Jin-Taek Seong

Click-through rate (CTR) prediction requires balancing predictive performance with model interpretability, which is essential for accountability in industrial advertising systems. While transformer-based architectures achieve high accuracy through implicit learning of complex feature interactions, they often operate as black boxes with limited transparency in decision-making. This paper presents FT-AFM, a hybrid framework that combines a transformer-based feature tokenization backbone for context-aware representation learning with an Attentional Factorization Machine (AFM) for explicit attention-weighted pairwise interaction modeling. Through shared embeddings across two coordinated branches, FT-AFM captures global field-wise dependencies while revealing the relative importance of individual feature interactions. Evaluations on three large-scale benchmarks—Criteo, Outbrain, and Avazu—show that FT-AFM achieves competitive AUC scores of 0.8076, 0.7003, and 0.7386, respectively, while remaining close to strong transformer-based baselines. We further introduce a faithfulness-based evaluation protocol to examine whether the learned attention weights are meaningfully related to the model’s reliance on specific interactions. Results suggest that FT-AFM offers a practical accuracy–interpretability trade-off for large-scale CTR prediction by enabling explicit interaction-level analysis while maintaining competitive predictive performance.

Combinatorial optimization within complex networks poses a substantial challenge in the realm of soft computing, particularly when decision spaces are characterized by dynamism, stochasticity, and risk weighting. This study presents an advanced Ant Colony Optimization (ACO) framework designed for sequential path optimization on weighted contact networks, where transmission pathways are conceptualized as a Traveling Salesman Problem (TSP). The proposed methodology integrates rank-based elitist reinforcement, greedy seeding, edge-weight normalization, and periodic 2-opt local refinement to improve convergence stability and solution quality. To address domain uncertainty, the pheromone update mechanism is integrated with a Markovian SEIR process, facilitating probabilistic state transitions that guide reinforcement towards feasible low-risk paths. Experiments conducted on synthetic graphs and real-world SocioPatterns proximity networks demonstrate that the enhanced ACO achieves rapid convergence within 2550 iterations, attains near-optimal path costs, and consistently surpasses traditional metaheuristics such as Genetic Algorithms, Tabu Search, and standard ACO variants. The findings highlight the efficacy of integrating soft computing optimization with stochastic state dynamics, providing a scalable and interpretable framework for risk-aware path optimization in complex networked systems, with epidemic transmission modeling serving as an illustrative application domain.

Integrating high-frequency sequential signals with low-frequency contextual descriptors into a unified deep encoder is a recurring challenge in computational modelling, exemplified by cross-sectional stock ranking where price dynamics must be jointly modelled with quarterly accounting fundamentals. Existing approaches use late concatenation, where the contextual signal influences only the final prediction head and cannot shape upstream feature extraction. We propose Feature-wise Linear Modulation (FiLM) as an intermediate conditioning mechanism: fundamentals generate per-channel scaling (gamma) and shifting (beta) parameters that affinely transform the encoder’s intermediate representations before aggregation. The same price sequence thus yields different temporal features depending on the firm’s fundamental profile, which we hypothesise reduces signal variability across heterogeneous market regimes by allowing the encoder to amplify or suppress patterns based on contextual quality. We instantiate FiLM across recurrent (LSTM), convolutional (TCN), and attention-based (iTransformer) encoders, evaluated on China A-share equities (2010–2024). Across the convolutional and recurrent encoder families, the primary benefit of FiLM conditioning is improved signal stability—formally measured as the standard deviation of RankIC across rebalancing dates—and risk-adjusted performance, rather than mean predictive accuracy. The gain depends critically on where modulation is applied: pre-aggregation conditioning on temporally-rich representations produces the largest variance reduction. FiLM-TCN, which modulates the convolutional feature map before pooling, achieves RankIC of 0.1415, annualised Sharpe of 1.633, and IC hit rate of 80.4% net of transaction costs. The insight that intermediate conditioning improves signal stability rather than raw accuracy may inform analogous fusion problems in other sequential modelling domains.

