Imbalanced data is characterized by the severe difference in observation frequency between classes and has received a lot of attention in data mining research. The prediction performances usually deteriorate as classifiers learn from imbalanced data, as most classifiers assume the class distribution is balanced or the costs for different types of classification errors are equal. Although several methods have been devised to deal with imbalance problems, it is still difficult to generalize those methods to achieve stable improvement in most cases. In this study, we propose a novel framework called model-based synthetic sampling (MBS) to cope with imbalance problems, in which we integrate modeling and sampling techniques to generate synthetic data. The key idea behind the proposed method is to use regression models to capture the relationship between features and to consider data diversity in the process of data generation. We conduct experiments on 13 datasets and compare the proposed method with 10 methods. The experimental results indicate that the proposed method is not only comparative but also stable. We also provide detailed investigations and visualizations of the proposed method to empirically demonstrate why it could generate good data samples.

Early classification on multivariate time series is an important research topic in data mining with wide applications to various domains like medical diagnosis, motion detection and financial prediction, etc. Shapelet is probably one of the most commonly used approaches to tackle early classification problem, but one drawback of shaplet is its inefficiency. More importantly, the extracted shapelets may not be applicable to every test case at any time point. This work focuses on early classification of multivariate time series and proposes a novel framework named Multi-Domain Deep Neural Network (MDDNN), in which convolutional neural network (CNN) and long-short term memory (LSTM) are incorporated to learn feature representation and relationship embedding in the long sequences with long time lags. The proposed model can make predictions at any time point of a multivariate time series with the help of a truncation process. We conducted experiments on four real datasets and compared with state-of-the-art algorithms. The experimental results indicate that the proposed method outperforms the alternatives significantly on both of earliness and accuracy. Detailed analysis about the proposed model is also provided in this work. To the best of our knowledge, this is the first work that incorporates deep neural network methods (CNN and LSTM) and multi-domain approach to boost the problem of early classification on multivariate time series.

The imbalanced data problem occurs in many application domains and is considered to be a challenging problem in machine learning and data mining. Most resampling methods for synthetic data focus on minority class without considering the data distribution of major classes. In contrast to previous works, the proposed method considers both majority classes and minority classes to learn feature embeddings and utilizes appropriate loss functions to make feature embedding as discriminative as possible. The proposed method is a comprehensive framework and different deep learning feature extractors can be utilized for different domains. We conduct experiments utilizing seven numerical datasets and one image dataset based on multiclass classification tasks. The experimental results indicate that the proposed method provides accurate and stable results.

Processing large volumes of streaming data in near-real-time is becoming increasingly important as the Internet, sensor networks and network traffic grow. Online machine learning is a typical means of dealing with streaming data, since it allows the classification model to learn one instance of data at a time. Although many online learning methods have been developed since the development of the Perceptron algorithm, existing online methods assume that the number of classes is available in advance of classification process. However, this assumption is unrealistic for large scale or streaming data sets. This work proposes an online Chinese restaurant process (CRP) algorithm, which is an online and nonparametric algorithm, to tackle this problem. This work proposes a relaxing function as part of the prior and updates the parameters with the likelihood function in terms of the consistency between the true label information and predicted result. This work presents two Gibbs sampling algorithms to perform posterior inference. In the experiments, the online CRP is applied to three massive data sets, and compared with several online learning and batch learning algorithms. One of the data sets is obtained from Wikipedia, which comprises approximately two million documents. The experimental results reveal that the proposed online CRP performs well and efficiently on massive data sets. Finally, this work proposes two methods to update the hyperparameter $\alpha$ of the online CRP. The first method is based on the posterior distribution of $\alpha$, and the second exploits the property of online learning, namely adapting to change, to adjust $\alpha$ dynamically.