Institution profile

Japan Advanced Institute of Science and Technology

Academic institutionasia · jp
Official website
Research library138linked papers
Opportunities0open roles
Selected work

Representative Papers

Estimating the Optimal Number of Clusters in Categorical Data Clustering by Silhouette Coefficient

Nov 29, 2019Communications in Computer and Information Science

Addressing the challenge of automatic optimal cluster number selection in categorical data clustering, this paper introduces the first unsupervised, parameter-free method that adapts the silhouette coefficient to purely categorical spaces. The core methodological contribution lies in designing a dissimilarity measure grounded in Hamming distance and attribute matching, reformulating intra- and inter-cluster silhouette computations accordingly, and establishing a comprehensive clustering evaluation framework tailored specifically for categorical data. Extensive experiments across multiple real-world categorical datasets demonstrate that the proposed approach significantly outperforms mainstream indices—including the Calinski–Harabasz (CH) index and Gap Statistic—reducing average cluster number estimation error by 37%. The method exhibits strong robustness, intrinsic interpretability, and broad generalizability across diverse categorical domains.

107 citations4 influentialRead paper

Lower Bound on the Error Rate of Genie-Aided Lattice Decoding

Jun 26, 2022International Symposium on Information Theory

This paper addresses the fundamental performance limits of lattice decoders under finite power constraints. We propose a genie-aided decoding framework and jointly optimize the real-valued scaling factor α and Voronoi-region covering-sphere geometry to derive, for the first time, a theoretical word error rate (WER) lower bound applicable to arbitrary lattices—including classical ones such as E₈ and BW₁₆—along with a closed-form high-SNR asymptotic expression. Furthermore, we introduce an efficient WER estimation method based on the effective sphere, significantly improving prediction accuracy: at a target WER of 10⁻⁴, E₈ and BW₁₆ achieve coding gains of 0.5 dB and 0.4 dB, respectively. Finally, integrating α-MMSE scaling, CRC embedding, and polar lattice codes, we design a prototype decoder with blocklength n = 128, experimentally validating the achievability of the derived theoretical bounds in practical constructions.

3 citationsRead paper

A Distributed Multi-Modal Sensing Approach for Human Activity Recognition in Real-Time Human-Robot Collaboration

Oct 01, 2025IEEE Robotics and Automation Letters

This work addresses the challenge of accurately and reliably recognizing human hand activity intentions in real time during human-robot collaboration. To this end, the authors propose a distributed multimodal perception architecture that fuses data from a wearable inertial measurement unit (IMU)-based data glove and vision-based tactile sensors to enable high-precision, real-time identification of dynamic hand activities during physical human-robot interaction. The system employs a modular design combined with a real-time sequential classification algorithm, achieving consistently high recognition accuracy across offline evaluation, static real-time testing, and realistic collaborative scenarios. Experimental results validate the effectiveness and practicality of the proposed approach, demonstrating its potential to serve as a robust perceptual foundation for dynamic human-robot collaboration.

1 citationsRead paper

Data Augmented Pipeline for Legal Information Extraction and Reasoning

Jun 16, 2025International Conference on Artificial Intelligence and Law

Legal information extraction faces significant challenges due to the high cost of manual annotation and data scarcity. To address these issues, this work proposes a concise and general-purpose data augmentation method based on large language models (LLMs), which constructs an end-to-end augmentation pipeline to automatically generate high-quality training samples, thereby reducing reliance on human-annotated data. The proposed approach not only substantially improves the performance, robustness, and generalization capability of legal information extraction systems but also demonstrates strong transferability to other natural language processing tasks, highlighting its versatility and practical utility.

1 citationsRead paper
Recent publications

Latest Papers