The explosion of data volume has gradually transformed big data processing from the static batch mode to the online streaming model. Streaming data can be divided into instance streams (feature space remains fixed while instances increase over time), feature streams (instance space is fixed while features arrive over time), or both. Generally, online streaming data learning has two main challenges: infinite length and concept changing. Recently, feature stream learning has received much attention. However, existing feature stream learning methods focus on feature selection or classification but ignore the concept changing over time. To the best of our knowledge, this is the first work that studies concept evolution detection over feature streams. Specifically, we first give the formal definition of concept evolution over feature streams, which include three different types: concept emerging, concept drift, and concept forgetting. Then, we design a novel framework to detect the concept evolution over feature streams that consists of a sliding window, an improved density peak-based clustering algorithm, and a weighted bipartite graph-based concept detecting method. Extensive experiments have been conducted on several synthetic and high-dimensional datasets to indicate our new method’s ability to cluster and detect concept evolution over feature streams.

Undersampling is one of the most popular techniques for dealing with class-imbalance problems. Various undersampling methods have emerged over the past few decades. Each of them exhibits the superiority in some scenarios. However, selecting representative majority-class samples such that the structures of the selected groups are maintained according to the underlying imbalanced distribution remains a challenge. For this purpose, this paper proposes Spatial Distribution-based UnderSampling (SDUS) for imbalanced learning. SDUS uses a supervised constructive process to learn majority-class local patterns in terms of sphere neighborhoods (SPN). Two sample selection strategies, specifically, a top-down strategy and a bottom-up strategy, are proposed for maintaining the distribution pattern of original data in selecting majority-class sample subsets from different perspectives. SDUS introduces an ensemble technique that improves learning performance by utilizing the diversity caused by the randomness of the local-pattern learning process. Numerical experiments on 38 typical datasets from KEEL repository and 13 state-of-the-art comparison methods demonstrate the effectiveness of SDUS in maintaining the underlying distribution characteristics for imbalanced undersampling.

Multi-class imbalanced problem widely exists in real-life applications and has been a challenging issue. Existing sampling methods including decomposition approaches and dedicated approaches have limitations in handling the complex mutual relationships along with data difficulty factors. Actually, the relative minorities are critical in mutual relationship, and the data difficulty factors are harmful for these minority classes. In this paper, we propose SHSampler, a similarity-based hybrid sampling to combat the mutuality by addressing data difficulty factors in multi-class imbalanced data. Specifically, SHSampler firstly utilizes a sample similarity and dissimilarity estimation to identify data difficulty factors. Then, SHSampler conducts a relative majority weakening undersampling and a relative minority strengthening oversampling to reduce the negative impact of data difficulty factors and highlight the importance of the minorities. Extensive experiments over 20 typical datasets demonstrate the superiority of SHSampler in terms of MAUC and mGM when compared with 6 state-of-the-art methods.

Feature selection is one of the core concepts in machine learning, which hugely impacts the model’s performance. For some real-world applications, features may exist in a stream mode that arrives one by one over time, while we cannot know the exact number of features before learning. Online streaming feature selection aims at selecting optimal stream features at each timestamp on the fly. Without the global information of the entire feature space, most of the existing methods select stream features in terms of individual feature information or the comparison of features in pairs. This article proposes a new online scalable streaming feature selection framework from the dynamic decision perspective that is scalable on running time and selected features by dynamic threshold adjustment. Regarding the philosophy of “Thinking-in-Threes”, we classify each new arrival feature as selecting, discarding, or delaying, aiming at minimizing the overall decision risks. With the dynamic updating of global statistical information, we add the selecting features into the candidate feature subset, ignore the discarding features, cache the delaying features into the undetermined feature subset, and wait for more information. Meanwhile, we perform the redundancy analysis for the candidate features and uncertainty analysis for the undetermined features. Extensive experiments on eleven real-world datasets demonstrate the efficiency and scalability of our new framework compared with state-of-the-art algorithms.