Imbalanced Distribution Research Articles

Large-Scale Object Detection in the Wild With Imbalanced Data Distribution, and Multi-Labels.

Training with more data has always been the most stable and effective way of improving performance in the deep learning era. The Open Images dataset, the largest object detection dataset, presents significant opportunities and challenges for general and sophisticated scenarios. However, its semi-automatic collection and labeling process, designed to manage the huge data scale, leads to label-related problems, including explicit or implicit multiple labels per object and highly imbalanced label distribution. In this work, we quantitatively analyze the major problems in large-scale object detection and provide a detailed yet comprehensive demonstration of our solutions. First, we design a concurrent softmax to handle the multi-label problems in object detection and propose a soft-balance sampling method with a hybrid training scheduler to address the label imbalance. This approach yields a notable improvement of 3.34 points, achieving the best single-model performance with a mAP of 60.90% on the public object detection test set of Open Images. Then, we introduce a well-designed ensemble mechanism that substantially enhances the performance of the single model, achieving an overall mAP of 67.17%, which is 4.29 points higher than the best result from the Open Images public test 2018.

Triplet Adaptation Framework for Robust Semi-Supervised Learning.

Semi-supervised learning (SSL) suffers from severe performance degradation when labeled and unlabeled data come from inconsistent and imbalanced distribution. Nonetheless, there is a lack of theoretical guidance regarding a remedy for this issue. To bridge the gap between theoretical insights and practical solutions, we embark to an analysis of generalization bound of classic SSL algorithms. This analysis reveals that distribution inconsistency between unlabeled and labeled data can cause a significant generalization error bound. Motivated by this theoretical insight, we present a Triplet Adaptation Framework (TAF) to reduce the distribution divergence and improve the generalization of SSL models. TAF comprises three adapters: Balanced Residual Adapter, aiming to map the class distribution of labeled and unlabeled data to a uniform distribution for reducing class distribution divergence; Representation Adapter, aiming to map the representation distribution of unlabeled data to labeled one for reducing representation distribution divergence; and Pseudo-Label Adapter, aiming to align the predicted pseudo-labels with the class distribution of unlabeled data, thereby preventing erroneous pseudo-labels from exacerbating representation divergence. These three adapters collaborate synergistically to reduce the generalization bound, ultimately achieving a more robust and generalizable SSL model. Extensive experiments across various robust SSL scenarios validate the efficacy of our method.

Instance Consistency Regularization for Semi-Supervised 3D Instance Segmentation.

Large-scale datasets with point-wise semantic and instance labels are crucial to 3D instance segmentation but also expensive. To leverage unlabeled data, previous semi-supervised 3D instance segmentation approaches have explored self-training frameworks, which rely on high-quality pseudo labels for consistency regularization. They intuitively utilize both instance and semantic pseudo labels in a joint learning manner. However, semantic pseudo labels contain numerous noise derived from the imbalanced category distribution and natural confusion of similar but distinct categories, which leads to severe collapses in self-training. Motivated by the observation that 3D instances are non-overlapping and spatially separable, we ask whether we can solely rely on instance consistency regularization for improved semi-supervised segmentation. To this end, we propose a novel self-training network InsTeacher3D to explore and exploit pure instance knowledge from unlabeled data. We first build a parallel base 3D instance segmentation model DKNet, which distinguishes each instance from the others via discriminative instance kernels without reliance on semantic segmentation. Based on DKNet, we further design a novel instance consistency regularization framework to generate and leverage high-quality instance pseudo labels. Experimental results on multiple large-scale datasets show that the InsTeacher3D significantly outperforms prior state-of-the-art semi-supervised approaches.

Many-Objective Jaccard-Based Evolutionary Feature Selection for High-Dimensional Imbalanced Data Classification.

