Neo4J Graph Data Science Library (GDS)

Key Takeaways

Section 5.2 graph data science library. Neo4j's GDS plugin brings 65 plus production-grade graph algorithms directly into your cipher workflow. The Neo4j graph data science library GDS is what makes Neo4j competitive with dedicated graph analytics platforms. It gives you algorithms that would take months to implement correctly. packaged as simple cipher call statements. The library is organized into four main algorithm families. Centrality algorithms measure the relative importance of nodes. Page rank is the most famous. It measures global influence based on the structure of incoming connections. Betweenness centrality identifies bridge nodes. Closeness and degree centrality give you proximity and connectivity measures. Community detection algorithms find clusters of densely connected nodes. Luvet is the most accurate. It optimizes modularity. Label propagation is faster and more scalable. Weekly connected components tells you which parts of the graph are reachable from which other parts. Similarity algorithms find nodes that look alike based on their connections or embeddings. Node similarity uses the jacquard coefficient. Shared neighbors divided by total neighbors. KN&N uses embedding vectors for cosine similarity. Pathf finding algorithms find optimal routes through the graph. Dystra and AAR for weighted shortest paths. Random walk for sampling. Minimum spanning tree for network coverage problems. The GDS workflow always has the same shape. Project an in-memory subgraph from Neo4j, run your algorithm against it, and get results back in one of three modes: stream, mutate, or write. We'll cover projections and modes in detail in section 5.4. Let's look at the two most important centrality algorithms, page rank and betweenness centrality. The left code block shows page rank on a drug interaction graph. First you project an in-memory graph. Here we project drug nodes and their interacts with relationships as undirected. Then you call GDS.pager rank.stream on that projection. The yield clause pulls out node ID and score. You convert the node ID back to a node with GDS.youutil.node and return the drug name and its score ordered highest to lowest. The right block shows between this centrality on a location graph with a distance weight. Same pattern project, call, yield, return. The difference is the algorithm and the interpretation. So, what do these scores actually mean? PageRank asks, if you started at a random node and kept following relationships randomly, how often would you end up at each node? Nodes that many other high-scoring nodes point to score highest in a drug interaction graph. A high page rank drug is the one that's most broadly connected, co-prescribed with many other commonly co-prescribed drugs. In a citation graph, it's the most influential paper. Between this centrality asks something different. How many shortest paths between all other pairs of nodes pass through this node? A high betweenness score means this node is a critical connector, a bridge. Remove it and many pairs of nodes lose their most direct route to each other. In supply chain graphs, these are your critical logistics hubs. In biological networks, they're the regulatory genes that connect multiple pathways. Community detection algorithms answer the question, which nodes naturally cluster together. They're the foundation of customer segmentation, fraud ring detection, and scientific topic clustering. The left block shows Lou using right mode. We're writing the community assignment back to each drug node as a property called community. After the write, we query those properties to see which drugs were grouped together and how large each community is. The modularity score tells you how well separated the communities are. Higher is better. The right block shows label propagation using stream mode. Label propagation is faster. Each node starts with its own unique label and iteratively adopts the label most common among its neighbors. It converges quickly which makes it useful for very large graphs or when you need results in seconds rather than minutes. The comparison table at the bottom summarizes the trade-offs. Luane produces higher quality communities because it explicitly optimizes modularity, but it's slower on very large graphs. Label propagation sacrifices some quality for speed. It's highly parallelizable and great for scale. In practice, many teams use label propagation for fast exploration, then luane on the interesting subgraph to get more precise community boundaries. The two algorithms complement each other. Well, similarity algorithms find nodes that look alike based on either the structure of their connections or the content of their feature vectors. They're the engine behind recommendation systems, drug repurposing pipelines, and fraud cluster detection. The left block shows node similarity with the jacard coefficient. Two nodes are similar if they share many of the same neighbors. The formula is the number of shared neighbors divided by the total number of distinct neighbors between the two nodes. A score of one means identical neighborhoods. Zero means no overlap at all. We're filtering to only return pairs with similarity above 0.5, a useful threshold to avoid noise. The right block shows a two-step pipeline. First, generate node embeddings with fast RP, a fast random projection algorithm that encodes each node's structural position in the graph as a vector. Then run K nearest neighbors on those embedding vectors to find each drug's five most similar neighbors by cosine similarity. The difference between these two approaches matters. Node similarity is based purely on graph structure shared connections. KNN on embeddings captures higher order structural similarity. Nodes that play similar roles in the graph even if they don't share direct neighbors. Both are valuable and which to use depends on what kind of similarity is meaningful in your domain. Three practical use cases. Drug repurposing. Find drugs with similar interaction profiles to a known effective drug. They become candidates for treating new indications. Recommendation engines. Products sharing the most co-purchasers by jakard coefficient are good also bought suggestions. Fraud detection. Accounts with similar transaction counterparties and device fingerprints form clusters that a human analyst might miss, but a KNN query surfaces in seconds.