Last updated 07-28-2026
Category:
Reviews:
Join thousands of AI enthusiasts in the World of AI!
Cleora.AI
Cleora turns edge lists into dense vector embeddings for every node in a graph, without GPUs or random walk sampling. The pycleora Python library ships a Rust core that computes full walk distributions through sparse Markov matrix powers, then exposes classification, similarity search, and link prediction helpers. It targets recommendation systems, fraud graphs, knowledge triples, and other jobs where relational structure matters more than text features.
Walk-based libraries like DeepWalk and Node2Vec approximate neighborhoods by sampling paths and training skip-gram models. Cleora skips both steps and aggregates every k-hop neighborhood in one deterministic pass, which is why Zomato reported embedding a customer-restaurant graph in under five minutes after GraphSAGE took about twenty hours on the same setup. The trade-off is scope: it is a graph-structure toolkit, not a general feature store for unstructured data.
Data science teams install pycleora with pip and only need NumPy and SciPy alongside the roughly 5 MB package. The MIT-licensed library includes eight baseline algorithms, heterogeneous graph support, MLP and label-propagation classifiers, and a CLI for batch jobs. Published benchmarks on SNAP, Planetoid, and DGL datasets report top accuracy on graphs up to about two million nodes while competitor methods time out or run out of memory.
pip install pycleora pulls a ~5 MB package that needs only NumPy and SciPy, no GPU drivers or CUDA stack
Sparse Markov matrix powers compute all walk distributions exactly, with no random walks or negative sampling
Eight embedding algorithms in one API: Cleora, DeepWalk, Node2Vec, ProNE, RandNE, HOPE, NetMF, and GraRep
Zomato cut customer-restaurant embedding time from about 20 hours with GraphSAGE to under five minutes on the same data
Benchmarks on roadNet-CA embed 1,965,206 nodes in 31.5 seconds using about 4.1 GB RAM on one CPU core
Heterogeneous hypergraphs accept TSV edge files with typed columns like complex::reflexive::product
Open-source MIT license with no API keys or usage caps
Deterministic output: same graph and settings always produce identical embeddings
Embeds million-node graphs on CPU when walk-based libraries hit memory limits
Built-in classifiers, metrics, and CLI avoid spinning up PyTorch or TensorFlow for evaluation
Python library focused on graph structure, not a managed cloud embedding API
Largest published scale test used about 4.1 GB RAM for a 1.96 million-node road network
Heterogeneous TSV column syntax takes documentation reading before complex graphs import cleanly
Is Cleora free to use?
Yes. Cleora is open-source software under the MIT License, so there are no license fees, API keys, or usage limits. Install it with pip install pycleora and run it on your own CPU hardware.
Does Cleora need a GPU?
No. Cleora runs on CPU only and does not require CUDA, cuDNN, or any GPU drivers. Its Rust core uses sparse matrix operations with adaptive parallelism across CPU cores.
How do you install Cleora?
Run pip install pycleora from PyPI. Cleora depends on NumPy and SciPy only, and the install footprint is about 5 MB. A CLI supports embed, info, benchmark, and similar commands without writing Python.
What can you build with Cleora embeddings?
Cleora embeddings support recommendation systems, fraud detection graphs, community detection, link prediction, knowledge graph completion, and customer lookalike models. The use-cases docs include working Python examples for each pattern.
How does Cleora compare to DeepWalk and Node2Vec?
DeepWalk and Node2Vec sample random walks and train skip-gram models with negative sampling. Cleora computes exact walk distributions through sparse Markov matrix powers, producing deterministic embeddings without sampling noise.
What graph formats does Cleora accept?
Cleora builds graphs from TSV or space-separated edge lists with typed column strings such as complex::reflexive::product. It also exports to NetworkX, PyG, and DGL, and can save embeddings as NPZ files.