Abstract
Manual configuration of machine learning (ML) workflows is a labor-intensive process that often lacks reusability and scalability. This paper introduces an ontology-based framework designed to automate the generation of adaptive data processing workflows. By leveraging Semantic Web technologies and Semantic Web Rule Language (SWRL), the system dynamically infers optimal task sequences and model parameters based on data characteristics and user requirements. Evaluation on the New York Citi-bike dataset demonstrates that the automated approach matches the performance of expert manual configurations while significantly reducing development overhead.
Introduction
Modern data science requires "adaptive" systems capable of selecting appropriate models and parameters as data evolves. Traditional workflow management tools, such as Apache Oozie, rely on rigid, manual configurations. This approach suffers from high overhead, limited reusability, and increasing complexity as the library of available algorithms grows. To address these challenges, we propose a "smart" mechanism that uses logical reasoning to generate workflows dynamically.
Framework Design
The core of the system is a semantic ontology developed in Protégé. It is structured around three primary pillars: ML_Process (the procedural steps), ML_Process_Service (the specific tools and algorithms), and Parameters (the variables required for model execution).
The ontology defines 14 object properties and 7 data properties to capture the implicit relationships between different ML tasks. For instance, properties like hasC or hasGamma guide the selection of Support Vector Machine (SVM) parameters. To handle complex logic beyond standard OWL capabilities, we utilize SWRL rules. These rules act as the "reasoning engine," dictating workflow transitions based on goals like high accuracy or specific data formats.
Case Study and Results
We evaluated the framework using the New York Citi-bike parking quantity dataset, aiming to predict bike rental and return patterns across various station clusters. The system successfully instantiated a specific case for "citibike" and used its reasoner to generate a complete workflow.
The resulting sequence—data_preprocess → feature_preprocess → modeling → evaluation—was generated automatically. The system further selected specific models (e.g., SVM or GBRT) and their corresponding parameters. Crucially, the predictive performance of this automatically generated workflow was identical to that of an "optimal" manual configuration, validating the framework's effectiveness.
Conclusion
By transforming implicit ML task relationships into explicit, machine-readable ontologies, we can automate the design of complex workflows. While future work is needed to address the scalability of reasoning and the complexity of rule creation, this approach represents a significant step toward reducing the manual burden in data science.
Summary updated