Create High-Quality Knowledge Graphs

Connected Data London 2025 Masterclass:
Combining Data from Structured and Unstructured Sources to create High-Quality Knowledge Graphs

Thu Nov 20, 2025
09:30-11:30 GMT
https://2025.connected-data.london/talks/combining-data-from-structured-and-unstructured-sources-to-create-high-quality-knowledge-graphs/

• Paco Nathan and Gurpinder Dhillon @ Senzing.com
• slides: https://drive.proton.me/urls/S7352D48GG#eGkDq6kWhLFs

Overview

Integrating structured and unstructured data sources into high-quality knowledge graphs is an incredibly common need in production use cases. Downstream there may be several patterns of usage for a high-quality KG, such as graph analytics, dashboards, GraphRAG, question/answer chat bots, agents, tools, memory, planners, and so on.

Throughout this masterclass, we will leverage insights from the tutorial steps described below to identify patterns of tradecraft within a graph, as a fraud analyst team at a bank would do.

Overall, we will show how to use open datasets with contemporary entity resolution to enhance AI applications for more trusted outcomes, streamlined governance, better customer experiences, and accelerated innovation.

A tutorial in three parts:

1. Part 1: Visualize fraud networks, using Senzing, OpenSanctions, Open Ownership, Polars, Maplib, NetworkX, yWorks
2. Part 2: Entity embeddings in graphs: blend structured + unstructured data, using Senzing, LanceDB, RDFlib, NetworkX, DSPy, Ollama
3. Part 3: How to become a money launderer

Course Goals

Gain hands-on experience in Python using high-quality knowledge graphs with entity resolution, computable semantics, entity embeddings, graph algorithms, interactive visualization, plus context engineering to augment the graph and enhance downstream AI applications.

Target Audience

• Data Scientists, Machine Learning Engineers
• Data Engineers, MLOps
• Financial Fraud Analysts
• Team leads and managers for the roles above

Prerequisites

• Level: Beginner - Intermediate
• Some experience coding in Python
• Familiarity with popular packages such as Jupyter and Docker

Important: You must have Docker, Git, and Python 3.13+ downloaded and installed to run this tutorial.

Before going to the conference, download this container onto your laptop:

docker pull senzing/serve-grpc:latest

NB: due to sudden changes for KùzuDB, we've had to adjust the content of this course, and unfortunately Prashanth Rao will not be able to join.

We will include additional content using LanceDB and DSPy for context engineering, plus a section on "how to become a money launderer" -- leveraging graph analytics to examine OCCRP leaked data from the "Azerbaijani Laundromat" case.

---

Part 1: Visualize Fraud Networks

• Duration: 55 minutes
• GitHub repository: https://github.com/DerwenAI/cdl-2025-masterclass

We'll start with a brief intro lecture covering the background for leveraging these technologies together with open data in an anti-fraud use cases.

1. Download the Docker container for Senzing gRPC server (during the intro)
2. Launch the Senzing container and run it in the background for a gRPC server.
3. Initialize the Python environment on your laptop using uv to load the library dependencies.
4. Download slices of the OpenSanctions and Open Ownership datasets.
5. Run entity resolution in Senzing to merge the datasets and generate graph "building blocks" in RDF, as a domain-specific thesaurus.
6. Review the metadata application profile (MAP) for the SKOS taxonomy used, how it integrates with NIEM, FollowTheMoney, BODS, and so on.
7. Also transform records from the datasets into Polars dataframes.
8. Use SPARQL queries in Maplib to transform the RDF into Polars dataframes.
9. Load tables into a graph database from the dataframes.
10. Leverage graph algorithms in NetworkX: partitioning to identify subgraphs as potential fraud rings within the graph, and betweenness centrality to rank individuals of interest within each subgraph.
11. Run visualizations using yFiles to examine the 2021 South London Papa Johns tax evasion case.
12. Q&A discussion.

Tutorial: PART1 instructions

Links to components

• https://github.com/senzing-garage/sz-semantics
• https://github.com/senzing-garage/serve-grpc
• https://www.opensanctions.org/
• https://www.openownership.org/
• https://github.com/DataTreehouse/maplib
• https://pola.rs/
• https://networkx.org/
• https://jupyter.org/
• https://www.yworks.com/products/yfiles-graphs-for-jupyter
• https://github.com/yWorks/yfiles-jupyter-graphs-for-kuzu

---

Part 2: Entity Embeddings in Graphs

• Duration: 30 minutes
• GitHub repository: https://github.com/DerwenAI/strwythura

Let's build on what we covered in Part 1, and examine how to use entity embeddings in graphs, to blend structured and unstructured data, using Senzing, LanceDB, RDFlib, NetworkX, DSPy, Ollama -- applying a more intentional approach to developing "context" for context engineering.

This uses entity resolution (ER) with computable semantics, starting from a taxonomy curated per use case, then generating a domain-specific thesaurus from structured data sources. In other words, resolve identify information about people and organizations along with relations among the entities, their synonyms, and so on, and use this as a "backbone" for working with the graph.

