ICGK 2025

Opening

Keynote: Managing Knowledge Space

Speaker: Steffen Staab

• Data Spaces: Similar to Data Lake $\to$ to Knowledge Spaces - require semantics, applications + data space
• Knowledge Space: contains multiple KGs/Applications - combine, integrate through semantics
• Context:
  • DFG Cluster of Excellence – construction industry, novel materials, concepts, …
    • Knowledge Space: architects $\to$ engineers $\to$ … / past: exchange using paper (or pdf) - works for waterfall process (does not work if anything changes!)
  • Circular Factory for future (CRC)
    • re-use of products by inspection
    • find defects using CT/… - very individual tasks!
    • Difficult to do in high-paying country (DE, EU)
    • $\approx$ 20 profs with students (a lot of perspectives!)
    • Knowledge managed in ontology + KG $\to$ store data in an agile manner
      • allow easy modification
      • hierarchy of linked ontologies (very interesting for our HEREDITARY use case?); but share data easily!
      • Model statistics using distribution definitions (i.e. add Gaussian variance property to number literal)
  • Supporting people over years - SHACL Constraints as Programming Constraints
    • infer types from SHACL constraints within SPARQL query
    • combining KGs - CONSTRUCT query for derived shapes; derived shapes from query
    • Conflicting Shapes / Epistemic (conflicting statements) - resolve Bilattice $\mathcal{FO}\mathcal{UR}$
  • Graph Embeddings for Graph Recommendations (TransE - embed in vector spaces by operations in ${\ell }_2$ space)
    • They extend that concept to boxes/sets/affine spaces - be extended to concepts within KGs
  • KG foundation models
    • learn/pre-train model for graph embeddings $\to$ reason on unseen KG based on pattern
    • their work: extend structure by text embedding
    • Oven benchmark: not only Q&A - but unique ID for disambiguation
  • Will LMs “eat” the KGs?
    • Most use cases: Q&A on DBpedia
    • But other use case: finance, factories, construction, medicine - require accurate and precise software! - KGs with correct relations are the best there! - combine in RAG/Foundational models Quote:

Instead of starting with a goal for a paper, rather start with a project and once that one start writing the paper!

Session 1: Knowledge Graph Construction

Generic and domain-specific AI-powered knowledge graph construction

Authors: Irene Kilanioti and George Angelos Papadopoulos

• Creating SDG from dataset, enhance it with ML methods

Topic-Enhanced Instruction Tuning for Automatic Knowledge Graph Construction

Authors: Xiaoyu Sun, Ke Liang, Sihang Zhou, and Jie Chen (Online)

• More accurate extraction using LM to reduce search space

Design of Cybersecurity Knowledge Graph Systems Based on Large Language Models

Authors: Yichun Li and Fei Xiong

LLM-Based Construction of Knowledge Graphs for the Analysis of Human Smuggling Networks

Authors: Dipak Meher, Carlotta Domeniconi, and Guadalupe Correa-Cabrera

Mind the Context: Enriching Knowledge Graphs with Rules and Mappings

Authors: Veronica Santos, Daniel de Oliveira, Daniel Schwabe, Edward Hermann Haeusler, Fernanda Baiao, and Sergio Lifschitz

FAIR GraphRAG: A Retrieval-Augmented Generation Approach for Semantic Data Analysis

Authors: Marlena Flüh, Soo-Yon Kim, Carolin Victoria Schneider, and Sandra Geisler

• Graph generation of FAIR datasets with LM integration

Session 3: Multimodal KG Completion & Understanding

DMF-MH: Dual-stage Modality-aware Fusion for Modality Heterogeneity in Multimodal Knowledge Graph Completion

Authors: Ronghua Tian and Hong Yu

FICHAD: Fusion of Image Context and Human-Annotated Descriptions for Multi-Modal Knowledge Graph Completion

Authors: Haodi Ma, Dzmitry Kasinets, and Daisy Wang

• Application: Recommendation of products for ads based on KG
• challenge: align data of different sources

SSGR-AR: Semantic-enhanced Scene Graph Reasoning for Robust Video Action Recognition

Authors: Daxu Shi, Fan Qi, and Changsheng Xu

Efficient Text-video Aligner method for text-video retrieval

Authors: Huaiqi Li and Chunxiao Fan

Improving Graph Embeddings in Machine Learning Using Knowledge Completion with Validation in a Case Study on COVID-19 Spread

Authors: Rosario Napoli, Gabriele Morabito, Antonio Celesti, Massimo Villari, and Maria Fazio

• Introduce Priors to network, analyze propagation network
• Case study on COVID-spread in contact network with transitive networks - add transitive properties to network
• Use PageRank as proxy measure of centrality
• Look at change in embeddings - Q: movement of individual embeddings, performance; A: no space in paper, next paper!

Session 4: Recommendations & Social Graphs

Chair: David Woollard, Standard.ai

Preference-aware Intent Fusion in Multi-Behavior Recommendation

Authors: Shuqing Sun, Peijie Sun, Ruijie Liu, and Dan Guo

ReviewGraph: A Knowledge Graph Embedding Based Framework for Review Rating Prediction with Sentiment Features

Authors: Albert J.W. de Vink, Natalia Amat-Lefort, and Lifeng HAN

• Improve interpretability of review prediction models using KG embeddings
• Compare to more costly LM methods
• Imbalanced class problem - Q: baseline of 60%? How good is random choice? $\to$ Probably also around 60 percent, i.e. what is the metric (weighted accuracy/$F_1$)? - have dummy option (random baseline)
• Sampling approach matters for their scores - oversampling best here…
• This is bachelor thesis, but still quite thorough

Session 5: Ontologies, Semantic & Interoperability

Aligning Tag-based Building Operating Systems with RealEstateCore Ontology for Interoperability, Digital Twin Knowledge Graphs and Energy Management

