Data Freshness

Mean Time Since Update (MTSU) is a foundational latency metric that calculates the average elapsed time between the current moment and the last modification timestamp for all facts or entities in a specified subgraph. It provides a high-level view of data staleness.

• Calculation: Sum(Current Time - Last Update Timestamp) / Number of Facts.
• Use Case: Monitoring overall graph currency for compliance with Service Level Objectives (SLOs). A high MTSU indicates a potentially outdated knowledge base.
• Example: An enterprise product knowledge graph with an MTSU of 30 days suggests inventory and pricing data may be obsolete, impacting downstream e-commerce applications.

Update Frequency Distribution analyzes the histogram of time intervals between consecutive updates for entities or facts, revealing the underlying refresh patterns and identifying lagging data sources.

• Identifies Bottlenecks: Visualizes whether updates are streaming in near-real-time (e.g., seconds/minutes) or arriving in large, infrequent batches (e.g., weekly/monthly).
• Key Insight: A bimodal distribution often indicates a hybrid system where some data sources (like IoT sensors) are live, while others (like legacy ERP systems) are batch-oriented.
• Actionable Output: Guides pipeline optimization efforts by pinpointing specific entity types or source systems with suboptimal refresh cycles.

The Source Latency Score is a weighted composite metric that assigns a freshness rating to each external data source feeding the knowledge graph, based on its extraction frequency, processing delay, and ingestion reliability.

• Components: Incorporates Extraction Lag (time from source event to data pull), Transformation Lag (ETL processing time), and Ingestion Success Rate.
• Purpose: Enables data governance teams to tier sources by reliability and prioritize modernization efforts. A low score triggers alerts for source health checks.
• Implementation: Often scored on a scale (e.g., 0-100), allowing for automated dashboards that highlight the weakest links in the data supply chain.

Temporal Validity Coverage measures the proportion of time-sensitive facts in the graph that are accompanied by explicit temporal metadata, such as validFrom and validTo properties, enabling precise historical querying and current-state reasoning.

• Critical for Dynamic Domains: Essential for financial instruments, employment records, clinical diagnoses, or contractual terms where facts have specific effective dates.
• Metric Formula: (Facts with Temporal Context / Total Time-Sensitive Facts) * 100.
• Low Coverage Impact: Without temporal context, the system cannot algorithmically determine if a fact like 'CEO is Alice' is still true today or was true five years ago, crippling analytical accuracy.

Staleness-Driven Error Rate is an application-level metric that correlates data freshness directly with business logic failures or user-facing inaccuracies, providing a direct measure of the cost of outdated information.

• Measurement: Tracks incidents where a query returns a fact that is technically in the graph but is contradicted by a more recent, verified external update (a false positive due to staleness).
• Example: A customer support agent sees an account status as 'Active' in the knowledge graph, but the billing system has since terminated it. The resulting erroneous support action is logged as a staleness error.
• Value: This metric translates technical data latency into tangible business risk, justifying investment in real-time update pipelines.

Event-to-Graph Latency measures the end-to-end delay between a real-world state change (an event) and its full representation as an updated, queryable fact within the knowledge graph, including all inference and propagation steps.

• The Ultimate Freshness Metric: This is the most comprehensive measure, encompassing source latency, processing time, and the speed of graph materialization and index updates.
• Benchmarking: Critical for use cases requiring real-time awareness, such as fraud detection (where a new suspicious transaction must be linked to existing entities within milliseconds) or autonomous supply chain orchestration.
• Monitoring: Typically measured as a percentile (P95, P99) to understand tail latencies that could cause operational disruptions.

A Temporal Knowledge Graph explicitly represents the time-varying nature of facts. Each triple (subject, predicate, object) is associated with a timestamp or time interval. This structure is foundational for measuring Data Freshness, as it allows systems to query the state of the world at a specific point in time and track how entities and relationships evolve. For example, a temporal KG can store (CompanyA, hasCEO, PersonB, [2020-01-01, 2023-12-31]) and (CompanyA, hasCEO, PersonC, [2024-01-01, Now]).

Drift Detection is the automated process of monitoring a knowledge graph for significant changes in its statistical or semantic properties over time. While Data Freshness measures the age of data, drift detection identifies changes in patterns that may indicate concept drift, data pipeline errors, or evolving real-world dynamics. Key techniques include:

• Monitoring the rate of new entity/relationship creation.
• Tracking changes in the distribution of entity types or relationship predicates.
• Detecting anomalies in the update frequency of specific subgraphs.

Provenance Tracking is the capability to record and query the origin, lineage, and transformation history of every fact in a knowledge graph. It provides the audit trail necessary to validate Data Freshness. Provenance metadata typically includes:

• Source System: The original database, API, or document.
• Extraction Timestamp: When the data was ingested.
• Transformation Logic: Any cleaning, mapping, or fusion rules applied.
• Update History: A log of modifications, including who or what process made the change and when.

Factual Consistency is the property that all stated facts within a knowledge graph are logically non-contradictory. It is intrinsically linked to Data Freshness. An outdated fact can directly cause inconsistency with a newer, more accurate fact. For instance, a graph containing both (PersonX, employedAt, CompanyY) (last updated 2022) and (PersonX, employedAt, CompanyZ) (last updated 2024) represents a consistency violation unless temporal context or status flags are used. Rule-based validation engines continuously check for such conflicts.

Query Answerability measures the practical utility of a knowledge graph by evaluating its ability to return complete and accurate results for a representative set of queries. Data Freshness is a direct driver of answerability. A query like "Who is the current CFO?" is only answerable if the hasCFO relationship is fresh. This metric is often tested with a Gold Standard set of questions and expected answers, measuring the percentage of questions for which the KG returns a correct, up-to-date response.

Semantic Data Governance is the framework of policies, standards, and tools for managing the lifecycle of semantic data assets, including knowledge graphs. It establishes the operational controls for Data Freshness. Key governance components include:

• Update SLAs: Service Level Agreements defining maximum allowable latency for data from source to graph.
• Staleness Policies: Rules for flagging, deprecating, or archiving data that exceeds a freshness threshold.
• Stewardship Models: Assigning ownership for maintaining the freshness of specific entity types or data domains.