Jun-Geol Baek

Tabular data is a data format commonly used in various industries, including finance, medical services, and manufacturing. However, deep learning research for this data type has lagged behind the progress seen in computer vision and natural language processing. Real-world applications often face severe label scarcity due to high labeling costs. Traditional Gradient Boosting Decision Trees (GBDTs) and existing deep learning architectures struggle to achieve high predictive performance under these constraints. This study proposes the Gated Attention Tabular Encoder (GATE) and a two-stage semi-supervised learning pipeline. GATE effectively captures complex feature interactions by using global context from self-attention to dynamically regulate the flow of feature information. The proposed pipeline uses a teacher-student self-distillation framework to extract robust representations from unlabeled data. Then, it employs SoftNCA-based metric learning alongside high-confidence pseudo-labeling to optimize the embedding space with minimal labeled instances. Experimental results on public benchmark data showed that the proposed framework consistently achieved superior average performance across all label-ratio scenarios. The framework achieved an average accuracy of 62.23% with only 1% of the labeled data. It also achieved the highest average accuracy of 89.26% in a fully supervised setting, confirming its universal competitiveness. Through additional experiments, we confirmed that the proposed pipeline constructs a structurally superior representation space and achieves greater class separation compared to supervised learning models. In addition, in terms of computational efficiency, the proposed pipeline achieved this performance by using only 50% of the training parameters and 33% of the FLOPs compared to the comparative transformer-based model, FT-Transformer, and using only one-hundredth of the number of parameters compared to complex SSL models such as SAINT. This study presents a highly effective and computationally efficient paradigm for learning tabular representations in label-deficient industrial environments. The code is available at https://github.com/yongeun-jang/GATE.

This paper proposes a framework that improves classification performance for multivariate time series data while providing an objective assessment of each variable’s influence, including interaction effects. While convolutional neural networks (CNNs) offer significant advantages in analyzing multivariate time series data, the structural limitations of CNNs have restricted their ability to detect statistical interactions. Our approach creatively modifies convolutional filters and layer structures, enabling feature extraction that captures the influence of pairwise interactions. These extracted features are processed by interpretable models to calculate feature importance, enabling in-depth causal analysis by quantifying both individual and pairwise variable effects. In addition, the proposed method enhances the overall classification performance of multivariate time series data. Synthetic data experiments verified that the proposed method effectively extracted relevant features that explain pairwise interactions. In addition, in the multivariate time series classification experiments using real data, the proposed method demonstrated superior performance compared to baseline methods. These results suggest that the proposed approach is a practical and interpretable solution for multivariate time series classification tasks in domains where variable interactions play a decisive role, such as healthcare, finance, and manufacturing.

This paper presents stable reinforcement learning through adversarial imitation learning (SRAIL), a two-phase learning framework for automatic proportional integral derivative (PID) tuning in the absence of high-quality demonstration data. First, adversarial imitation learning (AIL) is used to learn an initial control policy using imperfect demonstration data. Subsequently, reinforcement learning (RL) refines this initial policy to derive the optimal PID control policy. SRAIL is designed to ensure stable and safe exploration of the RL agent, even with imperfect demonstration data, thereby enabling efficient and automatic tuning in industrial environments. Conventional PID tuning methods are time-consuming owing to system complexity and lack of prior knowledge. Furthermore, the random exploratory nature of RL presents significant risks during the early training phase, limiting its practical application. SRAIL addresses these challenges by ensuring stable control performance despite the absence of high-quality demonstration data and by enhancing the convergence speed of RL. The experimental results demonstrate that SRAIL minimizes random exploration in the initial learning phase, ensuring safety, and highlighting its potential for implementation of RL in real-world industrial environments.

Accurate cycle time (CT) prediction is crucial in semiconductor manufacturing. Hybrid models integrating classification and prediction models can enhance CT prediction accuracy. However, existing methods have limitations, including challenges in capturing the dynamic conditions of the production line and optimizing job classification to ensure high CT prediction performance. In this study, we propose a novel hybrid framework for predicting process step-level CT in semiconductor wafer fabrication, thereby addressing the limitations of previous methods. Moreover, the paper formalizes and introduces dynamically changing manufacturing environment attributes as variables that contribute to CT. The proposed method combines deep embedded clustering (DEC) with a regression ensemble network. First, the DEC extracts cluster-friendly representative features from high-dimensional CT datasets and classifies jobs accordingly. Then, a weighted ensemble approach merges regression networks based on cluster membership probabilities. Unlike existing methods that separately handle feature extraction, job classification, and CT prediction, the proposed unified network synchronizes these processes. Experimental results using real-world operational data from a semiconductor manufacturing system indicate that the proposed prediction method considerably outperforms previous approaches in terms of prediction accuracy. To the best of our knowledge, this is the first study to integrate deep clustering with a regression ensemble network for predicting cycle time at the process step level in semiconductor manufacturing. By synchronizing feature extraction, clustering, and prediction tasks, the proposed framework achieves enhanced accuracy and robustness in dynamically changing manufacturing environments.