Quantum-enhanced Machine Learning Improves Rare Event Detection in Tabular Data.

QCLMixNet, a novel machine learning framework, improves rare event detection in imbalanced tabular data, crucial for expert systems. It utilises quantum inspired layers and adaptive data mixing, achieving superior performance in macroF1 and recall across 18 datasets, consistently exceeding 20 existing benchmarks and establishing a new standard.

The reliable identification of rare events constitutes a significant challenge for expert systems operating with real-world tabular data, particularly where imbalances between classes exist. Conventional methods frequently struggle with overfitting, noisy labels, and poor performance in data-sparse regions, hindering the accurate detection of critical, albeit infrequent, instances. Researchers now present a novel approach, QCL-MixNet, a quantum-informed contrastive learning framework designed to address these limitations through the integration of sinusoidal transformations, adaptive data augmentation, and a hybrid loss function. This work, detailed in a forthcoming publication, is the result of collaboration between Md Abrar Jahina of the University of Southern California, Adiba Abid from Khulna University of Engineering & Technology, and M. F. Mridha of American International University-Bangladesh, and is titled “Quantum-Informed Contrastive Learning with Dynamic Mixup Augmentation for Class-Imbalanced Expert Systems”.
Expert systems frequently encounter difficulties when analysing tabular data characterised by imbalanced classes, a situation where identifying rare, yet critical, instances is paramount for reliable operation. Researchers now present QCL-MixNet, a novel framework designed to improve classification performance in these scenarios, addressing limitations found in existing techniques such as cost-sensitive learning and graph neural networks. These conventional methods often struggle with overfitting, inaccuracies stemming from noisy labels, and poor performance when dealing with data sparsely distributed in feature space.

QCL-MixNet incorporates three key innovations to mitigate these issues, beginning with a quantum-inspired layer that models complex interactions between features. This layer utilises sinusoidal transformations and gated attention mechanisms, allowing the system to capture nuanced relationships within the data that simpler models often miss. Furthermore, the framework implements a sample-aware mixup strategy, a data augmentation technique that intelligently interpolates feature representations of similar instances, effectively enhancing the representation of the minority class and improving its detectability. Mixup creates synthetic data points by combining existing ones, thereby increasing the diversity of the training set.

The system further refines its performance through a hybrid loss function, combining focal reweighting, supervised contrastive learning, triplet margin loss, and variance regularization. Focal reweighting prioritises the learning of difficult-to-classify instances, while supervised contrastive learning and triplet margin loss encourage the model to learn discriminative feature embeddings. Feature embeddings represent data points as vectors in a multi-dimensional space, where similar points are closer together. Variance regularization promotes robustness and prevents overfitting, ultimately improving both the compactness of data within each class and the separation between different classes.

Extensive evaluation on eighteen real-world imbalanced datasets, encompassing both binary and multi-class problems, demonstrates QCL-MixNet’s superior performance. The framework consistently outperforms twenty state-of-the-art machine learning, deep learning, and graph neural network-based baselines, achieving substantial gains in macro-F1 score and recall. Macro-F1 score provides a balanced measure of precision and recall across all classes, while recall measures the ability to identify all relevant instances. Ablation studies confirm the importance of each architectural component, validating the design choices and establishing QCL-MixNet as a new benchmark for handling tabular imbalance in expert systems.

Researchers introduce QCL-MixNet, leveraging Informed Contrastive Learning and k-nearest neighbor (kNN) guided dynamic mixup for robust classification under imbalance. The quantum-inspired layer captures complex feature interactions, the sample-aware mixup strategy enhances minority class representation, and the hybrid loss function promotes robust and accurate classification.

Theoretical analysis and empirical results demonstrate that QCL-MixNet effectively addresses the challenges posed by imbalanced tabular data, offering a significant improvement over existing methods.