Modeling Connected Data with Graphs and Hypervectors

Companion notebooks for the Connected Data London (CDL) 2026 two-hour masterclass, led by Prashanth Rao and David Hughes from HDC Labs.

How do you find a useful precedent when you only know part of the story? And once a case is resolved, how can that one example help with the next one?

In this workshop, we'll investigate a delayed delivery for a London homeware retailer. A customer in Ireland is waiting for a cordless lamp, order SO-1042, shipped from the East London depot. Support knows the product, depot and country, but not what went wrong. We'll follow the connected records, retrieve similar past orders, and update a memory that helps decide who to call first: the depot, customs broker or carrier.

The workshop is for graph practitioners who are comfortable with basic Python and Cypher. No prior knowledge of hyperdimensional computing (HDC) is needed. We introduce hypervectors and their operators through small, runnable examples, with figures from the talk beside the relevant code.

A graph models connections explicitly. Hypervectors model them implicitly. We'll use the graph to inspect exact evidence, and hypervectors to compare patterns and learn from resolved cases.

What you'll work through

Run the notebooks in order. Each rebuilds the data it needs, so you can restart a notebook without depending on another notebook's kernel or database.

Notebook	What you'll do
1 · A sales-order graph	Store connected records in LanceDB, query them with Cypher, and find an exact precedent.
2 · The three HDC operators	Build hypervectors and experiment with binding ⊗, bundling ⊕ and permutation ρ.
3 · Motif search	Compose patterns from stored edge hypervectors at query time, rank partial matches, compare with Cypher scoring, and inspect the graph evidence.
4 · Learning from one example	Summarise cases into issue prototypes and add one newly resolved order to an existing prototype, then inspect the predictions before and after.

This is a small, constructed teaching dataset. “One example” means one additional confirmed case, not learning an entire class without prior examples. The notebooks include the assumptions and limits alongside the results.

Set up before the workshop

You'll need Git, uv, and internet access for the initial installation. uv manages the project's Python 3.12 environment and installs Python if needed. No GPU, API keys or separately running database server are required.

The pinned lance-csr release provides wheels for these platforms:

Platform	Requirement
Apple Silicon Mac	macOS 11 or later
Linux x86_64	glibc 2.28 or later, such as Ubuntu 20.04+

Intel Macs, native Windows and Linux ARM do not have wheels for this pinned release. Use a supported machine for the workshop. For other environments, consult the lance-csr source-build instructions before the session.

Clone the repository and install the locked dependencies:

git clone https://github.com/hyperdimensionalcomputingai/cdl-2026.git
cd cdl-2026
uv sync --locked --no-build-package lance-csr

The wheel-only option for lance-csr prevents an unexpected Rust source build. Linux uses CPU-only PyTorch wheels to avoid downloading CUDA dependencies. Allow the install to finish before arriving; the exercises use data included in the repository.

Open the notebooks

From the repository root, start JupyterLab:

uv run jupyter lab src/

Open the local URL printed in the terminal if your browser doesn't open automatically. Start with 01_sales_order_graph.ipynb and select the Python 3 kernel. Run cells from top to bottom with Shift+Enter, pausing to inspect each result. The workshop alternates between explanations in the presentation and hands-on work in these notebooks.

If you get a missing-package error, stop JupyterLab and relaunch it with the command above so it uses this project's environment. If you run cells out of order, restart the kernel and run from the top. Keep the notebooks in src/, beside the shared data module.

The notebooks create local *.lancedb directories, which Git ignores. Setup cells rebuild the teaching tables when rerun; use these notebooks with the supplied workshop data.

Check your environment

For a complete check before the session, execute all four notebooks without opening a browser:

uv run pytest

The tests run in temporary directories and leave the saved notebook outputs unchanged.

What's underneath

LanceDB stores the tables and hypervectors, lance-csr runs Cypher over the graph, and torchhd supplies the HDC operators. The shared dataset and helpers are in src/sales_orders.py; workshop figures are in img/.

See the CDL talk page for the session details. The scenario builds on Your data is connected.