During KG construction and updates, the domain-specific context focuses the named entity recognition (NER) and entity linking to extract entities from unstructured data sources and link these contextually into the KG.

Given the thesaurus, we'll construct the following bundle of assets from the unstructured sources:

• vector store with embeddings of text chunks, cross-linked with entities
• entity embedding model, built from sequences of lemmatized noun phrases
• lexical graph constructed from parsed text chunks using a textgraph algorithm

This provides better afforandances for human-in-the-loop (HITL) curation of extracted entities (vs. resolved entities) before their promotion into the KG, and also guides how to tranform RDF representation (which allows for validation, relation range constraints, etc.) into a property graph.

Working from these assets as a basis, we can promote graph elements (entities, relations, properties) into the resulting KG, retaining the cross-links with the vector store.

Downstream from the KG, during run-time use, the entity embeddings enhance LLM integrations in multiple ways to enhance outcomes in AI applications.

• query the entity embedding model, as a more robust alternative to Text2Cypher (or complementary)
• use semantic expansion and random walks so than entity reranking drives text chunk reranking
• include text definitions for classes in the taxonomy into the reranked text chunks

We will this section begin with a brief introduction to the topics plus overview of available (albeit less often discussed) technology components.

Then we'll have a live demo and code walk-through in the implementation:

1. Run entity resolution in Senzing to merge the datasets and generate graph "building blocks" in RDF, as a domain-specific thesaurus.
2. Construct a "backbone" for a knowledge graph in NetworkX from the thesaurus.
3. Crawl the related documents, chunking text from the unstructured content and loading chunks plus their embeddings into LanceDB.
4. Parse the text in each chunk using GLiNER zero-shot NER in a spaCy pipeline, based on semantic definitions from the thesaurus.
5. Use a textgraph algorithm to construct a lexical graph which links to the text chunks, plus ranking for the "most referenced" entities.
6. Curate semantics for optimizing the AI app outcomes within a specific domain.
7. Entity linking: promote entities extracted from the unstructured content into the KG, linked to ER results.
8. Build entity embeddings in GenSim, determining their neighborhoods semantic expansion.
9. Use NetworkX for semantic expansion and semantic random walks to rerank text chunks in LanceDB.
10. Implement a question/answer chat bot based on GraphRAG using DSPy and Ollama, running the gemma3:12b LLM locally.
11. Q&A discussion.

Tutorial: PART2 instructions

Links to components

• https://senzing.com/what-is-entity-resolution/
• https://lancedb.com/
• https://spacy.io/
• https://github.com/urchade/GLiNER
• https://networkx.org/
• https://radimrehurek.com/gensim/models/word2vec.html
• https://dspy.ai/
• https://ollama.com/
• https://huggingface.co/google/gemma-3-12b-it
• https://pyinstrument.readthedocs.io/

---

Part 3: How to become a money launderer

• Duration: 20 minutes
• GitHub repository: https://github.com/DerwenAI/cdl-2025-masterclass/

Let's build on what we covered in Part 1 and Part 2 to develop more sophisticated graph-based analysis for fraud tradecraft. Recall from the general-case data model used for anti-fraud that the event data is difficult to obtain, and we did not use it in Part 1.

So now we will introduce event data from whistleblower leaks, to analyze a real-world case. Then let's examine how to use leaked data to build simulations which generate synthetic data. This allows for better evaluation of anti-fraud approaches, both in terms of scale and diversity of patterns.

We'll start with a brief intro about the "Azerbaijani Laundromat" incident (~$3B money laundering) with real-world examples of how graph technologies empower anti-fraud investigations.

1. Review anti-money laundering points in summarized excerpts from the highly recommended The Dark Money Files by Graham Barrow and Ray Blake.
2. Demo of the notebook in the kleptosyn repo for forensic accounting, graph-based flow analysis, and interactive visualization of the Azerbaijani Laundromat leaked data from OCCRP.
3. Demo of the AML notebook in this repo to show how synthetic data for wire transfer transactions can simulate patterns of criminal tradecraft among fraud networks represented in a graph.
4. Review resources for how data practitioners can get involved: GraphGeeks community, learn more about AML, human trafficking certification, support whistleblowers, etc.
5. Q&A discussion.

Tutorial: PART3 instructions

Links to components

• https://www.thedarkmoneyfiles.com/
• https://www.occrp.org/en/project/the-azerbaijani-laundromat/the-raw-data
• https://github.com/DerwenAI/kleptosyn
• https://networkx.org/
• https://pyvis.readthedocs.io/
• https://seaborn.pydata.org/
• https://matplotlib.org/
• https://youtu.be/Gtp7U0iq-2I?feature=shared
• https://jupyter.org/

---

Kudos to @prrao87, @brianmacy, @jbutcher21, @docktermj, @cj2001, @louisguitton, @jesstalisman-ia, @pudo, @StephenAbbott, and the kind folks at GraphGeeks for their support.