Graphs can preserve the context and relationships in unstructured data, which makes more of an organization's information available for analysis.
2
Graph analysis works best when a large, detailed graph is narrowed into domain-specific subgraphs, abstracted into useful concepts, and passed through an investigation workflow.
3
Visual analytics helps people discover questions by mutating and exploring graphs, while graph algorithms and AI can automate parts of that process.
Summary
Paco Nathan and Weidong Yang discuss how graphs can extend analytics beyond structured tables. They start with sensitive financial investigations, where tokenizing personally identifiable information can let external AI process data without exposing the original identities. The conversation then turns to GraphRAG and the value of preserving context before the downstream question is known. Weidong argues that graphs can keep information flexible and nuanced, while domain-specific definitions can make selected parts precise. Paco describes a graph-thinking workflow: build a large graph, partition it into interesting subgraphs, apply graph algorithms to prioritize what matters, and send the result into a human case-management process. Both guests stress that graphs must be simplified and transformed for analysis, since a raw enterprise graph is too large for people to understand. They also discuss federation across domains, data quality, access control, and the role of visualization in discovering questions rather than merely displaying answers.
Sensitive investigations need privacy controls before AI processing
Paco Nathan explains why suspicious activity reports create a difficult balance for banks. A report can be required when activity looks criminal, yet sending too much information can create legal exposure. Processing one report can cost about $50,000, so a model that generates thousands of reports per day could create tens of millions of dollars in liability. For data sent to an external AI service, Paco describes replacing personally identifiable information with generated tokens, running the analysis, and restoring the confidential values inside the bank. Weidong Yang adds that even a document's structure can reveal sensitive facts, so anonymization has to consider more than names and identification numbers.
Local AI hardware matters when data cannot leave the organization
Paco Nathan says about 60% of his use cases are air-gapped, including public-sector environments that run in secure facilities. That limits the use of external AI services. He points to Useful Sensors and work by Pete Warden and colleagues, who demonstrated small, low-power hardware with a neural-network accelerator that could run multiple language models on battery power. His point is that inference does not always require a large GPU cluster. Local hardware can support low-power and low-latency use cases where sending data outside the environment is prohibited.
Graphs preserve context that may matter before anyone knows the question
Weidong Yang argues that language models work with contextual and fuzzy information in a way that differs from older systems based on exact symbolic matching and rigid data structures. He prefers data structures that preserve nuance because subtle details can become important later. His example is a Brazilian party invitation marked for 6:00 p.m., which may mean that the host starts preparing at that time and guests arrive several hours later. Paco connects this idea to the recursive references found in stories, speech, and books. Both argue that early data processing should avoid discarding context before the eventual use is known.
Ontology is useful inside a domain when its boundaries are explicit
Weidong Yang compares ontology with a scientific theory. It can describe important truths within a defined context, but it should always be tested at its boundaries. He says traditional knowledge graphs often cause problems when an ontology is treated as if it applies everywhere. Paco Nathan agrees that enterprise-wide ontologies rarely work across domains. A term such as "customer" can mean different things to sales and procurement. In areas such as news analysis, it can be better to collect relationships and evidence first, then apply shared definitions where the domain becomes clear.
Graph investigation reduces a huge network into something people can inspect
Paco Nathan describes a repeated graph-thinking workflow for financial crime investigations. First, build a graph that may contain hundreds of millions or billions of nodes. Next, partition it into smaller subgraphs with patterns that deserve attention, such as a promising customer or a money-mule scheme. Then apply graph algorithms such as Louvain community detection or betweenness centrality to find connected or influential parts. Finally, place the result into a work process where an analyst explores the graph, possibly with help from agents that summarize or retrieve information. Weidong adds that the graph usually has to be projected into a simpler, domain-specific form before centrality calculations are meaningful.
Graph data can expose structural errors that tables hide
Weidong Yang contrasts graph capture with a wide tabular flight dataset that allocates a fixed group of columns for diversions. A graph can represent any number of diversions without deciding in advance how many columns are needed. He also describes combining datasets from different departments. Teams may know that other teams' records contain errors, but they cannot easily reconcile them in tables. Once the data is connected in a graph, duplicate records and mismatched identifiers become visible through their dependencies. Paco gives the example of a supposedly unique Social Security number that appears thousands of times because one application later replaced real values with the same dummy number.
Visualization helps people discover questions before they can state them
Weidong Yang says that knowing which question to ask can account for most of the work in an exploratory investigation. Visual perception can reveal patterns before a person has words for them. A graph gives those associations a form that people can inspect, compare, and revisit. Paco Nathan describes complex investigations as a process of poking at unknowns that cannot be fully charted in advance. He compares this with spreadsheets, which show rows and columns while hiding the dependency graph that makes their calculations work. Visualization therefore supports exploration, rather than simply presenting the result of a predefined query.
A graph must be transformed for human analysis instead of shown raw
Weidong Yang separates a graph used to capture information from a graph used to support human thinking. The capture graph should preserve as much truth and detail as possible, but that raw structure can overwhelm a person. Analysis needs operations that simplify, aggregate, and abstract the graph within a particular domain. A detailed set of relationships between Paco Nathan and Weidong Yang could be reduced to the concise statement that they know each other, while the underlying evidence remains available. Weidong describes visual operators that let a domain expert mutate the graph without writing Python or graph queries. Those visual transformations can also turn an observed pattern into a feature for later analysis.
"The more complex the problem becomes, the more that those unknowns are not something that can really be charted. They have to be sort of poked at and explored."Paco Nathan29:48
Who should watch
You are building analytics over emails, reports, investigations, or other unstructured sources and need to preserve context before the use case is fully known.
Your organization has large graphs that are difficult to inspect, or teams disagree about definitions and data quality across business domains.
You are evaluating GraphRAG and want a workflow that includes privacy controls, graph algorithms, visual exploration, and human investigation.
