Abstracts Track 2026


Nr: 170
Title:

Publication-Centred Knowledge Management for Reconstructing Programme-Level Knowledge and Establishing an Evidence Layer for Strategic Evidence-Based Evaluation and Management

Authors:

Kotaro Ochi, Hiroshi Usami and Koji Okamoto

Abstract: The Eichi Programme is Japan's national research programme supporting the decommissioning of the Fukushima Daiichi Nuclear Power Station. Since 2011, more than 210 projects have been conducted through fundamental research, problem-solving research, international collaboration, and human resource development. Although these projects have generated outputs, research knowledge remains fragmented at project and publication levels, limiting programme-level understanding of evidence accumulation and uncertainty reduction. Existing knowledge management approaches organise research outputs at the project or publication level, providing limited support for integrating distributed scientific evidence into programme-level knowledge. This study proposes Publication-centred Relational Integrated Scientific Mapping (PRISM), a knowledge management framework for reconstructing programme-level knowledge from publication-based evidence. Eighty JAEA Review reports and project information were integrated into a KnowledgeBase containing research objectives, outcomes, future challenges, and normalised knowledge tags. Projects were classified using three knowledge dimensions—Core Technology, Technical Function, and Research Need Category—for reproducible programme-level analysis. Analysis of the KnowledgeBase revealed three findings: (1) a multidisciplinary technology portfolio spanning eight major scientific and engineering domains; (2) a Technical Function classification linking scientific disciplines with engineering capabilities; and (3) complementary relationships between Core Technologies and Research Need Categories for reducing decommissioning uncertainty. These findings show the programme functions as an interconnected knowledge system rather than isolated projects. The KnowledgeBase provides an Evidence Layer linking scientific evidence, technological capabilities, and decommissioning uncertainties.

Nr: 193
Title:

Joint Model Matching (JMM): Trajectory-Based Matching for Prognostic Confounding by Longitudinal Biomarkers in Observational Survival Analysis

Authors:

Seong Pyo Kim, Minseon Park and Hyung-Jin Yoon

Abstract: Introduction Causal effects in longitudinal medical data are often estimated by time-dependent propensity score matching within a Cox framework (TD-Cox PSM), which matches on the current biomarker value. When that biomarker changes over time, this ignores the subject-specific trajectory that predicts the outcome. For example, in renin-angiotensin system inhibitor (RASi) discontinuation for advanced chronic kidney disease, TD-Cox PSM misses the estimated glomerular filtration rate (eGFR) trajectory, which predicts kidney failure and drives discontinuation. No established method matches on the pre-treatment trajectory of such a prognostic biomarker. Methods We developed Joint Model Matching (JMM), which fits a shared-random-effects joint model linking longitudinal and survival submodels, and matches on estimated trajectory features (subject-specific slopes and random effects) rather than a cross-sectional propensity score. JMM was compared with TD-Cox PSM in a plasmode simulation based on KNOW-CKD, a Korean cohort of 2,238 pre-dialysis CKD patients, estimating RASi discontinuation's effect on a renal composite outcome (true HR = 1.28 from an RCT). Confounding strength and measurement error were each varied at two levels; bias, RMSE, and 95% CI coverage were compared (Table 1). Results Under weak confounding, the methods performed comparably. Under strong confounding, JMM was better on all three metrics. The gap was widest with added measurement error: TD-Cox PSM coverage fell to 63.6% (lowest across scenarios) versus 86.0% for JMM, with lower bias (0.063 vs 0.133) and RMSE (0.110 vs 0.162). Conclusions JMM matches on the prognostic trajectory rather than the current biomarker value, and its advantage grows with time-varying confounding and measurement error—the latter because the joint model absorbs observation noise that degrades raw-biomarker methods. It suits settings where confounding operates through a repeatedly measured prognostic biomarker.

