Addressing the Problems of Data Variations, Quality, and Scarcity in Training Deep Neural Networks

An Imbalance-Resilient Deep Classification Framework for Reliable Analytics in Real-World Data Streams

Data streams in the real world often display severe class imbalance, with a critical under-representation of minority-class instances, resulting in biased and unreliable classification results from conventional deep learning frameworks. We introduce a new Imbalance-Resilient Deep Classification Framework that is defined for robust and stable analytics deployed on dynamic non-stationary data streams. The proposed framework combines adaptive resampling strategies, cost-sensitive learning mechanisms, and deep neural architectures to address the adverse impact of imbalanced class distributions. To cope with this, we propose a new dynamic threshold calibration module which gradually updates decision boundaries according to the changing characteristics of data streams in order to ensure relatively high classification accuracy overtime. The framework also includes an ensemble-driven feature extraction pipeline, enabling to learn discriminative representation for minority classes without compromising majority-class accuracy. The experiments conducted on benchmark and real-world imbalanced streaming datasets show that the proposed approach outperforms existing state-of-the-art algorithms, both in G-Mean, AUC-ROC and F1-Score metrics. Our findings validate the convenience, scalability and broad applicability of our framework in important applications like fraud detection, medical diagnosis, and network intrusion where imbalance constantly arises as a prominent issue.

GAN-powered oversampling with strengthened sample distribution for class overlapping imbalanced data

A minority-friendly learning method for object detection: joint optimization of oversampling-driven data balancing and loss reconstruction

Class imbalance in object detection severely degrades the recognition of minority categories, leading to unstable training and unsatisfactory detection accuracy for rare objects. Existing methods usually address this issue either at the data level or at the loss level alone, while lacking effective coordination between the two. To this end, this paper proposes a minority-friendly learning method for object detection through the joint optimization of oversampling-driven data rebalancing and loss reconstruction. Specifically, a category-aware oversampling strategy is introduced to alleviate sample distribution bias, and a minority-oriented loss reformulation mechanism is designed to strengthen the contribution of underrepresented classes during training. The proposed method introduces little additional model complexity and can be integrated into the training pipeline without modifying the detector backbone or detection head. Experiments on an imbalanced MS COCO 2017 setting demonstrate consistent gains in mAP, Recall, and minority-class AP, while maintaining stable performance on majority classes.

Tackling extreme class imbalance in NIDS: a hybrid architecture integrating SMOTE and CNN-BiLSTM with focal loss

With the rapid development of network technologies, Network Intrusion Detection Systems (NIDS) play a critical role in securing networks against malicious attacks. However, existing machine learning and deep learning models often struggle to effectively identify minority class attacks, such as User-to-Root (U2R) and Remote-to-Local (R2L), when faced with highly imbalanced network traffic data. To tackle this extreme class imbalance challenge, this paper proposes a hybrid NIDS framework integrating Synthetic Minority Over-sampling Technique (SMOTE), a CNN-BiLSTM architecture, and Focal Loss. First, at the data level, the SMOTE algorithm is employed to synthetically oversample minority classes in the training set, establishing a balanced data distribution. Subsequently, at the algorithmic level, a Convolutional Neural Network-Bidirectional Long Short-Term Memory (CNN-BiLSTM) deep neural network is designed to jointly extract the local spatial topologies and global temporal features of network traffic. Finally, the Focal Loss function is introduced to further optimize the model by dynamically down-weighting easily classified majority samples, forcing the network to focus on hard-to-classify boundary samples. Extensive experiments on the benchmark NSL-KDD dataset demonstrate that the proposed method achieves a breakthrough in minority attack recognition. Specifically, the F1-Scores for R2L and U2R are elevated to 0.38 and 0.34, respectively, significantly outperforming traditional machine learning and standard deep learning baselines while maintaining robust detection capabilities for majority attacks.

Class-Balanced Softmax: A Bayes Theory-Based Method for Long-Tailed Recognition

Deep learning models using traditional softmax classifiers have achieved remarkable success in various classification tasks. However, their performance degrades significantly on imbalanced datasets. Although Balanced Softmax is widely adopted as a state-of-the-art rebalancing method, it possesses inherent limitations, such as yielding disproportionately lower testing accuracy for tail classes. To mitigate these shortcomings, we propose the Class-Balanced Softmax (CBS). Rooted in a theoretical Bayesian framework and a heuristic power-law assumption, the CBS is a simple logit adjustment that is computationally inexpensive and easily integrated into existing pipelines. Furthermore, we characterise a fundamental phenomenon in models trained on imbalanced data, termed the preference issue, wherein models exhibit higher training error and a larger generalisation gap for classes with limited data. To quantify this issue, we introduce a novel metric and demonstrate that CBS effectively mitigates the preference issue. Extensive experiments on large-scale benchmarks show that CBS is highly scalable and outperforms existing methods, including Balanced Softmax.

CSW-AL: A category-scaling weight schema for imbalanced multi-domain active learning.

This work introduces a comprehensive framework featuring a dynamic category-scaling weight mechanism that hierarchically addresses imbalance at the sample and domain levels, and builds a progressively refined "Ideal Domain" through the adaptive integration of well-performing domains.