See what connects.
Inspect paths, time and location in Link Charts. Use the DataWalk ML graph extension to calculate centrality, detect communities and derive network features.
Platform / Connected Analytics
Graph, relational and search analysis in one engine.
Start with curated, connected data. Test hypotheses with visual queries, graph algorithms and machine learning. Bring new sources into the analysis as your questions evolve.
Enterprise knowledge graph
Do these suppliers have more in common than their payment history?
The challenge
Too much setup between questions.
01
Rebuilding the dataset
Each project repeats the work of finding sources, reconciling identities and working out how the records connect.
02
Rewriting the query
Following another relationship often means another join, another run and another round of inspection. Exploration becomes query maintenance.
03
Hitting compute limits
Network features take memory and computation. As data grows, useful experiments can become infrastructure projects.
04
Leaving predictions unused
A machine learning model produces predictions. Getting them into the records and tools analysts use is another job.
Build on the preparation already done
Add a source.
Ask a new question.
Work with the entities, relationships and cleaning rules already defined in DataWalk. Connect another database, file or API to that shared model as your analysis develops.
Combine transactions, business records and events. Entity and relation extraction can bring information from documents into the analysis too.
Analytical tools
Follow the hypothesis.
Choose the method.
Analysts and data scientists can explore relationships, calculate across records and test hypotheses on the same connected data. Work visually or extend the analysis with code.
Filter, compare, calculate.
Explore distributions and aggregate across connected records. Build queries visually in Universe Viewer, change a condition and save the analysis for reuse.
Find your starting point.
Find records across configured fields and datasets. Use exact or configured fuzzy matching for name variations, then inspect the results and their connections.
Test a new feature.
Evaluate models with tabular and network features in JupyterLab. Supported algorithms execute in the database sandbox. Import predictions onto entities for analysts to query.
Make the analysis reusable.
Combine saved analyses with assigned weights to score entities. Configure an alert when entities enter or leave the results of a saved analysis.
From question to analysis
“Are payments concentrated among connected suppliers?”
A payment total tells you how much each supplier received. Directorship data lets you examine whether those suppliers are connected.
-
01
Add the missing context.
Connect company-registry records to the suppliers and payments already in DataWalk.
-
02
Test the hypothesis.
Select a period. Find suppliers with overlapping directorships and calculate their share of payments.
-
03
Examine the result.
Inspect the records behind the calculation. Change the period and compare. Save the analysis for reuse.
For machine learning, calculated properties such as shared-director counts can become candidate features.
Scale changes what you can attempt
Explore large networks.
Explore large networks.
Keep testing ideas.
DataWalk’s reported traversal benchmark tested graphs up to 20 TB. Bounded queries recorded median times below one second per hop, with essentially flat latency from hop 2 through hop 15.
20 TB graph · bounded traversal
0.24–0.60seconds per hop
Median range across steady-state hops.
Top 100 counterparties retained per hop. Six-node clusters, warm storage, one user. Broader traversals require more work.
In practice / ING Bank
From connected data to working customer profiles.
ING combined transaction and relationship features to build customer profiles, then automated monitoring as new data arrived. Scientists and analysts could work with the same behavioural descriptions.
Read the customer story
FAQ
Before you start.
How much preparation is already done?
Once sources are configured, analysts and data scientists reuse the shared data preparation and relationships. Adding a source requires mapping and refresh configuration. Each modelling task still needs its own feature selection and validation.
Do I need to code?
Search, visual queries and Link Charts support analysis without writing code. Python, JupyterLab and APIs let data scientists extend the analysis with their own methods.
Can I extract entities and relationships from text?
Yes. DataWalk supports GLiNER-based entity extraction and configured relation-extraction workflows. Link extracted information to existing entities so document content can contribute to search and analysis.
Does all machine learning run in the database?
Supported DataWalk ML algorithms run in the database sandbox. Custom Python libraries may need a separate execution environment. Imported predictions need a configured process to rerun the model and update the results.
Can we keep our warehouse and existing tools?
Yes. DataWalk ingests the data needed for connected analysis. Warehouses and operational systems can remain in place, with APIs and exports connecting the results to other tools.
Architecture & DeploymentWhat would you test next?
Bring a question and the sources you need to answer it.