Authors: Prerna Juhlin, Michael Kleefisch, Charles Steinmetz, Gösta Stomberg, Matthias Schlöder, Eike Fokken, and Philipp Bauer

• Renewable/Energy Savings for Building, local energy management integration into a tool
• Moving from tag-based onto to formal class-based onto (openBOS+)
• Used Azure Digital Twin Platform (Time series management on graphs)

Semantic and Structural Drift in Financial Knowledge Graphs: A Robustness Analysis of GNN-based Fraud Detectors

Authors: Rener S. de Menezes and Raimir H. Filho

Keynote 2: Changsheng Xu

Session 8 — Time-Series & Anomaly Detection I

Chair: Benedikt Kantz

M-TSINR: Multiscale Implicit Neural Representations via Mamba Encoder for Time Series Anomaly Detection

Authors: Ke Liu, Mengxuan Li, Qianqian Shen, Yang Gao, and Haishuai Wang

• Q:

  • Do you know the paper by Keogh (Time Series benchmarks are flawed), have you considered their UCR benchmark? x
  • Have you looked at multiple runs of your system (variability?) your results are really close to the other systems, why it might be interesting to evaluate that further? - They averaged their results and took to the mean

Large Language Models for Anomalous Event Detection from Temporal Point Processes

Authors: Qinming Zhuang, Peng Zhang, and Hong Yang

• Q:
  • How does this method compare to the simpler Poisson/Hawke approaches?
  • You talk about efficiency - how fast is your LM approach compared to the simple approaches?
• Sadly no answer in session

Adaptive-GraphSketch: Real-Time Edge Anomaly Detection via Multi-Layer Tensor Sketching and Temporal Decay

Authors: Ocheme Anthony Ekle and William Eberle

• Q:
  • $\mathcal{O}(1)$ is claimed, but also in respect to nodes/timesteps considered?
  • Why should one consider this system over AnoEdge-L as it seems to 10x more efficient?
• Sadly no answer in session

A Novel Transfer Learning Approach for Detecting Unseen Anomalies

Authors: Khan Mohammad Al Farabi and Gagan Agrawal

• Q
  • why are your scores almost perfect? - any intuition?
  • You use target domain data to relabel source data to label the source data, seems like test data leakage? - is this applicable to a real world dataset? (no, still weird results?)
  • Do you plan to release a Github repository? (Repository not really working…)

Survey: Generalization of Graph Anomaly Detection: From Transfer Learning to Foundation Models

Authors: Junjun Pan, Yu Zheng, Yue Tan, and Yixin Liu

• Q
  • Are there any advantages to the older methods? Can they still perform similarly to the newer methods if they are used in principled manners?

Session 10 — Graph Systems, Benchmarks & Applications

Chair: Prerna Juhli

KubeGraphBench: Benchmarking Graph Databases for Kubernetes Observability

Authors: Charles O’Brian, Tarek Zaarour, Ahmed Khalid, and Ahmed Zahran

• Benchmark dataset within DELL for observability of kubernetes
• Custom ontology
• Use KubeInsights/KWOK to simulate configuration and querying on a single machine for up to 1k nodes, 50 pods each
• Memgraph for smaller graphs with a lot of reads, neo4j for bigger graphs, memgraph has a lower memory footprint, neo4j has a higher memory usage.
• Did not look into indices
• Q: “significantly” did p-test of results? (tip graph labeling) - did not do statistical testing!

Intent-Driven Smart Manufacturing Integrating Knowledge Graphs and Large Language Models

Authors: Jradi Takoua, John Violos, Dimitrios Spatharakis, Lydia Mavraidi, Ioannis Dimolitsas, Aris Leivadeas, and Symeon Papavassiliou

• Extract intent from query, construct subgraph, and build graph from domain data
• Domains:
  • manufacturing,
  • IoT,
  • shipping

Memory-Efficient Information Filtering in Contrastive Learning for Temporal Knowledge Graph Reasoning

Authors: Eshani Fernando, Michael Adjeisah, Jian Chang, and Jian Jun Zhang

• Subsample graph to foster local-global analysis of Temporal KGs
• select based on importance based on prior events + degree of node
• Quite efficient, at little performance loss

Session 12: Causality, Queries & Explainability

Chair: Aris Leivadeas

CausaMap: A Semi-supervised Map For Causal Text Mining

Authors: Sami Diaf

• Steiner Trees?

On the Development of an Interactive Cause-Effect Learning System (CELS) for a Metal-Forming System Use Case

Authors: Josua Höfgen, Birgit Vogel-Heuser, Dominik Hujo-Lauer, Michael Lechner, and Marion Merklein

• Metal forming KG integration for wear/product quality
• Application: metal stamping/forming
• Problem: unclear terms/vocabulary $\to$ inconsistencies
  • starting point: excel list of 150 terms for terminology
• All stakeholders get access to system
• Cause-effect graph: two spectrums: computability (is the graph in some way computable?) / representation (from PPT to RDF/GraphDB data)
• Key goal: interactive system for of cause-effect graph for all stakeholders!

Computational Fact-Checking of Online Discourse: Scoring Scientific Accuracy in Climate Change Related News Articles

Authors: Tim Wittenborg, Constantin Sebastian Tremel, Oliver Karras, and Sören Auer

• Problem: IPCC produces a 1k pages long PDF, not in a structured format (but no existing KG!)

• KG creation from IPCC not a problem, but getting claims from videos is a magnitude slower

• Existing systems do not reuse/efficiency problems

• Their contribution:

  • use secondary literature
  • open-source pipeline to extract text, align them
  • extend KG using good sources, verify statements over KG
  • Expert interviews to improve system, gather feedback

• independently of their work: ClimateKG; further work: integration