Filters and wrappers represent two mainstream approaches to feature selection (FS). Although evolutionary wrapper-based FS outperforms filters in addressing real-world classification problems, extending these methods to high-dimensional, many-objective optimization problems with imbalanced data poses substantial challenges. Overcoming computational costs and identifying suitable performance metrics are vital for navigating search operation complexities. Here, we propose using the Jaccard similarity (JS) in a set-based evolutionary many-objective (JSEMO) FS search, addressing both evolutionary FS and imbalanced classifier choice concurrently. This study highlights the mutual influence between these aspects, impacting overall algorithm performance. JSEMO integrates JS into population initialization, reproduction, and elitism steps, enhancing diversity and avoiding duplicate solutions. The set-based variation operator utilizes intersection and union operators for compatibility with binary coding. We also introduce a double-weighted KNN (KNN2W) classifier with four supportive objectives as a many-objective FS problem to handle imbalanced distributions. Compared with 20 methods on 15 benchmark problems, JSEMO produces distinct optimal features, significantly improving overall accuracy, balance accuracy, and g-mean metrics with comparable feature set size and computational cost. The ablation study underscores the positive impact of all JSEMO components, highlighting the set-based variation operation with JS and KNN2W with relevant evaluation metrics as the most influential aspects.

TC-Sniffer: A Transformer-CNN Bibranch Framework Leveraging Auxiliary VOCs for Few-Shot UBC Diagnosis via Electronic Noses.

Utilizing electronic noses (e-noses) with pattern recognition algorithms offers a promising noninvasive method for the early detection of urinary bladder cancer (UBC). However, limited clinical samples often hinder existing artificial intelligence (AI)-assisted diagnosis. This paper proposes TC-Sniffer, a novel bibranch framework for few-shot UBC diagnosis, leveraging easily obtainable UBC-related volatile organic components (VOCs) as auxiliary classification categories. These VOCs are biomarkers of UBC, helping the model learn more UBC-specific features, reducing overfitting in small sample scenarios, and reflecting the imbalanced distribution of clinical samples. TC-Sniffer employs intensity-based augmentation to address small sample size issues and focal loss to alleviate model bias due to the class imbalance caused by auxiliary VOCs. The architecture combines transformers and temporal convolutional neural networks to capture long- and short-range dependencies, achieving comprehensive representation learning. Additionally, feature-level constraints further enhance the learning of distinctive features for each class. Experimental results using e-nose data collected from a custom-designed sensor array show that TC-Sniffer significantly surpasses existing approaches, achieving a mean accuracy of 92.95% with only five UBC training samples. Moreover, the fine-grained classification results show that the model can distinguish between nonmuscle-invasive bladder cancer (NMIBC) and muscle-invasive bladder cancer (MIBC), both of which are subtypes of UBC. The superior performance of TC-Sniffer highlights its potential for timely and accurate cancer diagnosis in challenging clinical settings.

Gyroscope in-assembly drift anomaly detection based on decision re-optimized deep auto-encoder

Abstract In dynamic servo gyroscope assembly, frequent drift performance overproof causes massive part re-assembly and write-off, for which in-assembly gyro drift anomaly detection is required. However, when utilizing common anomaly detection methods, imbalanced assembly data distribution causes severe accuracy reduction and false alarms. To tackle these problems, we first propose a decision re-optimized deep autoencoder model to conduct in-assembly drift anomaly detection under imbalanced assembly data distribution. Second, a decision-based training strategy is introduced to lower the false alarm rate in anomaly detection, for which models based on different training strategies are compared for better performance. Third, a modified SMOTE data augmentation method is utilized to settle the impact of data imbalance under small-sample conditions. The experimental results show that the proposed method can achieve in-assembly drift anomaly detection under imbalanced data distribution with high precision and outperforms all other existing methods, lowering the assembly repetition rate and improving assembly efficiency.

