Graph intelligence is a bold new movement encapsulating the best of graph technologies in today’s world of enterprise Artificial Intelligence (AI). It enables organizations to complete end-to-end workflows in a single platform, is fortified by a scalable graph computing engine, and eliminates unnecessary data movement.
The four pillars of graph intelligence are graph querying, graph mining, graph analytics, and graph AI. Although graph AI represents the upper end of the spectrum of graph intelligence’s potential for compounding business value across use cases, it’s predicated on the groundwork of the other elements — which is why a single platform for these capabilities is crucial.
Graph AI’s crowning achievement is its predictive and prescriptive capacity for presaging domain-specific events while perfecting businesses’ responses to them. For example, graph AI can reveal just how a cyber security breach occurred and what measures firms can take to prevent other attacks from succeeding.
Graph AI supports all aspects of AI (from data preparation to neural network-based prediction) within the graph intelligence framework. This quality decreases data movement while boosting efficiency. There are also several AI techniques — including Graph Neural Networks (GNNs), graph embeddings, and unsupervised learning approaches like clustering — that work best in graph environments.
As part of the larger graph intelligence construct, graph AI is the key to maximizing AI deployments because of graph’s natural benefits of connecting data and detecting relationships between them, culminating in contextualized, results-oriented AI.
Graph AI offers native support for the data engineering necessary to create AI models and perform reasoning techniques. All workloads run natively in graph intelligence servers, which is why scaling out with High-Performance Computing (HPC) methods that allow users to host data on multiple machines is integral to this paradigm. Specific graph AI workloads include data preparation, data discovery (enhanced by graph query), feature extraction, and predictive model building requisites for training models and annotating their data. Unlike other approaches requiring additional tools to annotate, clean, and train data, all the above steps occur in the single graph intelligence platform. Because graph AI is used to train and deploy models (or perform inferences) on the same servers that host data, users get fast responses for low latent applications like fraud detection — which requires minimal data movement for regulations.
Graph AI frequently utilizes the other three graph intelligence pillars to build AI models. Users can generate machine learning features from graph analytics results, which are often informed by graph querying and graph mining efforts, to create models with efficient workflows maximizing the productivity of valued data scientists. There’s also a tight Python integration so data scientists can orchestrate graphs directly in Python via its user interface. Building customized AI applications without learning new, proprietary tooling also optimizes data scientists’ efficiency. The Python functionality typifies the close interplay of graph intelligence’s querying, analytics, and AI capabilities. For instance, the first two are helpful for identifying features for machine learning applications built in Python. Graph AI also offers scalable training so users can train GNNs, for example, on hundreds of machines. Since storage is decoupled from compute, employees can always launch more servers to accelerate training.
AI in Graph Settings
Numerous AI approaches work best in graph settings, largely because their underlying models benefit from connections in the graph’s topology for superior accuracy. Unsupervised learning illustrates this fact well. It includes dimensionality reduction techniques as well as clustering, Principal Component Analysis, and more. The very nature of the graph representation makes this framework perfect for visualizing the results of clustering. Consequently, methods like Louvain clustering are gaining significant traction throughout FinTech for quickly building a hierarchical representation of data. Unsupervised learning lets users start learning right away without the time and cost of labeling data that supervised learning requires. Plus, the former’s learning is more authentic since it’s based on the data, not some human label.
GNNs naturally excel in graph environments. They work in unsupervised and supervised learning settings and let firms train neural networks on graph data to automatically extract knowledge from entities and their relationships. GNNs are adept at working on high-dimensionality data and performing node classification, graph classification, and link predictions between nodes. For example, GNNs can find links between parties and their funding to detect financial crimes for mandates like Anti-Money Laundering.
Graph embedding methods embed nodes and edges in a low dimensional space so they are easily processed by standard machine learning operations. This lets users find entity relationships that aren’t apparent when looking at the graph’s topology. While nodes with similar properties might not be close in the graph, after embedding, seemingly unrelated nodes are close together. Practical applications include using this approach to pinpoint machine learning features or link prediction in a medical knowledge graph. For this pharmaceutical use case, it’s expensive to try out new drugs in a laboratory. Graph embedding enables employees to narrow down different components for potential drugs to save money and time.
Graph AI in Action
The most meaningful graph AI use cases rely on its strengths of predicting events and prescribing action for them. This pillar often builds on graph analytics results. A good example includes capabilities for segmenting (and micro-segmenting) customers and their preferences. Netflix, for example, might identify the type of movies person A likes and those that person B likes to recommend movies both or a third person they know might like. Although this example may seem trivial, it’s applicable for issuing accurate recommendations for customer segmentation in verticals like financial services, for example, based on factors like whether they’re in rural or urban areas, or unobvious segmentations pertaining to customers’ spending habits. In the case of the former, one can employ graph AI to find points of commonality between them for comprehensive marketing campaigns appealing to both or a third group with homes in both areas. Graph AI provides similar advantages for finding factors contributing to churn and prescribing ways to decrease it.
For security analytics, graph AI is invaluable. It may leverage graph analytics, for example, to ascertain exactly what happened for a current or previous security event, then utilize that intelligence to predict what other events could happen and prescribe a credible defense. Graph AI constantly works with the other graph intelligence elements and its HPC architecture to solve problems of enormous scale. For precision medicine, the most accomplished doctor can only treat so many patients. Graph intelligence can consider patient information for millions of patients and additional information like literature and conference transcriptions to inform use cases like disease detection or drug research.
Graph AI relies on each of the other graph intelligence components for predictive and prescriptive insight into use cases across financial services, healthcare, network security, and other areas. It implements all aspects of the AI pipeline in one platform and maximizes the efficiency of the underlying system.
This article was originally published by Bo Wu, Principal Engineer at Katana Graph, on Medium.