Preface to the Special Issue “Advancement of Mathematical Methods in Feature Representation Learning for Artificial Intelligence, Data Mining and Robotics”—Special Issue Book

The feature representation learning is the basic task that plays an important role in artificial intelligence, data mining and robotics. With the recent rapid development of deep learning, many advanced methods have been proposed and have gained remarkable successes both in academia and in industry, such as auto-encoders, convolutional neural networks, generative adversarial networks, and so on. However, many questions remain unsolved. What makes one representation better than another? What are appropriate objectives for learning representations well? How can security and the algorithm be explained?

This special issue aims to highlight the latest results on the mathematical methods in feature representation learning for artificial intelligence, data mining and robotics, covering several recently reported methods.

The representation learning is the basic problem for computer vision. For example, the authors of [1] comprehensively reviewed the development of vehicle re-identification and revealed that representation learning plays a vital role in the vehicle re-identification. Furthermore, they classified the vehicle re-identification feature representation approaches into two parts: hand-crafted and deep learning based feature representations. In [2], semantic intelligent detection of vehicle color was studied under rainy conditions for jointly detaining and recognizing vehicle color. Specifically, the feature maps of the recovered clean image and the extracted feature maps of the input image are cascaded into the feature pyramid net (FPN) module to achieve joint semantic representation learning. Based on the YOLOX algorithm, works [3,4] proposed to learn representation features for high-performance head counting and garbage quantity identification, respectively.

Based on the fact that the low-level features contain small object information, while the high-level features contain accurate, large object information, the authors of [5] proposed an effective approach by integrating the characteristics of different stages on pedestrian detection. To explore the high-order information representation in vision tasks, the authors of [6] developed second-order spatial-temporal correlation filters for visual tracking, and the authors of [7] studied facial recognition via compact second-order image gradient orientations. In work [8], the authors proposed deep-learning based cyber & physical feature fusion for anomaly detection in industrial control systems.

In [9], discriminative multidimensional scaling based on pairwise constraints for a feature learning model was proposed considering both the topology of samples in the original space and the cluster structure in the new space. The authors of [10] proposed deep large-margin rank loss for feature learning for multi-label image classification. Ref. [11] proposed reinforcement learning-based representation approach for resource allocation in the elastic optical networks, and Ref. [12] presented a deep reinforcement learning based framework for gait adjustment for the patients suffer from physical disabilities.

Representation learning also plays an important role for the low-level image processing task. The authors of [13] studied blind image deblurring and proposed and learned an innovative sparse channel prior. The authors of [14] proposed a joint deep recovery model to efficiently address motion blur and resolution reduction simultaneously. The proposed multi-order attention mechanism comprehensively and hierarchically extracts multiple attention features and fuses them properly by drop-out gating. In [15], the authors reported an image aesthetic quality assessment and proposed a method that includes a representation learning step and a label propagation step. The authors of [16] developed a plug-and-play-based algorithm for mixed noise removal with the logarithm norm approximation model.

Since available source data are collected from related domains, multi-domain adaptation (MDA) has become increasingly popular. Although multiple source domains provide a significant amount of information, the processing of domain shifts becomes more challenging, especially in learning a common domain-invariant representation for all domains. In [17], due to the ambiguity of the category boundary, the authors proposed Dempster–Shafer evidence theory (DST) to reduce category boundary ambiguity and output reasonable decisions by combining adaptation outputs based on uncertainty. Inspired by generative adversarial networks (GANs), the authors of [18] proposed a novel adversarial domain adaptation method with an initial state fusion strategy followed by a domain similarity strategy based on information entropy. In [19], the authors adopt domain adaptation strategy to solve the remaining useful life (RUL) prediction caused by insufficient sample data of equipment under complex operating conditions. The authors of [20] proposed a geometric metric learning method for multi-output learning.

Sentiment classification is an important task in natural language processing. Traditional word-level vector representations provide the same representation for words that express different sentiment polarities in various domains. In [21], the authors proposed a dual-word embedding model considering syntactic information for cross-domain sentiment classification. The authors of [22] reported a graph convolutional network for aspect-based sentiment analysis considering the dependencies between words and the types of these dependencies simultaneously. The authors of [23] proposed a knowledge-enhanced dual-channel GCN for aspect-based sentiment analysis. In [24], the authors developed a triplet contrastive learning network to coordinate syntactic and semantic information for the domain of aspect-level sentiment classification. Works [25,26] show that the effectiveness of the knowledge enhanced sentiment feature learning for aspect-level the sentiment classification and hate speech detection. [27] studied the embedding representation learning for the uncertain temporal knowledge graph while [28] studied Tensor Affinity Learning for Hyperorder Graph Matching.

Some other representative works also show the importance of the feature representation learning. Such as, Ref. [29] studied the 3D reconstruction of self-rotating objects, Ref. [30] presented a fusion verification method cross-site scripting attacks. Ref. [31] proposed a novel feature transformation-based method to improve the robustness of adversarial example by transforming the features of data. Ref. [32] studied the requirement analysis for complex mechanical products scheme design, while Ref. Ref. [33] studied stability of switched systems with time-varying delays.

Briefly, this Special Issue received 65 submissions, 33 of which were published, including 32 research articles and 1 review article. All submissions covered topics from low-level vision feature learning to high-level semantic representation learning, including texts, images and videos from single domains to cross-domains. We believe that these will effectively boost the research on representation learning. We found the selection of papers for this Special Issue very inspiring and we thank the editorial staff and reviewers for their efforts and assistance during the process.