Abstract 4139194: Predicting Cholesterol Screening Behavior After Age 50 Using Machine Learning: Insights from the Health and Retirement Study

Background: In the U.S., about 8% of adults never received cholesterol screening. Although machine learning (ML) has been used to develop decision tools for Atherosclerotic Cardiovascular Disease (ASCVD) risk prediction, its application in behavioral forecasting has not yet been explored in the context of cholesterol screening behaviors. This study aimed to examine the performance and accuracy of ML algorithms in forecasting cholesterol screening behaviors in adults after age 50. Methods: This analysis used deidentified data from the Health and Retirement Study (HRS) 2004-2018. HRS is a longitudinal survey among 23,000 households in the U.S. Participants were excluded from the current analysis if they passed away by 2019, ever had ASCVD or stroke, were under age 50 at baseline, or had missing data in self-reported cholesterol screening. In total, 7176 participants (mean age [SD]=62 [8]) met the inclusion criteria; participants were randomly split into a training set (80%) and a testing set (20%). The synthetic minority oversampling technique was used to solve the imbalance distribution of the rare event. Five ML algorithms were used: random forest, gradient boosting machine (GBM), XGBoost, Support Vector Machine (SVM), and logistic regression. Accuracy, AUROC, and positive predictive value (PPV) were used to compare model performance. The average gain was evaluated for feature importance in the demographic and health domains. Results: In total, 232 (3.2%) respondents did not receive any cholesterol screening from 2008 to 2018. Experiments with five ML algorithms suggested that XGBoost with deeper trees and learning rate performed better in classifying those who did not screen for cholesterol levels over 10 years. Adding prior cholesterol screening history (2004-2006) into the model significantly improved model performance. Hypertension, self-rated health, and smoking were the major health features, while insurance, poverty, and work status were the major demographic features in the predictive model (accuracy=0.97; AUROC=0.88; PPV=0.42). Conclusion: Findings underscore the potential utility of ML models in predicting cholesterol screening behaviors after age 50. This could be the basis for developing decision tools for clinicians to identify those with a lower chance of cholesterol screening or make reminders accordingly. The low-cost predictive model might improve the uptake of preventive screening behaviors in middle-aged and older adults.

A Comprehensive Survey on Rare Event Prediction

Rare event prediction involves identifying and forecasting events with a low probability using machine learning (ML) and data analysis. Due to the imbalanced data distributions, where the frequency of common events vastly outweighs that of rare events, it requires using specialized methods within each step of the ML pipeline, that is, from data processing to algorithms to evaluation protocols. Predicting the occurrences of rare events is important for real-world applications, such as Industry 4.0, and is an active research area in statistical and ML. This article comprehensively reviews the current approaches for rare event prediction along four dimensions: rare event data, data processing, algorithmic approaches, and evaluation approaches. Specifically, we consider 73 datasets from different modalities (i.e., numerical, image, text, and audio), four major categories of data processing, five major algorithmic groupings, and two broader evaluation approaches. This article aims to identify gaps in the current literature and highlight the challenges of predicting rare events. It also suggests potential research directions, which can help guide practitioners and researchers.

Multi‐Objective Federated Averaging Algorithm

ABSTRACTThe recent global trend is the convergence of information and communications technology (ICT). By applying ICT in various fields such as the humanities, new types of products and services are created, and new values that help people's lives can be created. AI can be selected as a representative technology in such convergence ICT. However, applying AI technology to actual production requires ensuring data security. Federated learning (FL) can achieve secure sharing of data, where all parties participate in model training locally and upload it to the server for aggregation. The data never leaves the parties involved, thus solving the problems of data privacy and data silos. However, FL faces issues such as high communication cost, imbalanced performance distribution among participants, and low privacy protection. To achieve a balance between model accuracy, communication cost, fairness, and privacy, this paper proposes a multi‐objective optimization‐based FL algorithm (M‐FedAvg). The multi‐objective optimization problem of maximising the accuracy of the global model, minimising the communication cost, minimising the variance of the accuracy, and minimising the privacy budget is solved by NSGA‐III. The experimental results show that the algorithm proposed can effectively reduce the communication cost of FL and achieve privacy protection for participants without affecting the accuracy of the global model.

