and determining all the weights of the neural network. The AEIQ-R algorithm determines the voting weight of each neural network ensemble member, and you can use voting by linear combination and voting by weighted or simple majority. Four different approaches of NEVE are implemented and they differ from one another by the way of detecting and treating occurring drifts. The work also presents results of experiments conducted with the DetectA method and with the NEVE model in real and artificial databases. The results show that the detector has proved efficient and suitable for data bases with high-dimensionality, intermediate sized blocks, any proportion of drifts and with any class balancing. Comparing the accuracy of NEVE with other consolidated models in the literature, it appears that NEVE had higher accuracy in most cases. This reinforces that the neuroevolution ensemble approach is a robust choice to situations in which the databases are subject to sudden changes in behavior.