Neo4j Inc. is targeting artificial intelligence and machine learning applications with a new version of its graph database announced today.

Graph engines differ from conventional relational and NoSQL databases in that they document connections between data elements. That enables organizations to map relationships that would be impractical or impossible to represent in other database engines, such as those among all the characters in “Game of Thrones” (pictured).

Graph databases are increasingly being used in AI scenarios because of this unique capability, said Philip Rathle, vice president of products at Neo4j. “The problem with NoSQL databases is that they give you data but not the connections,” he said. “People are now looking to bridge connections to make better decisions with context.” For example, AI-driven voice-response systems are more effective when they understand the context of a command.

Graph engines are also well-suited to scenarios in which decisions made by an algorithm must be explained, Rathle said. For example, if a bank customer contests a denied loan application, “both the bank and regulators are going to want to know why the machine made that decision,” he said. “Many machine learning and deep learning processes [based upon conventional databases] are so complex that it can take weeks to understand.”

Version 3.5 adds full-text indexing to enable text-intensive applications such as metadata management and bill-of-materials processing. The feature, which is based on the open-source Lucene engine, “is one of the most highly requested features we’ve had for years,” Rathle said.

Native index support has been expanded to include spatial, temporal and Boolean values along with composite indexing. The result is up to a fivefold performance improvement, the company claimed. “It turned out that 80 percent of Neo4J’s time was being spent on indexing during large data ingestions,” Rathle said. Native indexes are now also used for sorting operations for queries written in Neo4j’s Cypher language.

The second most-common user request has been for improved handling of large write transactions, so a new memory subsystem was added for that purpose. The so-called off-heap transaction subsystem, coupled with clustering, uses native memory more efficiently to propagate large writes throughout the cluster. Off-heap memory management bypasses the garbage cleanup process in Java that can create processing pauses. “We’ve implemented some specific low-level memory management rather than relying on Java,” Rathle said.

Other new features include a driver for the increasingly popular Go programming language developed by Google LLC and graph algorithms for unsupervised learning methods such as Random Walks, Personalized PageRank, Similarities, DeepGL and DeepWalk.

Image: Neo4j