GraphNews

Graphina is a graph data science library for Rust. It provides common data structures and graph algorithms for analyzing real-world networks, such as social, transportation, and biological networks. It also helps power Onager – graph analytics on DuckDB – which we featured last week.

Compared to other Rust graph libraries, Graphina aims to provide a more high-level API and a wide range of ready-to-use algorithms for network analysis and graph mining tasks. Graphina aims to be as feature-rich as NetworkX but with the speed and performance benefits of Rust. And if you’re a Pythonista, the PyGraphina Python library allows users to use Graphina in Python.

🚀 Rethinking Portfolio Diversification with Graph Data Science

Traditional correlation matrices only tell us where markets have been—not where they're going. In today’s hyper-connected financial landscape, that’s not enough.

In the latest work by Nuno Pedro L., we use Neo4j Graph Data Science to model equities as a dynamic network and apply Link Prediction to anticipate future relationships between assets. Instead of reacting to correlations after they form, we can predict them—uncovering hidden risks, emerging clusters, and new opportunities for statistical arbitrage before they appear in traditional models.

🔍 Why it matters:

• Captures non-linear, evolving market structures
• Reveals early signals of contagion or co-movement
• Supports smarter diversification and proactive risk management

If you’re exploring the future of quantitative finance, network analytics, or portfolio intelligence, this approach is a game-changer.

📈 Graph data science isn’t just descriptive—it’s predictive.

One of my biggest contributions to the GraphFrames project is scalable graph embeddings. While not perfect, my implementation is inexpensive to compute and horizontally scalable. It uses a combination of random walks and Hash2Vec, an algorithm based on random projection theory.

In the post, I provide the full code and an explanation of all the engineering decisions I made. For example, I explain why I used Reservoir Sampling for neighbor aggregation or Map Partitions instead of the DataFrame API.

The pull request (PR) has not been merged yet, so if you have any ideas on how to improve the approach, I would love to hear them! Overall, it appears to be a good, inexpensive way to create scalable embeddings of graph vertices that can easily be incorporated into existing classification or recommender system pipelines. Finally, GraphFrames will have real capabilities for graph data science! At least, I hope so. :)