Article
Computer Science, Artificial Intelligence
Xiuyi Jia, Xiaoxia Shen, Weiwei Li, Yunan Lu, Jihua Zhu
Summary: Label distribution learning (LDL) is a novel machine learning paradigm that quantitatively describes the relevance between labels and unknown instances, as well as the ranking relation between labels. However, existing LDL models only utilize the quantitative advantage and ignore the label ranking relation. Therefore, we propose a novel algorithm with a ranking loss function to address this issue, and introduce two popular ranking evaluation metrics to comprehensively evaluate LDL algorithms.
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING
(2023)
Article
Computer Science, Artificial Intelligence
Ning Xu, Yun-Peng Liu, Xin Geng
Summary: Label distribution learning covers a certain number of labels and the process of recovering label distributions can enhance the supervision information in training sets, leading to better learning performance.
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING
(2021)
Article
Environmental Sciences
Jianqiao Luo, Yihan Wang, Yang Ou, Biao He, Bailin Li
Summary: The paper introduces a neighbor-based label distribution learning method for aerial scene classification, which addresses label ambiguity by capturing local similarity and label correlations.
Article
Computer Science, Artificial Intelligence
Xingyu Zhao, Yuexuan An, Ning Xu, Xin Geng
Summary: This paper proposes a novel approach called Continuous Label Distribution Learning (CLDL) to explicitly and effectively capture the continuous distribution of different labels and extract high-order correlations among them.
PATTERN RECOGNITION
(2023)
Article
Automation & Control Systems
Chao Xu, Hong Tao, Jing Zhang, Dewen Hu, Chenping Hou
Summary: With the evolution of data collection ways, label ambiguity has become increasingly common. This paper proposes a new framework (Label Distribution Changing Learning) to tackle the problem of label distribution changing with sample space expanding. By re-scaling distributions and using constraint factors to estimate the labels of new classes, this approach achieves effective results in addressing label ambiguity and estimating emotions.
JOURNAL OF MACHINE LEARNING RESEARCH
(2023)
Article
Computer Science, Artificial Intelligence
Yunan Lu, Weiwei Li, Huaxiong Li, Xiuyi Jia
Summary: Label distribution learning (LDL) addresses label ambiguity by transforming logical labels into label distributions. We propose a generative label enhancement model that utilizes variational Bayes inference to infer label distributions while preserving label ranking and correlation. Extensive experiments validate the effectiveness of our method.
Article
Computer Science, Artificial Intelligence
Jintao Huang, Chi-Man Vong, Guangtai Wang, Wenbin Qian, Yimin Zhou, C. L. Philip Chen
Summary: This article proposes a novel unified LE-LDL learning framework called SGP-FBLS, which improves feature mapping ability using a polynomial-based fuzzy system, mines potential LS using a graph regularized-based objective function (GP-FBLS), and transfers LS and weighted parameters using a weight stacked strategy, thus enhancing the accuracy and efficiency of LDL.
IEEE TRANSACTIONS ON FUZZY SYSTEMS
(2023)
Article
Computer Science, Artificial Intelligence
Xiangyu Li, Xinjie Liang, Gongning Luo, Wei Wang, Kuanquan Wang, Shuo Li
Summary: In this study, we propose a novel framework based on label distribution learning (LDL) paradigm to estimate tumor cellularity (TC). The proposed framework addresses the challenges of inter-rater ambiguity exploitation, label distribution generation, and accurate TC value recovery. It achieves superior performance compared to existing methods on the TC estimation task.