Multi-classification prediction of PM2.5 concentration based on improved adaptive boosting rotation forest

Developing efficient, stable, and user-friendly methods and technologies for predicting air quality has contributed to environmental research and management. Most traditional machine learning (ML) models often struggle to efficiently process extensive air quality data and grapple with the challenge of imbalanced data distributions. To this end, we introduced a novel multi-strategy collaborative approach that incorporates weighted feature selection, an adaptive enhanced rotation forest algorithm, and Bayesian Optimization for parameter tuning. Moreover, to improve the transparency in black box ML models, the novel Shapley Additive Explanations (SHAP) method was applied to interpret the outputs and analyze the importance of individual variables. Through experiments on real air quality datasets, we have verified the accuracy of our proposed method in classifying air quality levels. Notably, our model demonstrated an average test set accuracy improvement of approximately 10 % in cities like Beijing, Tianjin, Shijiazhuang, and Baoding after hyperparameter optimization using Bayesian methods. Furthermore, compared to alternative algorithms, our model showed an improvement of 1–2 % in evaluation metrics. By introducing novel methodologies and tools, our research contributes significantly to the advancement of air quality classification technology and holds profound implications for informing environmental policy decisions.

A general image classification model for agricultural machinery trajectory mode recognition

Field-road trajectory classification is a crucial task for agricultural machinery behavior mode recognition, aiming to distinguish field operation mode and road driving mode automatically. However, the imbalanced distribution of agricultural machine trajectories brings challenges for the field-road trajectory classification task. Additionally, most existing field-road trajectory classification methods have certain shortcomings. For instance, they encounter difficulties in accurately representing the state of agricultural machinery movement using the current features. The data transformation process often leads to information loss, and the model’s generalization capabilities are limited. The performance of the models is constrained by each of these elements. To address these shortcomings, this paper introduces a general image classification model for agricultural machinery trajectory mode recognition named ATRNet. First, to address the issue of imbalanced field-road proportions in agricultural machinery trajectory data, a Conditional Tabular Generative Adversarial Network (CTGAN) is employed to generate quasi trajectories, balancing the distribution of positive and negative samples in the data. This step aims to eliminate biases during the model training process. Second, to accurately characterize the motion status of agricultural machinery, we propose a multiangle feature enhancement method to extract rich spatiotemporal features from trajectory data. Finally, different from conventional field-road trajectory classification models that primarily rely on spatial and temporal information for identifying trajectories, we present a lossless trajectory data representation paradigm. This paradigm maps each trajectory point into a “feature map” and uses an image classification model to capture latent feature representations of trajectory points for the recognition of different behavior modes of agricultural machinery. This paradigm can generalize image classification networks to the field-road trajectory classification task, providing a general vision model solution for agricultural machinery trajectory mode recognition. To validate the effectiveness of the ATRNet model, experiments were conducted on real corn and wheat harvester trajectory datasets. The results demonstrate that the proposed model achieves remarkable performance improvements over the state-of-the-art (SOTA) models. In the corn harvester trajectory dataset, ATRNet achieves an accuracy of 92.36% and an F1-score of 92.34%, surpassing existing SOTA models by 3.12% and 12.46%, respectively. Similarly, in the wheat harvester trajectory dataset, ATRNet achieves an accuracy of 92.36% and an F1-score of 92.33%, outperforming the existing optimal algorithm by 4.76% and 18.18%, respectively.

Separating Noisy Samples From Tail Classes for Long-Tailed Image Classification With Label Noise.

Most existing methods that cope with noisy labels usually assume that the classwise data distributions are well balanced. They are difficult to deal with the practical scenarios where training samples have imbalanced distributions, since they are not able to differentiate noisy samples from tail classes' clean samples. This article makes an early effort to tackle the image classification task in which the provided labels are noisy and have a long-tailed distribution. To deal with this problem, we propose a new learning paradigm which can screen out noisy samples by matching between inferences on weak and strong data augmentations. A leave-noise-out regularization (LNOR) is further introduced to eliminate the effect of the recognized noisy samples. Besides, we propose a prediction penalty based on the online classwise confidence levels to avoid the bias toward easy classes which are dominated by head classes. Extensive experiments on five datasets including CIFAR-10, CIFAR-100, MNIST, FashionMNIST, and Clothing1M demonstrate that the proposed method outperforms the existing algorithms for learning with long-tailed distribution and label noise.

