GraphConnect 2020

When do you use graphs for machine learning, what domains can they be used in, and how do you get started. Real world examples and use cases to show the steps from getting started with a knowledge graph through to graph native learning. Graphs - or information about the relationships, connection, and topology of data points - are transforming machine learning. We'll walk through real world examples of how to get transform your tabular data into a graph and how to get started with graph AI.

This talk will provide an overview of how we to incorporate graph based features into traditional machine learning pipelines, create graph embeddings to better describe your graph topology, and give you a preview of approaches for graph native learning using graph neural networks. We'll talk about relevant, real world case studies in financial crime detection, recommendations, and drug discovery.

This talk is intended to introduce the concept of graph based AI to beginners, as well as help practitioners understand new techniques and applications. Key take aways: how graph data can improve machine learning, when graphs are relevant to data science applications, what graph native learning is and how to get started.

Leveraging Neo4j, Meredith has been able to create an Enterprise wide Identity Graph with 12.3 Billion Nodes and 18.9 Billion Relationships that represents its Digital presence over the last 18 Months across 30+ brands. Neo4js Graph Algorithms enable unique Digital Profiles for Premium Ad Sales. Digital Marketing and Custom Audience Advertising is a focus of any Media Publisher. Meredith Corporation is leading the field of first party data retargeting using Neo4j for an in-house Identity Graph containing every digital cookie seen across 30+ brands and applications, including People, Entertainment Weekly,Real Simple, Eating Well, and All Recipes. With 100 million unduplicated consumers, the Meredith Database is the largest U.S. consumer database of any media company, and includes 7 in 10 women and 8 in 10 homeowners. The Identity Graph solution has leveraged 10's of Billions page views across multiple data streams to create a graph comprising of 12.2 Billion Cookies and 18.9 Billion Relationships between them to identify recurring individuals across domains and devices to improve Custom Audience Advertising and Profiling.

Lockheed Martin Aeronautics has integrated graph technology into its technology landscape empowering end users to build and visually explore their data models more effectively than traditional methods. During our graph implementation journey, we developed a self-service operating and support model to enable our users. Equipped with Kettle, Neo4j and Linkurious, business users have been able to develop their own graph models to answer business questions, leveraging developer-consultants for complex solutions. This has proven to be successful in creating a graph community within Lockheed Martin Aeronautics.

At Healthfirst, we're on a journey to use graph to improve the health of our members. We began with a small pilot and scaled up to a KG, allowing us to analyze connections between members, claims, and providers: we’re using Neo4j for fraud detection, reducing costs, and improving care patterns.    

We'll start the talk with an overview of Healthfirst and our business model. We'll discuss the business problems outlined to us by stakeholders, and how we helped them frame their questions in a way that could be answered with graph analytics.

We'll review our very first use case (fraud detection), which started with a Neo4j trial off the side of our desks, the results of which were powerful and gained us the support to scale up to a Neo4j server and several licenses, expand the fraud detection use case, and launch a new use case for out of network utilization analysis.

We will focus the talk how we collaborated w/ business partners to execute the use cases and how the results are being used to improve quality of care for our members, but will also touch on the graph schema, data sources/ ingestion patterns, and graph algorithms used for the analyses."

When you grow corn, yield is paramount. How much a corn seed will yield is encoded in its genome. If you visualized that genome in ASCII characters, you would see a seemingly random string of 25 million A, C, G, and T characters. Somewhere in that string you could find the interesting portions that governed how much corn that seed will yield, as well as others that provide useful properties to the corn plant making it resilient to different pressures.

At Bayer Crop Science, we track these interesting portions of the Corn genome as numeric intervals: start and stop indices within the string of 25 million characters. Our goal is to find relationships between intervals of interest to determine how to breed plants to produce the greatest yield and be resilient for different conditions around the world.

