Offline reinforcement learning (RL) suffers from the out-of-distribution (OOD) problem, which significantly limits its deployment in real-world scenarios. A critical yet under-explored aspect of this challenge is OOD states: once an agent enters an unfamiliar state, it may fall into a trajectory of successive OOD states, resulting in severe performance degradation. We propose an effective and simple solution: Addressing out-of-distribution States for offline reinforcement learning with interference robustness (ASIR). At its core, ASIR introduces a transition-constrained policy, enforced through a regularization term derived from an inverse dynamics model (IDM). The IDM pre-trained on the offline dataset, captures the underlying dynamics of valid state transitions. By encouraging the policy to prefer actions that lead to familiar and predictable next states, ASIR keeps the agent within the data manifold and away from OOD regions. Theoretically, our method not only corrects for OOD states but also applies to OOD actions. ASIR also demonstrates robust performance against disturbances. In experiments, our method achieves state-of-the-art results on the D4RL and NeoRL benchmark tasks. Furthermore, to evaluate robustness under real-world disturbances, we inject Gaussian noise into the actions to push the agent toward OOD states. As the duration of the perturbation increases, ASIR exhibits significantly smaller performance degradation and lower variance, outperforming existing benchmark methods in challenging scenarios.
Early identification of borrowers at risk of transitioning into long-term delinquency is a critical challenge in credit risk management. In real-world financial settings, risk prediction is often constrained by short observation windows and highly aggregated tabular credit data, limiting the applicability of conventional time-series models and handcrafted feature-based approaches. To address these limitations, this study proposes a deep feature extraction framework that employs representation learning as a complementary enhancement to traditional classifiers rather than as a standalone predictive model. The proposed framework integrates self-attention-based global interaction learning with TabNet-inspired soft feature selection to construct latent representations that effectively capture delinquency-related risk signals from low-information credit data. Extensive experiments on a real-world multi-debtor credit dataset demonstrate the potential of the proposed feature extraction method to enhance long-term delinquency transition predictions. Beyond predictive accuracy, analyses of feature-space distances and model explainability reveal enhanced class separability and provide interpretable insights into the underlying drivers of delinquency transitions. Overall, the results suggest that the proposed deep representation learning offers a practical, interpretable, and effective solution for early delinquency transition detection in realistic tabular credit data environments with limited temporal information.
The growing adoption of ML/AI systems in high-stakes contexts has raised concerns about the quality, fairness and privacy of training data. Although synthetic data offers a promising way to address these concerns, the imbalance of sensitive categorical features still represents a source of bias. This paper presents two alternative approaches for a Variational AutoEncoder (VAE) training loss function that penalise deviations from a balanced distribution of sensitive feature classes. These are an additive balance component ( ℬℒ_term ) and a Simpson-index-based reweighting component ( ℬℒ_weight ). In comparison with other approaches in literature, which alter the architecture of the VAE, our method is a lightweight and tunable mechanism to steer the generation towards balance. An empirical evaluation was conducted on 1,260 synthetic datasets generated from three widely used benchmark datasets (Adult, South German Credit, and COMPAS), across two batch-epoch configurations, seven loss function configurations and thirty train/evaluation splits. The results show that ℬℒ_term reduces sensitive feature imbalance and improves statistical parity, while preserving utility and privacy at levels comparable to the baseline. The ℬℒ_weight formulation, although capable of reducing imbalance, exhibits a weaker impact on fairness and may degrade utility and privacy at high magnitudes. The proposed contribution offers practitioners a simple, tunable mechanism for improving the balance and fairness of synthetic tabular data generation without compromising utility or privacy.
The use of Machine Learning (ML) algorithms in decision-making processes has increased significantly in recent years, providing alternatives to biased human decisions. However, ML algorithms can also exhibit bias, leading to discrimination against individuals or groups that are identified by sensitive attributes such as gender or race. This bias often arises from the imbalanced representation of demographic groups in the training data, particularly in the distribution of favorable class labels. Mitigating representation bias during the training phase is crucial to ensure fair application of ML models in decision-making processes. Synthetic data generation has been considered an effective technique for mitigating representation bias. Existing techniques fail to accurately identify the regions in the data space where adding more synthetic examples is effective. In this paper, we present a preprocessing framework for generating synthetic examples for the minority groups to create a balanced and fair dataset for training fair ML models. The proposed framework identifies subgroups with lower ratios in getting the favorable class label and identifies the boundaries of the group examples with favorable class labels. Subsequently, SMOTE-NC is used for generating synthetic examples around the group boundary until the group reaches a target imbalance ratio. Synthetic data generation is done carefully to minimize the effect on the performance of the ML models. The performance and fairness of the proposed framework are evaluated using standard performance and fairness measures. The results indicate that the proposed framework successfully reduces representation bias and enhances fairness across demographic groups, while maintaining competitive predictive performance relative to state-of-the-art bias mitigation techniques.
Unmanned Aerial Vehicles (UAVs) increasingly collaborate across Flying Ad-Hoc Networks (FANETs) to detect distributed network anomalies. However, deploying collaborative intrusion detection under emerging privacy regulations (e.g., GDPR “Right to be Forgotten”) faces two fundamental bottlenecks: (i) structural feature heterogeneity arising from diverse onboard sensor suites, which causes standard Federated Learning (FedAvg) to fail; and (ii) the severe computational burden of retraining models from scratch when a drone leaves or requests data retraction. To overcome these challenges, this paper proposes FedUnlearn-HetFL, a novel federated unlearning framework tailored for feature-heterogeneous UAV networks. The framework integrates localized attention-gated autoencoders that project asymmetric client feature dimensions (40, 45, 50, and 62 features) into a unified 64-dimensional latent space, coupled with a shared 1D-CNN backbone. When an unlearning request is initiated, FedUnlearn-HetFL executes targeted unlearning via selective layer-frozen gradient ascent on the classification head, resets local input projections, and applies rapid stabilization fine-tuning across retained nodes. Evaluated on the ISOT Drone benchmark under Dirichlet non-IID partitioning, the unlearned model collapses target node accuracy from 99.16 × execution speedup and reduces communication rounds by 95.0