Privacy-Preserving Incipient Fault Identification in Distribution Networks Under Small Sample and Imbalanced Data Distribution Conditions

Privacy-Preserving Incipient Fault Identification in Distribution Networks Under Small Sample and Imbalanced Data Distribution Conditions

A Study on the Impact of Overtaking Lane-Changing Behaviour in Expressway Interchange Weaving Areas

This study investigates the overtaking lane-changing (OLC) behaviour in expressway interchange weaving areas, aiming to analyse these behaviours’ causes and potential impacts. Field data are utilised to analyse the statistical characteristics of lane-changing points and spatio-temporal utilisation in weaving areas. A modified NS model, which considers the distribution pattern of vehicle speeds, and a rigid lane-changing rule based on Gaussian distribution are proposed. Additionally, a cellular automaton simulation model is constructed to quantify the influence of OLC behaviour on traffic efficiency and spatio-temporal utilisation based on simulated data. The findings indicate that the imbalanced distribution of lane-changing points and spatio-temporal utilisation in weaving segments, caused by rigid lane-changing behaviour, is an objective factor that triggers OLC behaviour. When the traffic volume in weaving areas ranges from 500 to 1,100 pcu/5 min and the proportion of OLC behaviour is between 0.35 and 0.7, the behaviour will significantly enhance the average vehicle speeds of the outermost lane of the main road and normal rigid lane-changing (NRLC) vehicles, with increases of up to 48% and 51%, respectively. Moreover, OLC behaviour also improves the balance of spatio-temporal utilisation in weaving areas and reduces the average spatio-temporal utilisation. This study clarifies the positive impact of OLC behaviour on expressway interchange weaving areas and provides new research ideas for enhancing the efficiency of these areas.

Monocular Depth and Ego-motion Estimation with Scale Based on Superpixel and Normal Constraints

Three-dimensional perception in intelligent virtual and augmented reality (VR/AR) and autonomous vehicles (AV) applications is critical and attracting significant attention. The self-supervised monocular depth and ego-motion estimation serves as a more intelligent learning approach that provides the required scene depth and location for 3D perception. However, the existing self-supervised learning methods suffer from scale ambiguity, boundary blur, and imbalanced depth distribution, limiting the practical applications of VR/AR and AV. In this article, we propose a new self-supervised learning framework based on superpixel and normal constraints to address these problems. Specifically, we formulate a novel 3D edge structure consistency loss to alleviate the boundary blur of depth estimation. To address the scale ambiguity of estimated depth and ego-motion, we propose a novel surface normal network for efficient camera height estimation. The surface normal network is composed of a deep fusion module and a full-scale hierarchical feature aggregation module. Meanwhile, to realize the global smoothing and boundary discriminability of the predicted normal map, we introduce a novel fusion loss which is based on the consistency constraints of the normal in edge domains and superpixel regions. Experiments are conducted on several benchmarks, and the results illustrate that the proposed approach outperforms the state-of-the-art methods in depth, ego-motion, and surface normal estimation.

Learnable feature alignment with attention-based data augmentation for handling data issue in ancient documents

Recognizing ancient cursive handwritten characters presents unique challenges due to the diversity of writing styles and significant class imbalances, where some characters have disproportionately more samples than others. This imbalance leads to higher misclassification rates for minority classes compared to majority classes. To address these challenges, we propose a novel framework that integrates learnable channel and spatial attention modules to effectively align features between source and target domains for better representation. Our approach incorporates a learnable sequential feature alignment process that dynamically adjusts to the specific characteristics of the data, enhancing the transfer of knowledge across domains. Furthermore, we introduce an attention-based augmentation module to amplify the influence of tail classes. This module leverages class activation maps to identify and augment discriminative features, ensuring the model focuses on the most semantically rich regions, particularly for minority classes. As a result, it aligns the weight norms of minority classes with those of majority classes, effectively mitigating the limitations posed by imbalanced class distributions. This approach effectively mitigates the constraints posed by imbalanced character distributions in ancient handwritten documents. The proposed method increases the accuracy for the CCR, Hanja, Nancho, and Kuzushiji datasets.

