Graph embeddings without the training: Cleora averages its way past DeepWalk
Cleora embeds graphs and hypergraphs on a laptop CPU in seconds by replacing random walks with one deterministic averaging step, and it beats nine established baselines doing it.

What it does
Cleora turns a graph or hypergraph into one vector per node by repeating a single operation: every vector becomes the normalised average of its neighbours’ vectors. That is, in effect, one step of every random walk at once, without sampling any of them. No negative sampling, no learning rate, no epochs, no GPU — the only parameter that really matters is the number of iterations, and the default picks it from the graph’s density.
The interesting bit
The boring part is the value: because there is no stochastic training, the same input gives bit-for-bit identical embeddings regardless of thread count. No seeds to average over, no runs to repeat. It also handles hypergraphs natively (shopping baskets, sessions), and embed_inductive adds new nodes without touching existing vectors, so downstream models stay valid.
Key highlights
- ogbn-arxiv (169,343 nodes, 1.2M edges) embedded at 1024 dimensions in 1.4 s on a laptop CPU; baselines ranged from 10 s to 18 minutes.
- Best of 10 methods on 23 of 26 graphs at finding a node’s held-out neighbours; ahead in 457 of 488 head-to-head comparisons.
- Memory footprint is two embedding matrices and nothing else — 6 GB for Yelp where baselines needed 15–62 GB.
- Supports weighted edges, bipartite graphs, and propagating your own text/image features along the graph instead of random initialisation.
- The benchmark is one command (
benchmarks/replicate.sh) and every figure on the page is generated from the result files.
Caveats
- The old per-iteration whitening mode (the 3.2 default) needs many more entities than dimensions and is 30–350× slower; on small graphs it is clearly worse than the new default.
- The popularity prior for link prediction helps on two of three benchmark graphs — not a universal win.
Verdict
If you need entity embeddings for similarity search, classification features, or recommendations over relational data — especially on CPU-only hardware or very large graphs — this is worth a hard look. If you want a learned, task-tuned neural encoder, this is not that, and does not pretend to be.
Frequently asked
- What is BaseModelAI/cleora?
- Cleora embeds graphs and hypergraphs on a laptop CPU in seconds by replacing random walks with one deterministic averaging step, and it beats nine established baselines doing it.
- Is cleora open source?
- Yes — BaseModelAI/cleora is an open-source project tracked on heatdrop.
- What language is cleora written in?
- BaseModelAI/cleora is primarily written in Python.
- How popular is cleora?
- BaseModelAI/cleora has 547 stars on GitHub.
- Where can I find cleora?
- BaseModelAI/cleora is on GitHub at https://github.com/BaseModelAI/cleora.