Ontology-supported AI Model and Dataset Management

AI summary

Overview

Research area: AI asset management, semantic web technologies, and ontology-based metadata for machine learning models and datasets, with an industrial (automotive supply chain) framing.

Technical level: Intermediate. No deep machine learning mathematics is required, but familiarity with machine learning workflows, metadata, and basic ontology/linked-data concepts helps.

Scope: The paper introduces AIMDEP, a platform for registering, searching, deploying, and exporting AI models and datasets, built around AIMDEO, an OWL ontology that supplies explicit, shared semantics for their metadata, demonstrated through a real-time-critical systems use case.

What This Paper Is About

Organizations produce large numbers of AI models and datasets, but their descriptions are often semi-structured natural language text or inconsistent key-value metadata, which makes assets hard to discover, compare, and trust. This is especially problematic in an industrial supply chain such as automotive, where assets move between companies whose metadata formats differ or conflict. The paper asks what is required to develop and use an AI model collaboratively in an industrial context, and answers with a platform plus an ontology that keeps the meaning of metadata unambiguous when assets are exchanged.

Key Contributions

1. AIMDEP (AI Model and Dataset Exchange Platform): a platform for collaborative AI development and use, providing a central registry, REST API, web UI, semantic search, asset visualization, online deployment, and metadata export, all organized around a common ontology-backed asset description.
2. AIMDEO (AI Model and Dataset Exchange Ontology): an OWL ontology describing AI models and datasets, covering provenance, framework, task and sub-task, inputs, outputs, parameters, data sources, training/test splits, data points, quality criteria, and metric scores.
3. Four collaboration scenarios that guided the ontology design: Internal Use, Shared Data, Shared Models, and Model Services.
4. A use case evaluation in the context of real-time critical systems, in which a hardware developer, an AI expert, and a software developer exchange a dataset and a memory access time prediction model through the platform.

Main Findings

• Platform architecture: AIMDEP uses a client-server architecture with the server and Web UI implemented in Django; custom clients interact through a REST API. Because the platform leverages Django's Model-View-Template architecture, updates to the ontology can apply to the server's model component without disrupting overall functionality. Communication between server and clients is encrypted, and access to assets is restricted to privileged users by a server-side policy.
• Supported asset formats: datasets in CSV, Excel, JSON, and Parquet; models exported by Scikit-Learn, TensorFlow, and PyTorch. Custom datasets and models can be registered without specifying a framework, but then the platform cannot visualize the dataset or deploy the model.
• Registration workflow: assets pass through a register process that attempts to identify the required back-end, defines input and output features semi-automatically (detected by the platform, verified and extended by the user), and stores each feature as a Parameter based on AIMDEO. The asset is then stored in the AIMDEO-based central database.
• Asset operations: interactive dataset visualization as statistical tables, plots, and feature importance analyses (the latter only for numerical datasets), rendered server-side by Plotly and optionally limited to a random subset for very large datasets; online model deployment with inference through a REST API or a Gradio-based graphical interface generated from the model's task, sub-task, and input/output features; search and download powered by OpenSearch.
• Metadata export: downloaded assets are accompanied by metadata exported as micro-ontologies containing attributes and their values, as illustrated by an example of an AI model and its corresponding dataset metadata referring to AIMDEO.
• Ontology scale: AIMDEO is expressed as an OWL ontology with an axiom count of 414, logical axioms count of 178, declaration axioms count of 105, 47 classes, 29 object properties, 17 data properties, 5 individuals, and 18 annotation properties, with DL expressivity SHOIQ(D).
• Comparison with related work: an earlier assessment of machine learning asset management tools found that only MLflow supports asset registry and exchange without requiring data sharing with third parties, but it does not support ontologies for metadata description. MLEM offers a model registry that is very limited and does not collect metadata. Hugging Face hosts models and datasets and offers a deployment interface, but descriptions are semi-structured.
• Comparison with ITO and EMMM: the ITO is primarily aimed at describing AI tasks and at studying scientific research, whereas AIMDEO targets the exchange of AI models and datasets and additionally contains concepts for provenance and kind of AI model, model and dataset parameters, and model evaluation metrics; both support intended tasks and sub-tasks. The EMMM metamodel covers central concepts such as models, parameters, and dependencies and supports version control structures, but captures element metadata as arbitrary key-value mappings lacking a common format and explicit semantics.
• Use case result: in the memory access time prediction scenario, the exported model metadata records the model "RF Instruction Cache-Line Access Classifier" (version 1.0, framework scikit-learn), an average precision metric score of 0.9794 (dimensionless), a Replacement-Strategy parameter of LRU, a Cache-Size parameter of 2048 Byte, an input "set" of type uint8, an output "miss" of type bool, and a linked dataset named "Complex Programs Instruction Cache-Lines Set Sorted" (version 3.0).
• Reported benefits: semi-automatic feature recognition reduces susceptibility to errors such as forgetting features; keyword search reduces the effort of finding suitable assets; linked dataset descriptions let end users check whether their data matches the training data; and platform-side execution lets end users try a model without setting up a runtime environment.

