Cleora.ai

Cleora is a graph embedding engine designed around a single mathematical insight: instead of sampling random walks and training a skip-gram model, you can capture the full distribution of walks by repeatedly multiplying an embedding matrix by a sparse Markov transition matrix. Each iteration replaces every node's embedding with the L2-normalized average of its neighbors, producing embeddings that are stable, reproducible, and remarkably fast on commodity hardware.

Practically, Cleora ships as a compact Rust extension installed with pip, with only numpy and scipy as dependencies. It natively supports heterogeneous hypergraphs through typed TSV columns, so multi-relational graphs, bipartite graphs, and knowledge triples can be fed in directly without preprocessing. Typical benchmarks from the project and its users report large speedups versus deep-learning baselines — Zomato, for example, reports embedding generation dropping from roughly 20 hours with GraphSAGE to under five minutes.

The project also bundles a wide evaluation and comparison suite, including alternative algorithms (ProNE, RandNE, HOPE, NetMF, GraRep, DeepWalk, Node2Vec) and metrics such as AUC, MRR, Hits@K, MAP@K and nDCG, so teams can benchmark embedding quality without adding heavy dependencies. Because embeddings are inductive and stable, new nodes can be embedded without retraining the whole graph — useful in production recommendation, fraud detection, and entity resolution pipelines.