Cleora: CPU-Only Graph Embedding Library Achieves 240x Speedup Over GraphSAGE
Key Takeaways
- ▸Cleora eliminates random walk sampling and GPU requirements by computing exact walk distributions through sparse matrix powers, achieving 240x faster embedding generation than GraphSAGE
- ▸The library weighs only 5 MB with minimal dependencies (NumPy, SciPy only), making it dramatically lighter than competing frameworks like PyTorch Geometric and DGL
- ▸Zomato deployed Cleora in production to power 'People Like You' recommendations for 80M+ users across 500+ cities, reducing processing time from 20 hours to under 5 minutes
Summary
Cleora, a new CPU-only graph embedding library, has demonstrated dramatic performance improvements over existing graph neural network frameworks by eliminating random walk sampling and GPU requirements. The library computes exact distributions of all possible walks through a single sparse matrix power operation, achieving deterministic embeddings without the noise and approximation errors inherent in competitors like GraphSAGE, Node2Vec, and DeepWalk. With a minimal 5 MB footprint and no heavy dependencies, Cleora relies solely on NumPy and SciPy, making it significantly more lightweight than alternatives like PyTorch Geometric (500+ MB) and DGL (400+ MB).
Zomato, one of the world's largest food delivery and restaurant discovery platforms, has successfully deployed Cleora in production to power recommendations for over 80 million users across 500+ cities. The company reported reducing embedding generation time from 20 hours using GraphSAGE to under 5 minutes with Cleora—a 240x speedup—while maintaining higher accuracy on real-world graphs. Cleora natively handles heterogeneous hypergraphs with multi-type nodes and edges, supports inductive learning for new nodes without full retraining, and guarantees reproducible results through its deterministic algorithm design. This production validation demonstrates that the library is ready for enterprise-scale deployment in recommendation systems, search ranking, and other graph-based ML pipelines.
- Deterministic by design, Cleora guarantees reproducible embeddings and supports inductive learning for new nodes, making it production-ready without requiring retraining on entire graphs
Editorial Opinion
Cleora represents a significant departure from the current paradigm in graph machine learning by proving that sophisticated embeddings don't require sampling heuristics, stochastic approximations, or GPU acceleration. The mathematical elegance of computing exact walk distributions through matrix powers is compelling, and Zomato's 240x speedup validates the approach at scale. However, the broader ML community's adoption will depend on whether Cleora's deterministic, structure-only approach generalizes beyond recommendation systems to more complex downstream tasks where the inductive biases of learned neural networks may prove valuable.
