Explainable Artificial Intelligence over Knowledge Graph: from Reinforcement Learning to Generative Modeling

Course Aims and Scope

Intelligent systems shape many daily decisions, but their reasoning is often unclear. Among them, Recommender Systems personalize content but need transparency to build trust. This course explores how Knowledge Graphs (KGs) can enhance explainability by structuring knowledge and supporting reasoning. Students will integrate KGs with Reinforcement Learning and Generative Models , using tools like HuggingFace and Hopwise in hands-on sessions to build explainable recommender solutions.

What you will learn

Enhance model explainability by integrating Knowledge Graphs into intelligent systems
Use Large Language Models and Reinforcement Learning to provide explainable recommendations
Understand how Transformers and Large Language Models work, including their core architecture and mechanisms
Train and evaluate recommender systems with HuggingFace and Hopwise libraries

Who is it for

This course is designed for graduate students and advanced undergraduates in computer science, artificial intelligence, or data science who are interested in building intelligent systems that are not only effective but also transparent and explainable.

What you need

No prior experience with Knowledge Graphs, Reinforcement Learning, or Large Language Models is required. Just bring your curiosity and a laptop — everything else will be provided during the course through interactive notebooks. 😊

Program

Stay tuned—more details will be shared shortly! 😊

Date	Day	Sessions
21/07	Day 1: Foundation of Explainable AI over Knowledge Graph	Welcome and Course Overview Introduction to the course structure, objectives, and key topics. Core Concepts: Explainability and Knowledge Grounding Explore the foundational principles of explainable AI, focusing on how structured knowledge enables transparent and explainable recommendations. This session provides the theoretical basis for Graph-based Recommender Systems. Hands-on Session: Introduction to recommendation pipeline with hopwise Overview of the Hopwise library and its application in building recommendation models.
22/07	Day2: Explainable Recommendation Pipeline over Knowledge Graph	Welcome & Recap of Day 1 Explainable Recommendation Pipeline: from data processing to model evaluation End-to-end pipeline: from integrating knowledge graph data to training and evaluating explainable models. Hands-on Session: TransE for Knowledge Representation Training the TransE model and exploring its effectiveness in learning knowledge embeddings.
23/07	Day 3: Explainable Reinforcement Learning for Knowledge Graph Reasoning	Welcome & Recap of Day 2 Theoretical Foundation of Reinfocerment Learning Introduction to key concepts in reinforcement learning: agents, environments, states, actions, and rewards. Examination of how reinforcement learning guides path exploration to improve recommendation quality and explainability. Theoretical Foundation of PGPR model Overview of Policy-Guided Path Reasoning (PGPR) and its role in multi-hop reasoning over knowledge graphs. Hands-on Session: PGPR Training & Evaluation Training PGPR using the Hopwise library on knowledge graph datasets.
24/07	Day 4: Foundations of Generative Models	Welcome & Recap of Day 3 Introduction to Transformer Architecture Foundational overview of the Transformer architecture as introduced in “Attention is All You Need,” covering key components such as self-attention, positional encoding, and multi-head attention to build a solid understanding of modern generative models. Hands-on Session: Hugging Face Transformers Overview of the library’s structure and key components. Deep dive into tokenization, with a comparison of different tokenizers and their impact on model input representation. Exploration of the AutoModelForSeq2SeqLM architecture, showcasing how to load and utilize pre-trained models for sequence-to-sequence language modeling tasks.
25/07	Day 5: Explainable Generative Models over Knowledge Graph	Welcome & Recap of Day 4 Language Models for Explainable Recommendations Exploration of cutting-edge approaches that integrate knowledge graphs with language models to improve explainability, with a focus on PLM and PEARLM models. Hands-on Session: PEARLM Implementation and Usage Using path-languale modeling to generate top-k recommendations along with natural language explanations. Final Discussion & Conclusions Reflection and future directions.

Speakers

Francesca Maridina Malloci

Dr. Francesca Maridina Malloci
Assistant Professor, University of Cagliari
Department of Mathematics and Computer Science
(Italy)

Francesca Maridina Malloci is Assistant Professor at the Department of Mathematics and Computer Science of the University of Cagliari (Italy). Her research focuses on responsible artificial intelligence, with attention to decision-making systems, such as recommender systems, for multi-stakeholder contexts. She has co-authored papers in international journals and conferences. Francesca has chaired workshops and tutorials, including those on knowledge discovery at ACM UMAP,ECML-PKDD, and AIED, on explainable AI at ECIR, and on bias and fairness at SIGIR. She also serves as Associate Editor for Neural Processing Letters (Springer) and Information Processing in Agriculture (Elsevier).

Venue

Computer Science Department
Boise State University
1910 University Drive Boise,
ID 83725-2055 USA

Room: CPP-221
Time: 9:00 AM – 11:00 AM

Contacts

For general enquiries on the course, please send an email to francescam.malloci@unica.it