Neo4j GDS Certification key terms

An in-memory copy of part or all of a graph, built from node labels and relationship types, that algorithms run against so the stored graph is never modified during computation.

A projection created directly from stored labels and relationship types, loaded fast from Neo4j's own storage and supporting node and relationship property filters without running a query.

A projection defined by a Cypher query that returns node and relationship columns, allowing computed values, filtering, and unions that native projections cannot express.

The registry that stores named graph projections in memory, tracking which exist, their node and relationship counts, memory use, and which algorithms have been run on them.

Removing a named projection from the catalog to free heap memory, done explicitly or automatically when the database restarts or the projection ages out under configured limits.

Selecting node properties to carry into the in-memory graph so algorithms can read weights, labels, or feature values without querying the stored graph during execution.

A numeric relationship property used by weighted algorithms such as Dijkstra or weighted PageRank in place of uniform cost, changing which paths or scores the algorithm produces.

Persisting algorithm output to the stored graph as node properties or new relationships, using a write mode that mutates the database rather than only returning results.

An execution mode that returns algorithm results as rows to the client without changing the stored graph or the in-memory projection, useful for inspection and custom post-processing.

An execution mode that computes results and stores them into the database as node properties or relationships, while also reporting summary statistics such as nodes written and execution time.

An execution mode that writes algorithm results back into the in-memory projection only, enabling chained algorithms in one workflow without touching the stored graph or needing a new projection.

An execution mode that returns only summary statistics about an algorithm run, such as counts, timings, and configuration, without streaming rows or writing results anywhere.

A call that predicts the memory a projection or algorithm run will need, letting candidates size the heap and avoid out-of-memory failures before starting expensive work.

The JVM memory region where GDS stores projections and algorithm working state; insufficient heap is the most common cause of projection or algorithm execution failures.

A centrality algorithm that scores nodes by the number and importance of incoming links, iterating until scores converge; widely used to rank influence in directed graphs.

A PageRank parameter, typically around 0.85, controlling how much score flows along relationships versus being redistributed uniformly, affecting convergence and ranking stability.

A PageRank variant that reduces the influence of low-degree nodes, producing rankings better suited to graphs where highly connected nodes would otherwise dominate score propagation.

A simple centrality measure counting a node's relationships, optionally filtered by orientation or relationship type, indicating how directly connected a node is.

A centrality measure scoring nodes by how often they lie on shortest paths between other node pairs, identifying brokers or bridges in a network.

A centrality measure based on the average shortest-path distance from a node to all others, so nodes that can reach the rest of the graph quickly score higher.

A closeness variant that sums reciprocal distances instead of averaging them, so it handles disconnected graphs and unreachable nodes without producing undefined scores.

A modularity-optimizing community detection algorithm that iteratively groups nodes into communities, then aggregates them into a smaller graph until modularity stops improving.

A community detection algorithm improving on Louvain by refining communities to guarantee well-connected groups, avoiding badly split or internally disconnected communities.

A score measuring how much more densely connected a graph's communities are than expected by chance, used to compare community detection results and tune resolution.

A fast community detection algorithm where nodes repeatedly adopt the most frequent label among neighbors until labels stabilize, scaling well to large graphs at some accuracy cost.

An algorithm partitioning a graph into maximal sets of nodes reachable from each other when relationship direction is ignored, useful for isolating disconnected subgraphs.

An algorithm finding maximal node sets where every node can reach every other node following relationship direction, revealing cycles and tightly coupled regions in directed graphs.

A similarity algorithm comparing nodes by the overlap of their neighbor sets using Jaccard or overlap coefficients, commonly used for recommendations and link prediction.

A similarity coefficient dividing the size of the intersection of two node neighbor sets by the size of their union, ranging from zero to one.

A similarity algorithm that, for each node, keeps only the top-k most similar neighbors, reducing output size and enabling graph-based classification or recommendation.

A shortest-path algorithm computing the minimum-cost path between a source and target node using non-negative relationship weights, returning the path and total cost.

A heuristic shortest-path algorithm that uses a per-node estimate of remaining distance to the target, often exploring far fewer nodes than Dijkstra while returning the optimal path.

An algorithm returning the k lowest-cost loopless paths between two nodes, useful when the single best path is unavailable or alternatives must be compared.

An algorithm selecting a subset of relationships connecting all nodes at minimum total weight, used for network design and clustering tasks on weighted graphs.

A convergence parameter for iterative algorithms such as PageRank, specifying how small the change in scores between iterations must be before computation stops.

A cap on how many rounds an iterative algorithm performs, bounding runtime when convergence is slow or the graph is large and exact convergence is unnecessary.

A node property used to initialize community labels in algorithms like Louvain, enabling incremental or reproducible community detection across successive runs.

A configuration restricting which nodes an algorithm considers, letting candidates run the same algorithm on subsets of a projection without rebuilding it.

The statistics returned by every algorithm run, including node and relationship counts, execution time, configuration used, and whether results were written or streamed.