Nr: 195
Title:

Adaptive Knowledge-First Alternating Training for LightKG

Authors:

Yoki Fujiwara, Keiko Ono and Takuya Futagami

Abstract: Graph Neural Networks (GNNs) drive Knowledge Graph-aware Recommender Systems (KGRSs), propagating relational signals indispensable for cold-start and zero-shot recommendation. LightKG, the lightweight SOTA, matches this accuracy at lower cost, but never explicitly optimizes KG structure, allowing recommendation signals to dominate before relational knowledge is internalized. We propose a knowledge-first alternating training framework that improves relation utilization while preserving LightKG's lightweight architecture. The model first acquires KG structure through an explicit KG objective, then alternates recommendation and KG learning, ensuring downstream ranking is grounded in relational knowledge rather than overwriting it. To further balance the two objectives across datasets with different graph compositions, we introduce a ratio-aware gating mechanism that modulates KG training according to the ratio of KG triplets to user–item interactions. We evaluated our method on two long-tailed benchmarks at opposite ends of the KG/CF ratio spectrum: Last.FM (KG/CF ≈ 0.73) and Book-Crossing (KG/CF ≈ 6.15). On balanced Last.FM, our method yielded modest but consistent gains of roughly +1% relative over the LightKG SOTA baseline across every metric. On KG-heavy Book-Crossing, naive alternating training failed, but once ratio-aware gating throttled KG updates, the model achieved +3–5% relative improvement across all metrics — a larger gain that directly reflected the regime the gating is designed for, showing the improvements stemmed from better relation utilization, not brute-force KG weighting. Broader KG/CF regimes remain to be tested, but the takeaway is decisive: the degree to which the KG should drive representation learning is itself a dataset-dependent quantity ー and the proposed ratio-aware gating supplies exactly that training-time knob, keeping LightKG lightweight while lifting performance in zero-shot and long-tail settings.

Nr: 196
Title:

Domain Ontologies for Psychotherapy: The Case of Transference Modeling

Authors:

Weronika Teresa Adrian and Jadwiga Wojnicka

Abstract: There are several therapeutic approaches in psychotherapy. The problem is that the multitude of approaches makes it difficult to study and compare the therapeutic process and effects of therapy, or even establish a common nomenclature. Phenomena that are important in some modalities are marginalized in others, the same or similar effects can be achieved using different methods, and the same phenomena can be called differently. Because psychotherapy aspires to be a scientific method of treatment, it is thus necessary to organize the knowledge, standardize the nomenclature, and integrate the approaches, or at least provide a mapping among them that will facilitate semi-formal analysis and comparison. Creating domain ontologies and analyzing their interconnections is the basis for solving the above problem. In this work, we present our modeling experiments in the field of psychotherapy, with the focus on psychodynamic, cognitive-behavioral, and systemic psychotherapeutic approaches. We show an example of the phenomenon called transference. In the psychodynamic approach, it is considered an important diagnostic element that helps to identify internal, unconscious relational patterns during therapy; in the cognitive-behavioral approach, it is treated as a typical automatic thought pattern and little attention is paid to it; the systemic approach recognizes its existence, but focuses on the relationships between individual members of therapy instead of treating it as a tool. In our model, we capture different perspectives, and highlight the variety of granularity between different approaches, laying the ground for further analysis. A formal model of such a complex theory, although limited in its expressivity, could serve as help in education of future therapist. Our work fits into the current debate in Poland regarding the training of therapists and the requirements for the admission to the profession. We believe it can be an important voice of interdisciplinary research. This research was supported by the programme "The Excellence Initiative - Research University" for the AGH University of Krakow (no. 18442).

Nr: 197
Title:

The Essence of a Musical Work: An Ontological Approach to Copyright Protection in the Era of Generative AI

Authors:

Weronika Teresa Adrian and Katarzyna du Vall

Abstract: This study aims to model the ontological structure of a musical work to address contemporary intellectual property challenges in the age of generative AI. The proposed model seeks to identify the essential components of musical identity and originality, providing a foundation for copyright analysis and plagiarism detection. Building on Ingarden's ontology of artworks, the study develops a multilayer ontological model of musical works using formal ontology languages. The model represents musical identity through interconnected structural, expressive, performative, acoustic, and semantic dimensions. We believe that musical originality cannot be reduced to isolated musical features but emerges from specific configurations of ontologically relevant elements. The model provides a conceptual framework rather than a complete legal test. Further empirical validation using real-world musical corpora is required. The ontology may serve as the foundation for AI-assisted copyright analysis and early warning systems capable of identifying potentially infringing musical content before publication. While formal ontologies have been extensively applied to textual and knowledge domains, no comprehensive ontology of musical works has been developed specifically for copyright protection and plagiarism detection in the context of generative AI. The study contributes simultaneously to computer science, digital humanities, musicology, and intellectual property law. This research was supported by the programme "The Excellence Initiative - Research University" for the AGH University of Krakow (no. 18442).