GraphConnect 2018: Linking AI and Graph DB
Leading Graph Database company Neo4j organized GraphConnect 2018 on September 20 in New York City to introduce their latest products and developments — foremost among them the new Neo4j 3.5 Graph Platform. The conference gathered experts from Neo4j’s R&D, service, and marketing departments; along with customers and academics, to share Neo4j experiences and knowledge. Synced was on the scene to bring our readers the latest on Graph Databases (Graph DB) and Artificial Intelligence.
Rising Popularity of Graph DB
The conference keynote speech was given by Neo4j Fo-founder and CEO Emil Eifrem, who presented the latest DB-engines ranking, which show Graph DB as by far the fastest growing database category since 2014.
“Every single hotel booking reservation at Marriott Hotels — including Starwood — is done by Neo4j,” said Emil. “Ninety-nine percent of all airfare calculations in the world are being done with Neo4j, and twenty of the top twenty US banks use Neo4j today.”
What’s New in 2018 & Neo4j 3.5
Neo4j Product VP Philip Rathle took the stage to talk about “What’s new in 2018” for Neo4j. Philip first provided an overview of the company’s Neo4j 3.4 Graph platform, which was updated this May with Multi-Clustering, 3D Geo-Spatial, and improved performance. The Neo4j graph platform ecosystem was also introduced, together with recently released data visualization tool Neo4j Bloom and updates on Neo4j engineers’ work in aspects of graph analytics: Graph algorithms, Cypher for Apache Spark, and Graph-Enhanced AI & ML. Some of those features have already been implemented in the new Neo4j 3.5.
Philip said Neo4j 3.5 will be available in 2018 Q4, with three highlight features:
- Full-text search: “One of the most requested features” which applies to both nodes and relationships in the graph. Full-text search allows users to locate examples regardless of how they are worded, which enables more possibilities with regard to natural language processing (NLP) tasks.
- Accelerated Data Ingestion: Thanks to the removal of unnecessary Lucene features, end-to-end performance of data importation has been improved five times, which is important and helpful for connected feature extraction and machine learning training processes.
- Go Language Driver: Golang is a fast-growing programming language that supports CPU-level parallel processing very well. With its growing popularity in various applications including AI, the Go Language Driver will make Neo4j 3.5 more accessible and useful in the Go community.
Although the official version of Neo4j 3.5 will not be available until Q4, 2018, the latest working version — Neo4j 3.5.0-alpha01 can be downloaded now.
Graph Enhanced AI & ML
When Eifrem returned to the stage he talked about how Graph DB relate to machine learning (ML).
Eifrem explained that Graphs provide connections and context which better enable AI and ML to support a variety of tasks such as predictive/prescriptive analytics, NLP, and making recommendations.
There are four pillars which summarize how Graph tech is changing AI & ML. these were discussed in detail by Neo4j’s Jake Graham and Amy Hodler in one of the 13 talks regarding AI/ML and Graphs.
Knowledge Graphs — Context for Decisions: Graph DB is a connected, dynamic, and understandable repository of different data types. They can link siloed or external data sources in an intelligent way. There are in general three types of knowledge graphs: Context Rich Search, External Insight Sensing, and Enterprise NLP.
Connected Feature Extraction — Context for Accuracy: Graphs add highly predictive features (relationships) to machine learning models which usually rely on vectors, matrices, and tensors built from tables. Neo4j methods for connected feature extraction include:
- Engineered features (labelled and inferred relationships)
- Graph algorithms (e.g. centrality and community detection)
- Graph embeddings (DeepWalk, DeepGL, Node2Vec, etc.)
Graph Accelerated AI — Context for Efficiency: Graphs are useful for model optimization and accelerating the ML process. There are always sparse matrices involved in machine learning which are easier to compute and more resource cheap to operate as graphs. Common methods to do so include:
- Replace table joins with graph queries
- Replace sparse matrices and directional relationships with more efficient graph structures (i.e. collaborative filtering via cypher query vs. matrix factorization)
- Use subgraph filtering to accelerate ML pipelines (cypher queries, collaborative filtering, community detection, clustering, etc.)
AI Explainability — Context for Credibility: The recent development of AI in aspects of deep learning is often criticized for its lack of explainability. Graph architecture can be useful in providing some transparency and visibility for the black-box, especially with the help of the graph visualization tool Bloom. Three types of explainability can be addressed:
- Explainable Data: Graphs provide data lineage — when, where, and often why data was accessed
- Explainable Predictions: Associating nodes in a neural network to a labelled knowledge graph allows for traversing related documents to an explanation
- Explainable Algorithms: Early research shows that constructing tensors from graphs using weighted relationships may lead to explainable neural network algorithms
Neo4j is a NoSQL Graph Database developed by Neo4j Inc. (Neo Technology), and today’s number one platform for connected data. The company has more than 300 commercial customers including large enterprises such as Adobe, Walmart, Cisco, eBay, Microsoft, NASA, etc. Neo4j first hosted GraphConnect in 2012.
Author: Mos Zhang | Editor: Michael Sarazen
Follow us on Twitter @Synced_Global for more AI updates.
Subscribe to Synced Global AI Weekly to get insightful tech news, reviews and analysis! Click here !.