MEDICAL IMAGE ANALYSIS
(2023)
Article
Computer Science, Artificial Intelligence
Jing Wang, Xin Geng
Summary: In this article, a new label distribution learning method called LDL-LDM is proposed to efficiently exploit both global and local label correlations in a data-driven way. It learns the label distribution manifold to capture the correlations among labels and handles incomplete label distribution learning by jointly learning label distribution and label distribution manifold.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS
(2023)
Article
Computer Science, Artificial Intelligence
Qinghai Zheng, Jihua Zhu, Haoyu Tang, Xinyuan Liu, Zhongyu Li, Huimin Lu
Summary: Recently, label distribution learning (LDL) has gained attention in machine learning. Label distributions describe instances with multiple labels of varying intensities, accommodating more general scenes. To solve the problem of unavailable label distributions in real-world applications, two novel label enhancement methods, LESC and gLESC, are proposed. These methods utilize the correlations between samples in the feature and label space to boost label enhancement performance. Extensive experiments on 14 benchmark datasets demonstrate the effectiveness and superiority of the proposed methods.
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING
(2023)
Article
Computer Science, Artificial Intelligence
Jing Zhang, Hong Tao, Tingjin Luo, Chenping Hou
Summary: Label Distribution Learning (LDL) is a popular approach for addressing label ambiguity, but incomplete labels can degrade performance. To tackle this, we propose a Safe Incomplete LDL method (SILDL) that learns a classifier to prevent incomplete labeled instances from worsening performance.
PATTERN RECOGNITION
(2022)
Article
Computer Science, Artificial Intelligence
Ning Xu, Yong-Di Wu, Congyu Qiao, Yi Ren, Minxue Zhang, Xin Geng
Summary: This paper proposes a novel approach for multi-view partial multi-label learning, which learns predictive model and incorrect-labeling model jointly by incorporating the topological structure of the feature space. The experimental results on real-world datasets validate the effectiveness of the proposed approach for solving such learning problems.
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING
(2023)
Article
Computer Science, Artificial Intelligence
Xinyue Dong, Shilin Gu, Wenzhang Zhuge, Tingjin Luo, Chenping Hou
Summary: This paper discusses active learning methods in label distribution learning and proposes a strategy named Active Label Distribution Learning (ALDL) to select the most informative instances by quantifying the disagreement of unlabeled instances. The ALDL strategy maintains composing a committee with selected LDL algorithms to measure the value of unlabeled instances, and it uses a weight vector for both parts.
Article
Computer Science, Artificial Intelligence
Chao Tan, Sheng Chen, Xin Geng, Genlin Ji
Summary: In this paper, a novel label distribution manifold learning (LDML) method is proposed for accurately solving the multilabel distribution learning problem. Through manifold learning and multi-output kernel regression, accurate label distributions can be estimated and an enhanced maximum entropy model is formed. Experimental results demonstrate the advantages of the proposed LDML method in terms of learning accuracy.
PATTERN RECOGNITION
(2023)
Article
Computer Science, Artificial Intelligence
Manuel Gonzalez, German Gonzalez-Almagro, Isaac Triguero, Jose-Ramon Cano, Salvador Garcia
Summary: This paper introduces a decomposition-fusion method, DF-LDL, to address issues in label distribution learning. By utilizing a one-vs-one decomposition strategy and an effective fusion method, the algorithm shows significant improvements in performance.
INFORMATION FUSION
(2021)
Article
Engineering, Electrical & Electronic
Li Yu, Junyang Li, Farhad Pakdaman, Miaogen Ling, Moncef Gabbouj
Summary: No-Reference Image Quality Assessment aims to evaluate image perceptual quality based on human perception. Many studies have used Transformers to simulate the human visual system by assigning different self-attention mechanisms to distinguish image regions. However, the quadratic computational complexity of self-attention is time-consuming and expensive. We propose a lightweight attention mechanism using decomposed large-kernel convolutions to extract multiscale features, and a novel feature enhancement module to simulate the human visual system. Additionally, we compensate for information loss caused by image resizing with supplementary features from natural scene statistics. Experimental results on five standard datasets demonstrate that our proposed method outperforms existing approaches while significantly reducing computational costs.
IEEE SIGNAL PROCESSING LETTERS
(2023)