Deep learning models may not effectively generalize across under-represented or minority classes. We empirically study a convolutional neural network’s (CNN) internal representation of imbalanced image data and measure the generalization gap between a model’s feature embeddings in the training and test sets, showing that the gap is wider for minority classes. This insight enables us to design an efficient three-phase CNN training framework for imbalanced data. The framework involves training the network end-to-end on imbalanced data to learn feature embeddings, performing data augmentation in the learned embedding space to balance the training data distribution, and fine-tuning the classifier head on the embedded balanced training data. We develop Expansive Over-Sampling (EOS) as a data augmentation technique to utilize in the training framework. EOS forms synthetic training instances as convex combinations between the minority class samples and their nearest adversaries in the embedding space to reduce the generalization gap. The proposed framework improves the accuracy over leading cost-sensitive and resampling methods commonly used in imbalanced learning. Moreover, it is more computationally efficient than standard data pre-processing methods, such as SMOTE and GAN-based over-sampling, as it requires fewer parameters and less training time. The source code for the proposed framework is available at: https://github.com/dd1github/EOS.

Structural concept complexity, class overlap, and data scarcity are some of the most important factors influencing the performance of classifiers under class imbalance conditions. When these effects were uncovered in the early 2000s, understandably, the classifiers on which they were demonstrated belonged to the classical rather than Deep Learning categories of approaches. As Deep Learning is gaining ground over classical machine learning and is beginning to be used in critical applied settings, it is important to assess systematically how well they respond to the kind of challenges their classical counterparts have struggled with in the past two decades. The purpose of this paper is to study the behavior of deep learning systems in settings that have previously been deemed challenging to classical machine learning systems to find out whether the depth of the systems is an asset in such settings. The results in both artificial and real-world image datasets show that these settings remain mostly challenging for Deep Learning systems. Deeper architectures help with structural concept complexity but not with data scarcity and class overlap.

Data stream classification is one of the most vital areas of contemporary machine learning, as many real-life problems generate data continuously and in large volumes. However, most of research in this area focuses on vector-based representations, which are unsuitable for capturing properties of more complex multi-dimensional structures, such as images and video sequences. In this paper, we propose a novel methodology for learning adaptive decision trees from data streams of tensors. We introduce Chordal Kernel Decision Tree for continual learning from tensor data streams. In order to maintain the tensor characteristics, we propose to train and update classifiers in the kernel space designed to work with tensor representation. We use chordal distance to compute similarities between tensors and then apply it as a new feature space in which decision trees are trained. This allows for a direct decision tree induction on tensors. In order to accommodate the streaming and drifting nature of data, we propose a concept drift detection scheme based on tensor representation. It allows us to reconstruct the kernel feature space every time when change is detected. The proposed approach allows for fast and efficient induction of decision trees on streaming data with tensor representation. Experimental study, conducted on 4 real-world and 52 artificial large-scale tensor data streams, shows that using the native tensor feature space leads to more accurate classification than outperforms the vectorized representations.

Continual learning from data streams is among the most important topics in contemporary machine learning. One of the biggest challenges in this domain lies in creating algorithms that can continuously adapt to arriving data. However, previously learned knowledge may become outdated, as streams evolve over time. This phenomenon is known as concept drift and must be detected to facilitate efficient adaptation of the learning model. While there exists a plethora of drift detectors, all of them assume that we are dealing with roughly balanced classes. In the case of imbalanced data streams, those detectors will be biased towards the majority classes, ignoring changes happening in the minority ones. Furthermore, class imbalance may evolve over time and classes may change their roles (majority becoming minority and vice versa). This is especially challenging in the multi-class setting, where relationships among classes become complex. In this paper, we propose a detailed taxonomy of challenges posed by concept drift in multi-class imbalanced data streams, as well as a novel trainable concept drift detector based on Restricted Boltzmann Machine. It is capable of monitoring multiple classes at once and using reconstruction error to detect changes in each of them independently. Our detector utilizes a skew-insensitive loss function that allows it to handle multiple imbalanced distributions. Due to its trainable nature, it is capable of following changes in a stream and evolving class roles, as well as it can deal with local concept drift occurring in minority classes. Extensive experimental study on multi-class drifting data streams, enriched with a detailed analysis of the impact of local drifts and changing imbalance ratios, confirms the high efficacy of our approach.

