Cleora

Cleora is a Rust-powered graph embedding engine that produces exact, deterministic embeddings without random walks, negative sampling, or GPUs. Where other libraries approximate by sampling paths and training skip-gram models, Cleora computes the full distribution of every possible walk using a single sparse matrix power on one CPU core. The result: production-grade embeddings that are orders of magnitude faster to generate and perfectly reproducible. The library is tiny — about 5 MB with only numpy and scipy as dependencies — and installs with `pip install pycleora`.

Cleora works by constructing a sparse Markov transition matrix from your input graph (handling heterogeneous hypergraphs with multi-type nodes and edges natively). Each iteration multiplies the embedding matrix by that transition matrix, capturing the exact distribution of all walks of length k. L2-normalized propagation averages neighbor embeddings each iteration — 3-4 passes for co-occurrence similarity, 7+ for contextual similarity. No noise, no stochastic variation. Zomato replaced their GraphSAGE pipeline with Cleora and cut embedding generation time from 20 hours to under 5 minutes, scaling recommendations across 500+ cities. The library also includes 7 alternative algorithms (DeepWalk, Node2Vec, etc.) as baselines, along with evaluation metrics, graph sampling, and hyperparameter tuning.

Data scientists and ML engineers working on recommendation systems, knowledge graphs, entity resolution, or fraud detection will find Cleora a practical alternative to heavyweight deep learning frameworks. It runs on any machine with a CPU — no GPU needed — and produces embeddings stable enough for regulated environments and reproducible research. The tool is used in production by Synerise (e-commerce personalization) and Dailymotion (video recommendations), and has won top placements in KDD Cup and WSDM competitions. If your pipeline currently relies on random walks and skip-gram training, Cleora offers a ...