Date of Award

Summer 2025

Document Type

Open Access Dissertation

Department

Computer Science and Engineering

First Advisor

Amit Sheth

Abstract

Can I eat this food or not? Is this food suitable for diabetes and why? Which AI pipeline is best suited for a given task and dataset? How should an end-to-end pipeline be constructed? These questions differ from factual question-answering tasks. Recipes and AI pipelines are processes consisting of several entities interacting with each other. A recipe consists of ingredients, cooking methods, and their interactions, while an AI pipeline includes datasets, preprocessing techniques, models, hyperparameters, tasks, and results. Each entity must be analyzed individually, and collective inferencing is performed to derive the final decision. This decision-making process, known as compositional reasoning, is essential for assessing the relevance of processes to ensure appropriate recommendations.

Existing process recommendation approaches in the literature focus on ranking relevant processes to generate appropriate recommendations. However, these methods typically assume the availability of structured representations of processes. In reality, many process-oriented data, such as recipes and AI pipelines, lack structured representations, making it challenging to apply these algorithms directly. Currently, such data are treated as lengthy unstructured documents, often leading to information dilution due to the presence of noise in natural language text. Moreover, existing models struggle to capture essential interaction information, particularly when dealing with long input sequences.

Additionally, natural language descriptions of recipes and AI pipelines do not inherently provide the necessary domain knowledge. For example, a recipe description does not inherently specify that potatoes are classified as healthy carbohydrates with a high glycemic index. Similarly, the name PubMedQA does not explicitly convey that it is a question-answering dataset within the clinical domain. External domain knowledge is essential for contextual analysis. Since no structured representation currently exists for recipes and AI pipelines, mechanisms for entity extraction, knowledge integration, and reasoning are required to enable meaningful recommendations. Given the high-stakes nature of these domains, models must not only generate recommendations but also provide explanations that can be attributed to trusted sources.

While neural network models excel in pattern recognition, they struggle with compositional reasoning. Large Language Models (LLMs), despite their strong generalization capabilities, operate in vast embedding spaces, making it difficult to retrieve and apply relevant domain knowledge effectively. Addressing these challenges requires a structured, knowledge-driven approach that integrates reasoning and explainability into process recommendation systems.

To address these challenges, this work introduces a neurosymbolic explainable process recommendation framework leveraging Dynamic Multimodal Process Knowledge Graphs (DMPKGs). DMPKGs provide structured representations of processes and their entities, grounded in multi-contextual knowledge for high-order reasoning, explainability, and traceability. By decomposing recipes and AI pipelines into modular components, DMPKGs allow individual entity inference while capturing interactions and dependencies for compositional decision-making. The explanations can be provided for rationale behind explanations and can attributed to reliable sources, ensuring trustworthiness. The representation in DMPKGs elevated higher-order abstractive concepts enable symbolic reasoning to produce explainable recommendations, ensuring reliability and contextual awareness. DMPKGs are particularly well suited for dynamic procedural tasks, where components evolve over time. The modular structure allows for continuous updates without modifying the entire schema. Unlike traditional knowledge graphs, which store static relationships, Process Knowledge Graphs (PKGs) capture entity semantics within processes. Additionally, DMPKGs can store multimodal data, such as text, images, and embeddings, enabling efficient retrieval and ranking for recommendations.To demonstrate this approach, two use cases are explored: (1) recipe suitability analysis for dietary recommendations and (2) AI pipeline recommendation for selecting optimal configurations. DMPKGs bridge data-driven learning with symbolic reasoning, providing a robust, explainable, and dynamic approach to process recommendation that ensures accuracy, adaptability, and explainability.

Rights

© 2025, Revathy Venkataramanan

Download

Share

COinS