Learning from data streams is one of the most promising and challenging domains in modern machine learning. Proliferating online data sources provide us access to real-time knowledge we have never had before. At the same time, new obstacles emerge and we have to overcome them in order to fully and effectively utilize the potential of the data. Prohibitive time and memory constraints or non-stationary distributions are only some of the problems. When dealing with classification tasks, one has to remember that effective adaptation has to be achieved on weak foundations of partially labeled and often imbalanced data. In our work, we propose an online framework for binary classification, that aims to handle the complex problem of working with dynamic, sparsely labeled and imbalanced streams. The main part of it is a novel active learning strategy (MD-OAL) that is able to prioritize labeling of minority instances and, as a result, improve the balance of the learning process. We combine the strategy with a dynamic ensemble of base learners that can abstain from making decisions, if they are very uncertain. We adjust the abstaining mechanism in favor of minority instances, providing an effective method for handling remaining imbalance and a concept drift simultaneously. The conducted evaluation shows that in the challenging and realistic scenarios our framework outperforms state-of-the-art algorithms, providing higher resilience to the combined effect of limited labeling and imbalance.

Most machine learning methods work under the assumption that classes have a roughly balanced number of instances. However, in many real-life problems we may have some types of instances appearing predominantly more frequently than the others which causes a bias towards the majority class during classifier training. This becomes even more challenging when dealing with multiple classes, where relationships between them are not easily defined. Learning from multi-class imbalanced data has not been widely considered in the context of big data mining, despite the fact that this is a learning difficulty frequently appearing in this domain. In this paper, we address this challenge by proposing a comprehensive ensemble-based framework. We propose to analyze each class to extract instance-level characteristics describing their difficulty levels. We embed this information into the existing UnderBagging framework. Our ensemble samples instances with probabilities proportional to their difficulty levels. This allows us to focus the learning process on the most difficult instances, better capturing the properties of multi-class imbalanced problems. We implemented our framework on Apache Spark to allow for high-performance computing over big data sets. This experimental study shows that taking into account the instance-level difficulty leads to training of significantly more accurate ensembles.

Data stream mining is among the most contemporary branches of machine learning. The potentially infinite sources give us many opportunities and at the same time pose new challenges. To properly handle streaming data we need to improve our well-established methods, so they can work with dynamic data and under strict constraints. Supervised streaming machine learning algorithms require a certain number of labeled instances in order to stay up-to-date. Since high budgets dedicated for this purpose are usually infeasible, we have to limit the supervision as much as we can. One possible approach is to trigger labeling, only if a change is explicitly indicated by a detector. While there are several supervised algorithms dedicated for this purpose, the more practical unsupervised ones are still lacking a proper attention. In this paper, we propose a novel unsupervised ensemble drift detector that recognizes local changes in feature subspaces (EDFS) without additional supervision, using specialized committees of incremental Kolmogorov-Smirnov tests. We combine it with an adaptive classifier and update it, only if the drift detector signalizes a change. Conducted experiments show that our framework is able to efficiently adapt to various concept drifts and outperform other unsupervised algorithms.

In multi-label learning, data may simultaneously belong to more than one class. When multi-label data arrives as a stream, the challenges associated with multi-label learning are joined by those of data stream mining, including the need for algorithms that are fast and flexible, able to match both the speed and evolving nature of the stream. This article presents a punitive k nearest neighbors algorithm with a self-adjusting memory (MLSAMPkNN) for multi-label, drifting data streams. The memory adjusts in size to contain only the current concept and a novel punitive system identifies and penalizes errant data examples early, removing them from the window. By retaining and using only data that are both current and beneficial, MLSAMPkNN is able to adapt quickly and efficiently to changes within the data stream while still maintaining a low computational complexity. Additionally, the punitive removal mechanism offers increased robustness to various data-level difficulties present in data streams, such as class imbalance and noise. The experimental study compares the proposal to 24 algorithms using 30 real-world and 15 artificial multi-label data streams on six multi-label metrics, evaluation time, and memory consumption. The superior performance of the proposed method is validated through non-parametric statistical analysis, proving both high accuracy and low time complexity. MLSAMPkNN is a versatile classifier, capable of returning excellent performance in diverse stream scenarios.