Comparison of Machine Learning Models and Feature Importance Investigation of Intelligent Fault Diagnosis Methods for Robots Based on Datasets Across Various Distributions

Intelligent fault detection is an important component of the industrial and automation fields. However, conventional research on intelligent fault detection mainly focuses on industrial production equipment, while there is little research on intelligent fault detection scenarios for robots. Based on a hexapod robot joints dataset, this paper investigates intelligent fault diagnosis tasks in the field of robotics. Firstly, the dataset is clustered into two new datasets with different distributions through k-means method, which is used to simulate the practicality of imbalanced distribution between healthy and fault data. Afterwards, several multioutput machine learning classification models were established to predict robot joints with faults. In two datasets with different distributions, the larger one is used as the training set and the smaller one is used as the testing set. The article compares the performance of these models with prediction accuracy as the main indicator. And based on the results, the paper selects the model with the highest accuracy for further exploration of feature importance. Finally, the article explains the significance of the results and analyzes possible reasons. The experimental results show that the random forest model better overcomes the problem of data distribution differences and has the highest accuracy. In the random forest model, the importance of position data representing position error is higher than that of slope with respect to the axis data representing angle error. This result may be related to the distribution of feature values and the fact that position data contains more crucial information.

Improving long‐tail classification via decoupling and regularisation

AbstractReal‐world data always exhibit an imbalanced and long‐tailed distribution, which leads to poor performance for neural network‐based classification. Existing methods mainly tackle this problem by reweighting the loss function or rebalancing the classifier. However, one crucial aspect overlooked by previous research studies is the imbalanced feature space problem caused by the imbalanced angle distribution. In this paper, the authors shed light on the significance of the angle distribution in achieving a balanced feature space, which is essential for improving model performance under long‐tailed distributions. Nevertheless, it is challenging to effectively balance both the classifier norms and angle distribution due to problems such as the low feature norm. To tackle these challenges, the authors first thoroughly analyse the classifier and feature space by decoupling the classification logits into three key components: classifier norm (i.e. the magnitude of the classifier vector), feature norm (i.e. the magnitude of the feature vector), and cosine similarity between the classifier vector and feature vector. In this way, the authors analyse the change of each component in the training process and reveal three critical problems that should be solved, that is, the imbalanced angle distribution, the lack of feature discrimination, and the low feature norm. Drawing from this analysis, the authors propose a novel loss function that incorporates hyperspherical uniformity, additive angular margin, and feature norm regularisation. Each component of the loss function addresses a specific problem and synergistically contributes to achieving a balanced classifier and feature space. The authors conduct extensive experiments on three popular benchmark datasets including CIFAR‐10/100‐LT, ImageNet‐LT, and iNaturalist 2018. The experimental results demonstrate that the authors’ loss function outperforms several previous state‐of‐the‐art methods in addressing the challenges posed by imbalanced and long‐tailed datasets, that is, by improving upon the best‐performing baselines on CIFAR‐100‐LT by 1.34, 1.41, 1.41 and 1.33, respectively.

Robust Visual Question Answering utilizing Bias Instances and Label Imbalance

Visual Question Answering (VQA) models often suffer from bias issues which cause their predictions to rely on superficial correlations in datasets rather than the intrinsic properties of the task. However, most current debiasing methods primarily capture implicit biases but lack fine granularity, overlooking the need to address biases at a sample level. Additionally, these methods mostly fail to adequately utilize information about the imbalanced distribution of labels within the dataset. In this paper, we address these issues by utilizing Bias Instances and Label Imbalance for robust Visual Question Answering (BILI-VQA), which first identifies samples in the dataset that induce bias, defined as “Bias Instances”, and enhances their visual features during training to mitigate language shortcuts for achieving sample-level debiasing. The proposed method then introduces a new module that analyzes the label distribution, assigning higher weights to less frequent answer categories when computing the loss, thereby reducing the negative impact of label imbalance. Extensive experiments on the VQA-CP v2, VQA-CP v1, and VQA-CE datasets demonstrate the effectiveness of our BILI-VQA approach.

A Semantically Enhanced Label Prediction Method for Imbalanced POI Data Category Distribution

A Semantically Enhanced Label Prediction Method for Imbalanced POI Data Category Distribution