July 2023 Summaries

Neo4j is being used with Unreal Engine for a Conversational AI system that can have intentional dialog moves, meaning it has goals and tries to direct the graph configuration towards a desired one. The system uses Behavior Trees (BT) and Bayesian Networks (BNs) to manage decisions and select strategies to improve the graph configuration. BNs are integrated into Unreal using the aGRuM library, allowing them to dynamically assemble networks based on collected information. The system can ask questions to collect hard evidence about variables in the network and use this information to make decisions. It also presents inferential statements to support its positions, which may be considered argumentations if they aim to persuade the interlocutor or explanations otherwise. The system's structure is a computational interpretation of linguistic theories, allowing it to test such theories by deploying them in specific domains like movie recommendation tasks.

Building a Recommendation Engine Using Neo4j Hands-On — Part 2` is an in-depth guide on how to build a recommendation engine using Neo4j, a graph database. The article covers the process of loading data into Neo4j, creating relationships between nodes, and implementing a recommender system that answers three key questions: what does a user generally order most frequently, which items are frequently ordered along with an item amongst all users, and once a particular item has been added to the cart, which other items have been previously ordered by the user with it. The guide also discusses deployment options for Neo4j and serving recommendations through API endpoints.

Neo4j has integrated its graph database with Confluent Cloud's event streaming platform, enabling developers to build real-time graph analytics applications. This integration allows customers to stream data from Confluent Cloud into Neo4j AuraDB, unlocking valuable use cases in network analysis, digital twins, cybersecurity, fraud detection, and trustworthy AI experiences. With Custom Connectors, users can extend Confluent Cloud beyond the fully managed connectors available on their platform, providing flexibility to create custom plugins for their specific needs. This powerful combination enables developers to build graph-based streaming applications with real-time insights, simplifying data integration architecture, building robust and scalable data pipelines, and leveraging the strengths of Neo4j's graph database and Confluent Cloud to deliver enhanced user experiences. The integration is supported by Neo4j's support contract, and customers can start building their graph-powered cloud applications today by checking out the available resources on Neo4j's Connector for Confluent.

This week's newsletter covers various topics, including a blog on Entity Resolution in Reagent by Matthew Filbert, which delves into problem modeling and algorithms used to create an ML pipeline. The newsletter also recommends Artem Ryasik's blog, ArtGraph Cluster Analysis, for its detailed explanations of knowledge graph analytics and clustering algorithms. Additionally, it highlights upcoming events such as LiveStream: Discovering AuraDB Free with Michael & Alexander on July 17, 2023, and a Sydney Meetup: Exploring Power of Connected Data on July 20, 2023. The newsletter also features Ajmal Aziz, a Senior Solutions Engineer at Databricks, who presents a reference architecture that leverages Neo4j's strengths in connected data and big data. Furthermore, it provides a tutorial on Entity Resolution in Reagent and an introduction to Graph Data Science, as well as a tweet of the week from William Lyon about the latest episode of the Graph Data Science podcast.

Enterprises want to utilize Large Language Models (LLMs) in their critical applications, but the unpredictable nature of LLMs can lead to hallucinations or inaccuracies. Retrieval augmented generation is a key consideration for overcoming these challenges by grounding the LLM in facts. Knowledge graphs and vector databases are two primary contenders as potential solutions, but knowledge graphs offer more accurate, reliable, and explainable foundations for LLMs due to their ability to provide precise information based on traversing connected relationships. While vector databases can connect factual pieces of information together, they often rely on similarity scoring and predefined limits, leading to incomplete or irrelevant results. Knowledge graphs, with their human-readable representation of data, offer full transparency and the ability to identify misinformation, making them a better choice for backing LLMs in mission-critical applications.

NeoDash 2.3 introduces a fresh new look, improved performance for handling large dashboards, and exciting visualization features. The update also includes a natural language interface for Neo4j, allowing users to translate English to Cypher on the fly with the help of Generative AI models, such as OpenAI's LLMs. To set up this extension, users need a Neo4j database, an OpenAI account, and follow specific steps to configure the extension and start writing natural language queries. The extension provides features like prompt engineering, schema retrieval, retry logic, Cypher validation, and more, with best practices for using it effectively, such as choosing a visualization first, calling out node labels and relationship types, being precise about values, and specifying what you want to achieve.

