TLDR: This research paper introduces a novel method for integrating activity predictions into ontology-structured knowledge graphs. By leveraging frameworks like BFO and CCO, it demonstrates how historical data (e.g., fishing vessel movements) can be organized and queried to build Markov chain models for predicting future states. The paper critiques existing probability models and proposes an alternative where probabilities are understood as being about “process profiles” or “patterns of life,” allowing for auditable and justifiable predictions.
Predicting future events, from tomorrow’s weather to economic shifts, is a fundamental human need for planning and resource allocation. Traditionally, such predictions relied on data collected specifically for that purpose. However, a new approach leverages the power of knowledge graphs to make these predictions more efficient and auditable.
Researchers have proposed a method that integrates activity predictions directly into knowledge graphs, using a structured semantic framework. This approach allows data, once collected and organized, to be used for various analyses, including forecasting future events, without needing significant restructuring. This is a significant advantage over older database models.
The core of this method involves structuring information using established ontological frameworks like the Basic Formal Ontology (BFO) and Common Core Ontologies (CCO). These ontologies provide a logical vocabulary that gives meaning and structure to the data within the knowledge graph, enabling the integration of information from diverse sources.
To illustrate their methodology, the researchers used a hypothetical case: tracking the movements of a fishing vessel over 100 days. The vessel could be at one of three fixed locations, either staying put or moving to another location daily. This movement data was then organized within a knowledge graph, defining entities like the fishing vessel itself, its daily trips (processes), and its locations at specific “spatiotemporal instants.” Specialized queries, known as SPARQL queries, were then used to extract the vessel’s location history and sequences of movements.
Once the historical movement data is extracted, it’s used to build Markov chain models. A First-Order Markov Chain predicts the next location based solely on the vessel’s current location. For example, if the vessel is at location 3 today, the model can tell you the probability it will be at location 2 tomorrow. This is represented in a “transition matrix” where rows show the present location and columns show the next possible locations.
For more complex patterns, a Second-Order Markov Chain is employed. This model considers both the present and the immediate past location to predict the next state. While offering greater accuracy for movements influenced by recent history, it requires more data and results in a larger, more complex prediction matrix. The paper provides a detailed explanation of how these probabilities are calculated and presented in matrices.
A crucial aspect of this research is how these calculated probabilities are integrated back into the knowledge graph. The researchers critique existing ontological models of probability, particularly in CCO, which they argue incorrectly treat probabilities as being about future processes. Instead, they propose that probabilities are about “process profiles,” specifically “pattern process profiles” or “patterns of life.”
A “pattern of life” is defined as an observable regularity in the change of attributes of entities participating in a process. In the fishing vessel example, the vessel’s consistent pattern of location changes over 100 days establishes its “pattern of life.” The probabilities are then understood as measurements of the potential for certain dispositions (like moving from location A to B) to be realized, with these patterns of life serving as measurable proxies for these potentials. This new model allows for a more consistent and justifiable way to represent and use probabilities for future predictions within the knowledge graph.
The paper also outlines future research directions, including exploring Continuous-Time Markov Chains for more realistic, continuous movements, and Time-Inhomogeneous Markov Chains to account for external factors like weekdays vs. weekends. They also discuss methods for handling non-discrete locations (like geo-coordinates) by clustering them into defined areas. This work paves the way for generalizing this ontological model to other types of probabilities, such as Bayesian probabilities, and for linking probabilities to dispositions in BFO-compliant ontologies.
Also Read:
- A New Framework for Personalized AI: Dialogical Large Language Models
- Advancing Probabilistic Inference in Ordered Data
In conclusion, this research demonstrates a powerful methodology for using ontology-structured knowledge graphs to generate and integrate predictions about future events. By providing a standardized, machine-understandable representation of data and analytics, it offers an auditable trail for predictive models, moving away from “blackbox” algorithms. This is particularly valuable for decision-making processes that require transparent and justifiable insights. For a deeper dive into the technical details, you can read the full research paper here.