Data stream mining is a rapidly developing branch of contemporary machine learning. Ensemble approaches have proven themselves to be highly effective in this domain, due to their predictive power and capabilities for handling evolving data. One of the key aspects of ensemble learning is diversity among base classifiers - it improves accuracy and allows for anticipating and recovering from concept drifts. It has been shown that while diversity is desirable during changes, it may impede learning when data becomes stationary. In this paper, we present a novel ensemble technique that exploits the idea of dynamic diversification, which increases diversity during changes and reduces it when a stream becomes stable. The algorithm uses online clustering for this task by creating locally specialized base learners trained on spatially related instances. Three control strategies based on the novel range heuristic for managing a trade-off between error (a change indicator) and diversity are utilized. Additionally, two intensification strategies are proposed for exploitation of newly arriving instances, allowing for faster adaptation. Experimental study evaluates the general performance and diversity of the proposed algorithm, proving its capabilities to outperform state-of-the-art ensembles dedicated to drifting data stream mining.

Most of data stream classifier learning methods assume that a true class of an incoming object is available right after the instance has been processed and new and labeled instance may be used to update a classifier's model, drift detection or capturing novel concepts. However, assumption that we have an unlimited and infinite access to class labels is very naive and usually would require a very high labeling cost. Therefore the applicability of many supervised techniques is limited in real-life stream analytics scenarios. Active learning emerges as a potential solution to this problem, concentrating on selecting only the most valuable instances and learning an accurate predictive model with as few labeling queries as possible. However learning from data streams differ from online learning as distribution of examples may change over time. Therefore, an active learning strategy must be able to handle concept drift and quickly adapt to evolving nature of data. In this paper we present novel active learning strategies that are designed for effective tackling of such changes. We assume that most labeling effort is required when concept drift occurs, as we need a representative sample of new concept to retrain properly the predictive model. Therefore, we propose active learning strategies that are guided by drift detection module to save budget for difficult and evolving instances. Three proposed strategies are based on learner uncertainty, dynamic allocation of budget over time and search space randomization. Experimental evaluation of the proposed methods prove their usefulness for reducing labeling effort in learning from drifting data streams.

The class imbalance problem is a pervasive issue in many real-world domains. Oversampling methods that inflate the rare class by generating synthetic data are amongst the most popular techniques for resolving class imbalance. However, they concentrate on the characteristics of the minority class and use them to guide the oversampling process. By completely overlooking the majority class, they lose a global view on the classification problem and, while alleviating the class imbalance, may negatively impact learnability by generating borderline or overlapping instances. This becomes even more critical when facing extreme class imbalance, where the minority class is strongly underrepresented and on its own does not contain enough information to conduct the oversampling process. We propose a novel method for synthetic oversampling that uses the rich information inherent in the majority class to synthesize minority class data. This is done by generating synthetic data that is at the same Mahalanbois distance from the majority class as the known minority instances. We evaluate over 26 benchmark datasets, and show that our method offers a distinct performance improvement over the existing state-of-the-art in oversampling techniques.

Imbalanced classification is one of the most challenging machine learning problem. Recent studies show, that often the uneven ratio of objects in classes is not the biggest factor, determining the drop of classification accuracy. It is also related to some difficulties embedded in the nature of the data. In this paper we study the different types of minority class examples and distinguish four groups of objects - safe, borderline, rare and outliers. To deal with the imbalance problem, we use a one-class classification, that is focused on a proper identification of the minority class samples. We further augment this model by incorporating the knowledge about the minority object types in the training dataset. This is done applying weighted one-class classifier and adjusting weights assigned to minority class objects, depending on their type. A strategy for calculating the new weights for minority examples is proposed. Experimental analysis, carried on a set of benchmark datasets, confirms that the proposed model can achieve a satisfactory recognition rate and often outperform other state-of-the-art methods, dedicated to the imbalanced classification.