Knowledge graphs have significantly changed data accessibility with the rise of Large Language Models (LLMs). Retrieval-augmented LLM applications retrieve additional information from various sources to generate better and more accurate results. Vector similarity search is a strong bias in these applications, but structured information also has an important role to play in LLMs. Knowledge graphs can support LLM applications where users are interested in answering questions requiring highly-connected information, such as finding the shortest paths between data points or understanding complex biomedical relationships. They can also analyze supply chain scenarios and provide real-time insights into employee behavior and skills. The combination of structured and unstructured data retrieval paves the way for more accurate, reliable, and impactful results, extending beyond natural language answers into the realm of visually represented information.

To make a knowledge graph, it is useful to have a vocabulary in place, which is called an ontology. The Medical Subject Headings (MeSH) is one such ontology that includes many medical terms and can be downloaded as an RDF file, making it easy to import into Neo4j with neosemantics. Importing MeSH directly into Neo4j takes time due to the large number of nodes and relations. Descriptors, concepts, and terms are closely related in MeSH, where descriptors hold broad categories, concepts provide specific definitions within descriptors, and terms offer synonyms for concepts. Understanding the hierarchical structure is crucial for effective graph analysis, particularly with tree numbers being a more reliable way to establish relationships than broader descriptors. Leveraging MeSH as an ontology can be highly beneficial for constructing medical knowledge graphs, providing valuable insights from medical literature and data.

The Neo4j GraphQL Toolbox is a powerful tool for building GraphQL APIs using the @neo4j/graphql library. Released in July last year, it provides a developer user interface to quickly get started with an autogenerated GraphQL API for Neo4j databases. The toolbox has undergone significant updates, including a fresh and updated look, improved inner workings, and enhanced editors for type definitions and queries. These changes include the use of Needle's design system, a robust storage solution based on Zustand, and features such as auto-completion, code block collapse, and tabbed query pages. The toolbox is open-source and offers a free instance of Neo4j AuraDB to get started. The development team welcomes feedback and improvements to make the tool even better.

This week's newsletter highlights various topics in graph data science and related technologies. Alison Cossette shares her expertise on graph algorithms, including community detection, pathfinding, and graph sampling, with working code samples available for experimentation. Meanwhile, angrykoala provides tips on improving the performance of a Node.js GraphQL server using tools like @neo4j/graphql 3.19. Additionally, there's a live stream on Graph-Based Linguistics by Maria Di Maro, who explores the potential of graphs in extracting intricate information from text data. The newsletter also features a book review, podcast, and tweet of the week, as well as upcoming events and featured nodes speaker Donovan Bergin, an expert software engineer working with graphs at JB Hunt.

Neo4j Graph Data Science has introduced three new algorithms: K-Core Decomposition, Bellman-Ford shortest path algorithm (SPFA), and Common Neighbour Aware Random Walk (CNARW). These innovative algorithms can solve pain points such as negative weights in pathfinding, community detection with varying density levels, and efficient subgraph sampling. The K-Core Decomposition algorithm is used for community detection, allowing users to analyze community structures and identify influential nodes. The Bellman-Ford shortest path algorithm enables the calculation of the shortest paths in graphs with negative weights, which can be useful in IT networking, supply chain optimization, and finance. The Common Neighbour Aware Random Walk (CNARW) algorithm is a graph sampling technique that helps scale machine learning on large graphs by considering common neighbors of nodes. These new offerings provide Data Scientists with expanded tools to refine the ability to find the right algorithm for the right problem.

The text discusses building a recommendation engine using Neo4j, a graph database. The author aims to create a system that recommends items from a menu based on user preferences and order history. The system will use two types of filtering: collaborative filtering and content-based filtering. Collaborative filtering is based on the idea that users with similar interests will also like similar items, while content-based filtering is based on the idea that people generally prefer certain types of products or services. The author designs a data model to store user orders, menu items, and relationships between them. The system will recommend items based on a user's order history and information about co-occurrence of items among all users. The goal is to provide personalized recommendations to users when they open the app or website.