Special Issue on Advances in Deep and Shallow Machine Learning Approaches for Handling Data Irregularities
• 大类 : 工程技术 - 2区
• 小类 : 计算机：人工智能 - 2区
Performance of most of the well-known learning systems can considerably degrade if the data to be handled (e.g. the training examples for supervised learning) contain irregularities of various types. By data irregularity, we point to those situations where the distribution of data points, the sampling of data space for generating the training set, or the features describing each data point deviate from what could have been ideal; being biased, skewed, incomplete and/or misleading. Data irregularities can give rise to problems of class imbalance, small disjuncts (rarely covered sub-concepts within each class), class distribution skew, absent features, and missing features. While the three former problems are defined only for supervised learning systems, the latter two are defined for supervised as well as unsupervised learning systems. More than one of these assumptions may be violated together by a given dataset. Majority of the learning systems (both supervised and unsupervised) are quite sensitive to one of more of the data irregularities mentioned above, irrespectively of the systems?generalization ability or ability to learn representations. Unless special mechanisms are adopted either for data pre-processing or as a part of the learning system itself, such data irregularities considerably degrade the performance of traditional computer-based learning systems. This special issue aims to report the latest advances in the design of learning systems which can show considerable resilience to data irregularities of various forms which are frequently present in most of the real-life data.
The topics include, but are not limited to:
Handling class imbalance in supervised learning
Handling small disjuncts and sub-concepts within classes.
Handling class distribution skew and handling the disparate shape of class distributions
Missing and absent features problem in classification and clustering
Label noise and ambiguity resolution in supervised learning
Interrelation and co-occurrence of data irregularities
Effect of data irregularities on deep neural networks
Effect of data irregularities on adversarial machine learning
Theoretical studies on data preprocessing and learning algorithms tailor-made to handle data irregularities
New application areas giving rise to data irregularities and design of relevant learning systems including object detection, classification, recognition; image retrieval, segmentation, interpretation; document retrieval, categorization, topic model; signal processing, medical image analysis, bioinformatics; speech recognition, synthesis, anti-spoofing; fault detection/diagnosis, fraud detection, cyber-security; etc.