Our team, the developers of an internal genomics software stack within Bayer Crop Science, have been challenged to provide an API to our internal customers for efficiently finding these related intervals. In exploring solutions we came upon the interval tree data structure and implemented a means for storing and querying interval trees in Neo4j.

In this session we will discuss and demonstrate our approach as well as a set of Neo4j stored procedures created by Bayer Crop Science to effectively manage, retrieve and search interval tree data structures in large scale.

Molecules are graphs! When you change part of this graph, swap one part out for another, add something in here, remove a little there you change how that molecule behaves and interacts with your body. This talk models alterations of molecular graphs as a network and applies it to drug discovery!

**Molecules are graphs!** The nodes are atoms and the edges are bonds. What happens when we take these graphs and their properties and put them in a graph database?

**Drug discovery is expensive**. The cost of producing a new drug is estimated to cost $2.6 Bn with a significant chunk of that cost coming from research and development of the drug molecule. Reducing down the cost of developing new therapeutics is key to helping patients. If researchers can make smarter decisions earlier in R&D the timelines and cost of bringing new drugs to patients will be reduced.

**Matched molecular pair** analysis (MMPA) is a method used in chemoinformatics that compares the properties of two molecules that differ only by a single chemical transformation (or graph alteration). An example of this would be if you were looking at two molecules that only differed by the substitution of a hydrogen atom for a fluorine atom. These two molecules whilst almost identical could have vastly different chemical properties. These **pairs** of compounds are known as **matched molecular pairs** (MMP), and any change in the properties of these molecules can be modelled on an edge linking them together.

This talk will explore how combining biological assay data with these chemical transformations in a **matched molecular pair knowledge graph (MMPKG)** using Neo4J allows for powerful exploration of chemical data. Through the use of Neo4j browser, cypher, and graph algorithms library new insights can be gathered to help answer the question on most medicinal chemists lips... _""which molecule should I make next?""_ and exactly how the MMPKG can be used and applied to real drug discovery problems to help drive this decision making process."

The intelligence community has applied link analysis to everything from modeling call records to financial transactions. But what happens when you apply the same techniques to the technical artifacts of cyberattacks? How do you avoid overthinking your data model when modeling such complex data?    

Cybersecurity may be the ideal domain for graph analysis as the relationships between technical attributes are often more critical than the discrete values. For example, an attribute's maliciousness often depends on the surrounding context. This can include the presence or absence of other attributes, the behaviors that those attributes exhibited, and the similarity of that behavior with other attack vectors. Graphs and contextual link analysis are very effective mechanisms for identifying potentially malicious activity.

However, before performing any type of analysis, you need to create the data model! While many graph data model examples are reasonably straightforward, the modeling of cybersecurity data can become quite complex. You would ideally model the attributes of any real world artifact (e.g., an email or file), the occasions in which those attributes were seen together, the behavior that those attributes exhibited when they were observed, and the source of your knowledge about those relationships. But how much knowledge do you really need to encode in the graph? When should you rely on path traversals rather than leveraging more advanced graph algorithms? Do you need to create a hyper graph in order to capture the source of relationships? What are the performance implications? How do you expire data from the graph? And finally, how do you make some decisions and actually build something?

Money launderers use complicated schemes to wash dirty money, and financial institutes need to fight back with advanced techniques. Using graph analytics, it becomes feasible to connect dots in very complicated schemes that traditional methods cannot handle. Nowadays, money laundering involves leveraging different types of financial instruments with more complicated schemes. Preventing money laundering and terrorist financing has become high priority for financial institutions.

Traditional methods of monitoring for AML typically involve static, rule-based alerts built from previous experience. The biggest challenge faced by traditional approach is that money laundering schemes continuously changing in a way that is difficult to detect. With the power of graph analytics, it becomes feasible to connect dots in complicated schemes that traditional methods cannot handle. In this talk, we will share our experience using advanced rules traversing hidden patterns in the data, creating graph-based features for machine learning and finding similar patterns using graph embeddings.