ナレッジグラフ

A knowledge graph is a structured representation of real-world entities and the relationships between them. Unlike a standard database that stores data in rigid rows and columns, a knowledge graph organizes information as a network of nodes (representing objects, people, or concepts) and edges (representing the connections or interactions between those nodes). This structure mimics how humans organize information, allowing artificial intelligence (AI) systems to understand context, infer new facts, and reason about data in a more semantic and interconnected way.

Understanding the Structure

At the core of a knowledge graph are three main components that form "triples" (Subject-Predicate-Object):

• Nodes (Entities): These are the distinct data points, such as "London," "Python," or "Ultralytics YOLO26." In computer vision tasks, these might represent detected objects like a "Car" or a "Pedestrian."
• Edges (Relationships): These distinct lines connect nodes and define how they relate. For instance, an edge might label the relationship between "London" and "UK" as "is_capital_of."
• Attributes (Properties): Additional details describing a node, such as the population of a city or the confidence score of an object detection.

This web-like structure enables systems to perform semantic search, where the engine understands the user's intent rather than just matching keywords. For example, knowing that "Jaguar" is both an animal and a car brand allows the system to differentiate results based on context.

Integration with Machine Learning

Knowledge graphs are increasingly vital for enhancing machine learning (ML) models. While deep learning models excel at statistical pattern recognition, they often lack factual grounding. Integrating a knowledge graph allows models to access a verified "world view."

• Retrieval-Augmented Generation (RAG):Generative models can sometimes produce plausible but incorrect information. By grounding Large Language Models (LLMs) with a knowledge graph, AI agents can query a verified source of truth before generating a response. This significantly reduces hallucinations in LLMs and improves factual accuracy for enterprise applications.
• Recommendation Systems:In AI in retail, graphs map complex relationships between users and products. If a customer purchases a camera, the graph understands the functional link to "SD Cards" or "Tripods," enabling smarter suggestions than simple collaborative filtering.

コード例グラフのエンティティを抽出する

Computer vision models act as excellent entry points for populating knowledge graphs by identifying physical entities in the real world. The following Python snippet demonstrates how to use the Ultralytics YOLO26 model to detect objects in an image. These detected classes can act as nodes, which can then be linked in a graph database (like Neo4j or Amazon Neptune).

from ultralytics import YOLO

# Load the latest YOLO26 model (released Jan 2026)
model = YOLO("yolo26n.pt")

# Run inference on an image to find entities
results = model("https://ultralytics.com/images/bus.jpg")

# Extract unique detected object names to serve as graph nodes
# e.g., {'bus', 'person'}
detected_entities = {results[0].names[int(c)] for c in results[0].boxes.cls}

print(f"Graph Nodes: {detected_entities}")

実際のアプリケーション

1. 医療分野における創薬： 医療AIの研究者は知識グラフを活用し、生物学的相互作用をモデル化している。UniProtなどのデータベースから遺伝子、タンパク質、化学化合物といったエンティティを連結することで、アルゴリズムは潜在的な薬剤標的や副作用を予測し、新たな治療法の開発を加速させている。
2. サプライチェーンのデジタルツイン：物流企業はナレッジグラフを活用し、自社の業務の「デジタルツイン」を構築する。ノードはサプライヤー、倉庫、在庫を表し、エッジは輸送ルートと依存関係を表す。この構造はビッグデータ分析を促進し、管理者が遅延を予測し、動的にルートを最適化することを可能にする。

Knowledge Graphs vs. Relational Databases

It is important to distinguish a knowledge graph from a traditional Relational Database (RDBMS). A relational database stores data in rigid tables linked by foreign keys, which is efficient for structured, transactional data (like bank ledgers). However, querying complex relationships (e.g., "Find friends of friends who like sci-fi") requires expensive "join" operations.

In contrast, a knowledge graph (often stored in a Graph Database) treats the relationship as a first-class citizen. Traversing connections is instantaneous, making graphs superior for tasks involving highly interconnected data, such as fraud detection rings or social network analysis. While RDBMS excels at storage and retrieval of specific records, knowledge graphs excel at discovering patterns and hidden insights within the connections themselves.

Future Outlook with Multi-Modal AI

The future of knowledge graphs lies in multi-modal learning. As models like Ultralytics YOLO26 continue to advance in object detection and pose estimation, they will automatically feed visual context into graphs. This creates systems that not only "read" text but "see" the world, linking visual concepts to linguistic definitions. Using the Ultralytics Platform, developers can train these specialized vision models to recognize custom entities, effectively building the sensory organs for the next generation of knowledge-aware AI systems.