[Date]:2026/04/06～2026/04/06

Fukuoka Statistics DS workshop 2026｜2026a016

Overview

• How to hold: In-person
• Venue: IMI Conference Room (W1-D-414), West Zone 1, Ito campus, Kyushu University
• Main language: English
• Type/Category: Grant for General Research-Workshop（Ⅱ）
• Title of Research Project: Fukuoka Statistics DS workshop 2026
• Principal Investigator: Yuichi Goto（Kyushu University/ Assistant professor）
• Research Period: April 6, 2026.
• Open to the Public: April 6, 2026.
• Details of the Research Plan: https://joint2.imi.kyushu-u.ac.jp/en_research_chooses/view/2026a016

Program

April 6, 2026.

10:00-10:10　Opening

Yuichi Goto (Kyushu University)

10:10-10:35

Yuichi Goto (assistant Professor,  Kyushu University)

Title :Mixed difference integer-valued GARCH model for $\mathbb{Z}$-valued Time Series

Abstract: We propose a flexible modeling framework for integer-valued time series that admits both positive and negative values, which naturally arise, for instance, after detrending nonnegative integer-valued data. The model extends the INGARCH-type structure by separately modeling the positive and negative parts and by incorporating a switching mechanism governed by Bernoulli dynamics, thereby enabling the accommodation of bimodality and sign transitions with dependence on past observations.
We establish sufficient conditions for stationarity, ergodicity, and $\beta$-mixing of the proposed process, characterized by the spectral radius of a parameter matrix and by bounds on the switching probabilities. For parameter estimation, we introduce a mixed Poisson quasi-maximum likelihood estimator, and we show that it is consistent and asymptotically normally distributed under standard regularity assumptions.This talk is based on joint work with A. Aknouche (Qassim University) and C. Francq (CREST and University of Lille).

10:35-11:00

Takato Hashino (master Student, Kyushu University)

Title :Finiteness of the Shannon Entropy for Regularly Varying Distributions

Abstract: We will discuss the finiteness of the Shannon entropy for probability distributions defined on countably infinite classes. The Shannon entropy is a fundamental quantity in information theory and statistics, but it may diverge for heavy-tailed distributions. Clarifying conditions under which the Shannon entropy remains finite is an important theoretical issue, particularly in statistical models where heavy-tailed behavior naturally arises.

The talk focuses on discrete distributions whose tail probabilities exhibit regular variation. After reviewing basic concepts of regularly varying functions and distributions, we discuss general conditions that determine whether the Shannon entropy of such distributions is finite. In particular, the analysis highlights the role of the tail index in governing the convergence or divergence of the Shannon entropy, while boundary cases depend on the behavior of the slowly varying component.

As a brief application, we consider the marginal Pitman–Yor process. We show that this distribution belongs to the class of regularly varying distributions, thereby giving a clear explanation for the finiteness of its Shannon entropy. This example illustrates how the general result can be applied to a concrete heavy-tailed model. 

This work is based on the preprint: Hashino and Tsukuda (arXiv:2602.08347).

11:00-11:25

Koji Tsukuda (associate Professor, Kyushu University)

Title :Weak convergences in separable Hilbert spaces with applications to statistics

Abstract: This talk presents an overview of weak convergence for random processes in separable Hilbert spaces and its applications in statistics. I outline the basic framework for the weak convergence of Hilbert‑space‑valued random variables and show how limit theorems in separable Hilbert spaces provide a useful perspective on a range of statistical testing problems. Applications include parametric change‑point tests, nonparametric goodness‑of‑fit tests, tests in high‑dimensional linear models, and functional central limit theorems for selected combinatorial structures. Together, these examples show how weak convergence supports the asymptotic analysis of modern statistical procedures.

12:05-12:30

Sujin Park (Postdoc, Seoul National University）

Title :A Study on Time Series Forecasting Models Using Various Deep Learning Techniques

Abstract: This study proposes advanced time-series forecasting frameworks using diverse deep learning and machine learning architectures to address increasing volatility in energy and environmental sectors. The research is structured into three specialized domains: solar energy, particulate matter (PM), and electricity market prices.

