Cleora

Cleora is an open-source graph embedding library with a Rust core and Python API, installed via pip install pycleora. It builds a sparse Markov transition matrix from input graphs, including heterogeneous hypergraphs with typed edges read from TSV input, and iteratively propagates L2-normalized neighbor averages to produce dense, deterministic embeddings. The page states it requires only numpy and scipy, runs on CPU without GPUs, and includes built-in evaluation tools (MLP classifier, label propagation, metrics such as AUC and nDCG), graph sampling, hyperparameter tuning, a benchmarking suite, and a CLI. The website reports benchmark results against seven alternative algorithms and cites use by Zomato, Synerise, and Dailymotion, along with competition placements in KDD Cup 2021, WSDM WebTour 2021, and SIGIR eCom 2020. The page describes the library as free and MIT licensed.

Key features

• Sparse Markov transition matrix construction from input graphs
• Heterogeneous hypergraph and multi-relational edge support via TSV typed columns
• Deterministic embeddings: same input yields same output
• Iterative L2-normalized neighbor propagation (3-4 iterations for co-occurrence, 7+ for contextual similarity)
• CPU-only operation with no GPU required
• Rust core with PyO3 Python bindings and adaptive parallelism
• Minimal dependencies: numpy and scipy only
• Inductive embedding of new nodes without full retraining
• Seven alternative algorithms included for comparison: ProNE, RandNE, HOPE, NetMF, GraRep, DeepWalk, Node2Vec
• Built-in MLP classifier and Label Propagation
• Evaluation metrics: AUC, MRR, Hits@K, MAP@K, nDCG, ARI, Silhouette Score, k-fold cross-validation
• Graph sampling: neighborhood, subgraph, GraphSAINT mini-batching, negative sampling, train/test edge splits
• Grid search and random search hyperparameter tuning
• Benchmarking suite with time, memory, and accuracy metrics and 5 built-in graph generators
• Command-line interface (pycleora embed)
• Helper function find_most_similar for similarity lookup

Best suited for

• Recommendation systems
• Knowledge graphs
• Customer lookalikes and segmentation
• Entity resolution
• Fraud detection
• Social network analysis
• Drug discovery
• Supply chain analysis
• Teams wanting CPU-only, reproducible graph embeddings

Pricing

Open source and free to use under the MIT License; the page states no license fees, API keys, or usage limits. Cloud infrastructure costs are the user's own.