Methodology in Plain English

The authors first studied how AI assets are actually moved between parties. They analyzed existing platforms (MLflow, H2O, Ray, MLEM, Hugging Face) and existing metadata efforts (ITO, EMMM) and found that metadata either lacks explicit semantics or is not interoperable across organizations. From that analysis they defined four collaboration scenarios, ranging from keeping everything in-house to offering a model only as a remote service, and derived from those scenarios the concepts that any exchange description must capture. Those concepts were then formalized as an OWL ontology (AIMDEO), with classes and properties for general metadata, dataset metadata, and model characterization. Around the ontology they built a platform (AIMDEP) whose database model is generated from the ontology, so that ontology changes propagate into the server without breaking functionality. Finally, they exercised the whole chain on a concrete problem: predicting memory access time for real-time critical software, with three roles (hardware developer, AI expert, software developer) exchanging a dataset and a model through the platform.

Why This Matters

Impact on research: the work treats metadata as a first-class, semantically explicit artifact rather than free text or ad hoc key-value pairs, which makes AI asset comparisons and reuse reproducible across organizational boundaries. It also positions ontology design as a prerequisite for trustworthy model exchange, not an afterthought.

Real-world applications:

• Automotive supply chain exchange, where AI models and related assets move between tiers of different companies with differing metadata formats.
• Safety-critical and real-time systems engineering, where predicting software timing depends on accurately described hardware characteristics such as cache replacement strategy and cache size.
• Hardware procurement decisions, where an end user evaluates whether a described memory configuration justifies purchasing test hardware.
• Internal tool integration, where a REST API and ontology-based export let existing development tools consume and publish asset metadata.

Industry relevance: the paper targets the documented gap between where AI models are created and where they are used, and the absence of structured metadata with explicit semantics. By making registered assets searchable, analyzable, executable, and exportable with machine-readable descriptions, the approach lowers the barrier for organizations to reuse models without sharing raw data, which is a common constraint in commercial supply chains. The work was funded by BMWi within the project progressivKI (grant number 19A21006M).

Future Directions

• Extending EMMM to create a degree of compatibility, potentially storing AIMDEO metadata inside the metamodel's generic key-value pairs associated with resource types.
• Investigating further interoperability between AIMDEO and ITO.
• Determining how the platform and the ontology should be used separately versus together, given that both can be applied independently.
• Following through on the platform's premise that ontology updates can flow into the server's model component without disruption, which the paper asserts as a design property rather than as a measured result.

Target Audience

This paper benefits AI platform and infrastructure engineers, data and model governance specialists, semantic web and ontology practitioners, and industrial engineering teams in domains such as automotive or embedded systems that must move AI models and datasets across organizational boundaries. It is also relevant to researchers working on AI asset metadata standards, model registries, and reproducibility, and to readers comparing ontology-based approaches against conventional key-value metadata management.

Authors’ abstract

Recently, there has been a great deal of research into improving AI methods and their application. The main focus is on tracking progress, enabling transparent comparisons, and fostering a more profound understanding of AI. In that process, different organizations generate and use plenty of assets that need to be tracked, traced and managed. Moreover, it is important to discover assets relevant for the task at hand. This paper presents research aiming to contribute to answering the question of what is required to exchange and manage AI models and related assets effectively without semantic gaps in an industrial context. We introduce a platform for AI model exchange, which facilitates the usage, exchange, and analysis of AI models and datasets. The platform incorporates an ontology that can foster a more profound common understanding of what is required in these tasks and help tackle the issues mentioned above. Finally, we elucidate the utility of the platform through the illustration of a use case in the context of real-time critical systems.