This study re-evaluated the role of environmental, social, and governance (ESG) characteristics in China's A-share market over the period 2012Q1 to 2021Q3. Contrary to the prevailing view of ESG as a defensive, risk-mitigating attribute, we showed that passive high-ESG strategies fail to generate excess returns. To address this inefficiency, we introduced a meta-learning framework that integrates three heterogeneous base learners—LightGBM for nonlinear cross-sectional effects, ridge regression as a linear benchmark, and a temporal convolutional network (TCN) for sequential earnings dynamics—whose out-of-fold predictions were stacked via a second-stage ridge meta-learner trained on a held-out validation set. This architecture predicts next-quarter earnings per share (EPS-TTM) and ranks firms by fundamental growth potential within a high-ESG universe screened to the top tercile of ESG scores. The framework was trained on 2012Q1–2017Q4, validated on 2018Q1–2019Q1, and evaluated out-of-sample on 2019Q2–2021Q3 using a strict chronological split that eliminates look-ahead bias. We documented a clear performance hierarchy. While the broad high-ESG universe exhibits negative alpha relative to the CSI 300 benchmark, a refined top 50% portfolio largely neutralizes this underperformance, and economically meaningful excess returns are concentrated in the top 10% portfolio, with annualized alpha exceeding 40% over the out-of-sample evaluation period (Sharpe ratio: 1.11; market beta: 1.16). Industry-neutralized robustness tests confirmed that this alpha is attributable to stock-level fundamental prediction rather than passive sector concentration. These results support a risk-signaling interpretation of ESG in emerging markets: ESG-related return premia in China reflect compensation for growth-related systematic risk under conditions of disclosure heterogeneity and information asymmetry, rather than downside protection.

Classical image denoising methods remain relevant in practical scenarios where training data or noise models are unavailable, yet their performance is highly sensitive to parameter selection. Non-Local Means (NLM) is a representative example whose effectiveness depends critically on smoothing strength, patch size, and search window configuration. This paper formulates NLM parameter selection as a black-box optimization problem under unknown noise conditions and employs adaptive metaheuristic optimization strategies for this task. We propose an adaptive optimization framework that integrates rank-based perturbation, opposition-based learning, Lévy-flight exploration, and noise-aware parameter constraints to improve robustness and convergence. The proposed method is evaluated against fixed-parameter NLM and NLM optimized using standard evolutionary algorithms under identical protocols. Experiments on three sets of datatset demonstrate consistent improvements in PSNR and SSIM, highlighting the continued relevance of adaptive optimization for classical denoising.

This paper presents a hybrid deep learning framework for the automated optical inspection (AOI) of vertical cavity surface-emitting laser (VCSEL) semiconductor devices. Manual inspection is limited by subjective inconsistency and operator fatigue, while high-end commercial AOI systems impose substantial costs that are often impractical for small and medium sized manufacturers. The proposed system adopts a decision-level fusion architecture in which defect-specific binary classifiers are aggregated through an OR-gate. This design prioritizes recall so that any defect flagged by at least one sub-classifier triggers rejection, reducing the risk of defect escape. An industrial dataset of 22, 410 images with severe class imbalance (e.g., crack defects comprising less than 0.4% of all labels) was used for training, with targeted augmentation applied to minority classes. Five fold cross-validation yielded an overall accuracy of 98.7% and an F1-score of 93.5%. Deployment on an active production line reduced per-unit inspection time from 17.7 s to 1.5 s (an 89% reduction) and recorded a secondary defect escape rate of 0.00%.

Reliable defect detection in coarse-wavelength division multiplexing (CWDM) optical modules is critical for ensuring stable high-speed optical communication and minimizing network disruptions. Traditional inspection methods, such as manual video screening and optical testing, are labor-intensive, prone to human error, and often fail to identify underlying defect patterns, thereby necessitating multiple rounds of testing. To overcome these limitations, this paper proposes a deep learning-based multimodal learning framework for defect prediction that integrates manufacturing process data (tabular features) and optical eye diagram images from the module’s receiver. Leveraging the complementary nature of these data sources enhances defect detection performance, enabling the proposed model to surpass conventional single modal alternatives. The model was comprehensively evaluated on a real CWDM module dataset through cross-validation, based on performance metrics such as accuracy and F1-score. The multimodal learning schemes achieved an accuracy of 91.0% and an F1-score exceeding 90%, significantly outperforming both traditional single modal approaches and manual inspection techniques. These results demonstrate the effectiveness of multimodal learning in improving defect prediction accuracy, reducing reliance on repetitive testing, and lowering overall inspection costs. The proposed approach represents a scalable and efficient solution for automated quality control in optical module manufacturing, with potential applications in optical network maintenance and defect detection for other photonic components.