We present the Auto-Encoding Network with Semantic constraints (AE-NS) framework, a multi-task learning approach to fair classification on tabular data. The model jointly trains (i) a transformer encoder producing a shared latent representation Z , (ii) a classification head minimising cross-entropy, and (iii) a shallow reconstruction head preserving feature information in Z . A learnable Lagrangian dual multiplier λ adaptively penalises demographic-parity violation, while an auxiliary orthogonality loss explicitly decorrelates Z from the protected attribute. We frame AE-NS as a primal–dual saddle-point problem in the Fenchel-Lagrange sense, which lets us derive a standard O(1/√(T)) convergence guarantee and motivates future work on learning to optimise the dual step. We evaluate the framework on three benchmark datasets (Adult Income, COMPAS, Credit Default) under a strict fairness tolerance of ϵ =0.01 , using three random seeds and eight baselines. These baselines include five PyTorch models, such as adversarial debiasing and LAFAN, and three sklearn baselines, including Fairlearn's exponentiated-gradient reducer on top of XGBoost. Beyond standard metrics such as accuracy, F1, recall, AUC, DPD, EOD, and AOD, we report a Normalised Mutual Information (NMI) score between Z and the protected attribute, which directly measures latent disentanglement. A new classify-on- Z ablation shows that a logistic regression on the learned latent Z alone matches the accuracy of one trained on the raw features, and even exceeds it on Credit, where Z is 7.7 percentage points more accurate. This confirms that the reconstruction task preserves predictive signal. Training improvements in version 2, specifically label smoothing, gradual warm-up, and cosine annealing, boost Adult accuracy by 2 pp (0.760 → 0.781) and reduce DPD by 60 → 0.025). We further propose an AE-NS Z + Fairlearn post-processing pipeline that extracts the disentangled latent Z and applies Fairlearn's ThresholdOptimizer, achieving DPD = 0.009 on Adult (vs. Fairlearn-DP's 0.013) and DPD = 0.009 on Credit Default (vs. Prejudice Remover's 0.012), with 3-seed mean ± std.
The growing need for flexible techniques to detect multiple anomalous observations in modern data motivates this study. Traditional single-outlier tests are often inadequate because of masking and swamping. This paper proposes a new sequential framework for detecting multiple outliers by extending the relative range statistic. The approach iteratively removes the most extreme observation while the test statistic exceeds a critical threshold. Its performance is evaluated using simulated normal data with mean contamination and compared with the standardized range statistic and the extreme Studentized deviate method. Simulation results show that the relative range has greater power to identify outliers and is more robust than the standardized range statistic, especially at low contamination levels and small sample sizes. Furthermore, applications to real-world datasets confirm the reliability of the sequential relative range approach for detecting the true number of extreme observations.
This thesis delves deeply into and is committed to optimizing a cutting-edge and innovative classification technique, called pairwise comparison (Pcomp) classification. This technique was developed to address the widespread challenge of acquiring data labels in traditional binary classification tasks.To further enhance the performance of Pcomp classification, we introduce the Bounded Linear-Exponential (BLINEX) loss function for noise reduction. By setting a maximum loss limit, the BLINEX loss function effectively mitigates the impact of noise and outliers on the classifier, thereby improving the model's robustness. Additionally, this study incorporates Universum data to strengthen the model's generalization capability. Universum data, which consists of samples located near the classification boundary and are difficult to accurately categorize, provides additional useful information for the classifier. By making proper use of these hard-to-classify samples, we can offer more diverse training data for the classifier, thus enhancing its generalization performance and enabling it to make more accurate and reliable predictions when faced with new data. In summary, by introducing the BLINEX loss function and Universum data, this thesis has conducted a comprehensive optimization of the Pcomp classification method. The aim is to enhance its classification performance and generalization capability, ultimately resulting in the BL-Pcomp-U classification model.
Early identification of pneumonia through medical imaging is a cornerstone of effective patient care, and the professional experience required to interpret chest X-rays presents a significant clinical challenge. As deep learning has risen to occupy this space, the majority of existing models have become supercomputer-heavy. They require large volumes of data and the latest hardware, which are not available in most real clinics. To overcome such limitations, this paper proposes a hybrid classical-quantum system, where a sparsely weighted convolutional neural network (CNN) is combined with a quantum convolutional neural network (QCNN). This work begins by converting images to grayscale and extracting a sparse 16-dimensional set of features using a classical CNN, then by taking advantage of the high-dimensional properties of quantum state space, these features are encoded with angle encoding on a 16-qubit system. Further, a structured QCNN with hierarchical convolution and pooling is then applied to reduce the most significant diagnostic data to a small state. To complete the pipeline, the proposed method performs Pauli-Z measurements to obtain the final output and optimizes the system using classical backpropagation and the quantum parameter-shift rule. The model performs well in accuracy, F1-score, and ROC-AUC, achieving mid-tier to high-tier performance. This work finds that hybrid QCNN models are not merely an abstract curiosity; instead, they represent a viable and resource-efficient future for medical imaging in situations where data availability is constrained.
Understanding how machine learning models interpret mental health expressions across languages is essential for developing reliable, inclusive computational mental health systems. While transformer-based models have shown strong performance in English-language mental health detection, their interpretability and prediction behavior across languages with different linguistic characteristics remain insufficiently understood. This study examines cross-lingual interpretability in mental health natural language processing by analyzing how transformer models process and classify mental health expressions in English and Bengali social media text. To support this investigation, we introduce BMHD, a bilingual English-Bengali mental health dataset comprising 10,672 parallel text samples per language across 10 mental health classes and one neutral class, constructed through systematic data collection, translation, annotation, and human-evaluated validation. Multiple monolingual and multilingual transformer-based models are evaluated under English-only, Bengali-only, and bilingual settings. Beyond overall performance metrics, the analysis focuses on class-wise behavior, token-length effects, cross-lingual error patterns, attribution consistency, and agreement between multiple interpretability methods to explore differences in model interpretation across languages. The results reveal consistent cross-lingual disparities in prediction behavior despite semantic alignment between parallel texts. Models show reduced sensitivity to high-risk suicidal expressions and psychosis-related cues in Bengali, increased confusion between closely related mental health categories such as depression and suicidal thoughts, greater misclassification for a small number of unusually long sequences in the Bengali setting, and consistently low agreement among interpretability methods, particularly for Bengali predictions. These findings indicate that linguistic structure, translation effects, and representational imbalances between languages significantly shape how models interpret mental health signals, affecting both accuracy and interpretability. Overall, this study demonstrates that cross-lingual mental health classification cannot be treated as language-agnostic and highlights the need for linguistically aware evaluation frameworks. The dataset and empirical analyses presented in this work provide a valuable benchmark for future research on interpretable and responsible multilingual NLP for mental health, especially for low-resource languages such as Bengali.