First, for solar energy, we developed a CNN-Catboost hybrid model for short-term forecasting to capture complex spatial-temporal features, along with an SVR-based algorithm integrated with K-means clustering and PDF estimation for medium-term predictions. Second, a hybrid forecasting model for particulate matter (PM10) was established by combining various ensemble learning methods to improve accuracy in urban areas. Third, the study investigates System Marginal Price (SMP) forecasting utilizing UMAP for high-dimensional feature reduction, enabling robust analysis amidst market fluctuations.

Empirical results demonstrate that the proposed models outperform baseline methods in terms of MAE and RMSE. These findings provide a validated computational foundation for reducing uncertainty in energy and environmental systems, offering critical insights for efficient policy-making and industrial operations.

12:30-12:55

Sunmin Oh (PhD Student, Seoul National University）

Title :Relaxed Sparsest-Permutation Formulation for Causal Discovery at Scale

Abstract: Despite the growing availability of large datasets, causal structure learning remains computationally prohibitive at scale. We revisit sparsest-permutation learning for linear structural equation models and show that exact Cholesky factorization is unnecessary for structure recovery. This observation motivates a support-level relaxation that searches for sparse triangular factors over a precision-support screening graph. The relaxed formulation can be efficiently evaluated via masked zero-fill incomplete Cholesky factorization, enabling scalable comparison of candidate orderings. At the population level, we establish soundness for Markov equivalence class (MEC) recovery under no-cancellation and sparsest Markov representation assumptions, as well as robustness to ordering misspecification. Motivated by these guarantees, we introduce SCOPE, a sparse-Cholesky pipeline that provides a scalable implementation of the relaxed formulation. Experiments on synthetic and real datasets demonstrate that SCOPE matches the MEC recovery accuracy of substantially slower baselines, while achieving significantly reduced runtime and scaling to 10k variables.

14:30-15:10

Sanghun Jeong (assistant Professor, Changwon National University)

Title :Wasserstein Centroid-Based Binary Classification for Distributional Data.

Abstract: In this presentation, I will introduce a novel binary classification method specifically designed for analyzing random objects, such as distributional data, that reside in non-linear spaces. Unlike traditional classification approaches that rely on Euclidean distance, our proposed method utilizes the Wasserstein distance measured locally in a tangent space to accurately account for the dissimilarities between objects under intrinsic conditions. To resolve the mathematical challenge of computing group variances across different spaces, this approach employs logarithmic mapping and a parallel transport operator. Consequently, it effectively integrates the central and dispersion characteristics of these complex objects. Through repeated simulations and a practical application classifying glioblastoma multiforme MRI pixel intensity data, I will demonstrate how this Wasserstein centroid-based approach significantly outperforms conventional classification methods.

15:30-16:10

Il-Youp Kwak (associate Professor, Chung-Ang University)

Title :Advances in Sequence-Based Deep Learning for Gene Expression Prediction from Promoter Regions

Abstract: In this presentation, we will discuss our research applying deep learning methods to genomic data. The DREAM Challenge (https://dreamchallenges.org/), organized by IBM Research, provides open platforms where important problems in biomedical science are released with corresponding datasets, enabling researchers worldwide to collaboratively develop solutions and publish findings. Our team participated in the 2022 competition Predicting Gene Expression Using Millions of Random Promoter Sequences, which focused on predicting gene expression levels from promoter region DNA sequences. With access to approximately six million data samples, we developed a variant of the Transformer architecture, revising  Confermer architecture, and achieved third place in the challenge. In this presentation, we will outline how our model was constructed and refined during the competition, and describe how subsequent collaborations with the organizers and other top-performing teams have led to extended joint research efforts.

16:10-17:00　Open discussion

Participants 

Chung-Ang University

Il-Youp Kwak
Taehwan Kim

Seoul National University

Gunwoong Park
Donguk Shin
Sujin Park
Hyewon Park
Sunmin Oh 
Daehun Lee
Euijong Song
Seunghwan Noh
Byeongguk Kang
Junhyoung Chung

Changwon National University

Sanghun Jeong

Registration

Advance registration is required. 
Free participation fee.
(Registration also requires Organizing Committee members and speakers.)
Registration may be closed when the number of participants reaches the maximum.

＼Please go to the following link for registration.／