In the early 2000s, all major companies started focusing on the latest and greatest technologies and became tech-centric. The same companies further evolved themselves to being data-centric in the 2010s. Nowadays, in order to keep up with the evolving times, these companies have further progressed and are focused on being a decision-making company by gaining understanding of rich data relationships.

A Knowledge Graph (KG) is a natural data model in many real-world situations. KGs captures useful relationships and, with the help of graph embeddings, it aggregates a vast amount of knowledge(domain) into the lower dimensional representation (embedding spaces).

It’s not just the knowledge, but the interpretation of the knowledge. The ability to think and see how these pieces of information connect plays a huge role in extracting value from an organization's data.

‍

1. Introduction

A graph is a natural way to express a collection of objects and their connections. In a real-world scenario, relationships between an object with any other related object can be very easily defined using a graph. These relationships can be articulated by using graph topologies such as nodes, edges, and attributes, allowing for semantic searches or exploring similarities.

A common implementation of graphs is within social networks, for example, graphs can be used to view people who are linked/associated to each other in the network, such as finding a friend of friend, suggestions of new friends, finding events based on you and your friends’ common interests, or even capturing important dates such as the date you became friends or other shared life events.

Graphs are the most practical way to collect and connect the high-dimensional data points and depict and explore the relationships between all of the data points that are available from various sources and in different forms.

‍

2. Knowledge Graph Basics

A graph is a representation of schema-free objects (vertices or nodes) along with relationships between the objects (edges). It is a network of data points formed by vertices and edges. Formulating a domain-specific graph results in a “Knowledge Graph (KG)”, which represents all the information about nodes and relations between nodes in the given context.

‍

2.1 Components

A Knowledge Graph (KG) consists of the following core elements:

Nodes →

• Each entity in a graph is referred to as a node. Nodes are represented as circles (dots). The terms node and vertex are used synonymously. Nodes can be of different types, such as User or Transaction, and each of them can have unique attributes in a given context. Perhaps nodes can be:

• Connected with more than one other nodes via multiple connections.
• Connected to themselves (Self-looped).
• Disconnected from the graph, having no connecting edges.

‍

Edges → 

• The connections that exist between nodes, that are represented as lines or arcs. Edges, in other words, store information about the relationships between nodes. They can have properties like key/value and can be directed.

Graph → 

• A graph is the type of data structure that represents the relations (edges) between a collection of entities (nodes).

‍

Graph Neural Network → 

• Graph Neural Networks (GNN’s) are special types of neural networks capable of working with a graph data structure. A graph is passed through a series of Neural Network (NN) layers. The input graph structure is converted into graph embedding, allowing us to maintain information on nodes, edges, and global context. GNN’s can further be used for multiple downstream tasks such as:

• Node Classification: this task uses neighboring node labels to predict missing node labels in a graph.
• Link Prediction: predicts the link between a pair of nodes in a graph. It is commonly used for social networks.
• Community detection (Knowledge Spaces): divides nodes into various clusters based on node/edge similarities and topological graph structures.
• Semantic Search: it performs search operations on the graph embeddings to capture contextually similar nodes and edges.

Graph Embeddings → 

• are the projections of nodes and edges into a continuous low-dimensional space.

‍

3. Knowledge Graph System

KG incorporates the domain knowledge by consuming the SQL-like data structure from the enterprise data (primary and foreign key relationships) and transforming it into a Graph-like representation.

‍

We further learn the relationship between multiple entities by training the graph data on state-of-the-art Graph Neural Network (GNN) algorithms in a supervised way. During the learning process, it encapsulates relevant relations and context around each entity by summarizing them into Graph Embeddings in vector form.

Each node and edge can be visualized in a 2D plot. Users can interact with this 2D plot to examine, explore, annotate and take relevant actions such as creating clusters (knowledge spaces) and transforming them into features or rules.

Knowledge Spaces

A segment of enterprise data that captures common behavior among entities can be grouped together. Users can use the “Graph Annotation” flow to create such Knowledge Spaces. It will also recommend some high-quality spaces which can be relevant for the users such as spaces for good/bad users, etc.

‍

‍

Each of these spaces has something unique to narrate about the enterprise data. Users can group various data behaviors, such as “high-paying male customers” into one space for exploration;

Users can also group model behaviors, such as “all the False Positive predictions” into features or rules for data and model enrichment.

‍

4. Knowledge Graph and Beyond

Knowledge graph with explainable AI capabilities, offering a wider spectrum of semantic search, new cohort exploration, impactful feature adoption, a model improvement over time with a human in the loop for feedback integrations, and more.

• Explainability (XAI): Explainability (also known as “interpretability”) refers to the idea that a machine learning model and its output can be explained in a way that “makes sense” to a human at an acceptable level. It provides a platform that is built on the core pillar of transparency, making every decision transparent by describing the contributions of each feature in the decision-making process both globally and locally.
• Semantic Explorer: The Knowledge Graph is a semantic database in which information is structured so that rich knowledge can be generated from it. It uses this as its second core pillar to build foundations that are semantically rich in nature, and by using a great interface, users can access/explore all of the data in one place.
• Space Recommendations: KG can automatically detect high-fidelity information which creates interesting cohorts for business, and hence can recommend such spaces which add value to the better model performance and explore new cohorts in the business.
• Automatic Feature engineering: The platform generates new engineered features as a byproduct. KG can extract useful and meaningful features from any problem by utilizing the KG and XAI components. This enables users to be more productive by allowing them to spend more time on other meaningful tasks.
• Reporting: KG generates meaningful reports for the stakeholders to take a deep-dive and overview of all the model performance metrics, explainability metrics, correlation between data, and many more.

‍

5. Identifying use cases

5.1 Use cases for knowledge graph systems

• Fraud Detection — Detecting fraud necessitates complex pattern matching, which also requires the graph structure involving connections (e.g. a disproportionate number of connections between various entities and accounts, IP addresses, and so on) as well as joins, associative queries, and statistical analysis in many cases. A knowledge graph can be used to model the process of assembling and integrating a large amount of data.
• Enterprise Knowledge Graph — Enterprise Knowledge Graphs (EKGs) are becoming increasingly popular as extremely valuable tools for harmonizing internal and external data relevant to an organization into a common semantic model to improve enterprise operational efficiency and competitive advantage for business units.
• Recommendation Engine — There are numerous approaches and techniques for generating recommendations, and the majority of them work perfectly with knowledge graphs.
• Social Media Management — Social networks are an excellent example of large, highly connected graphs, and they frequently involve graph algorithms and graph traversal.

‍