Projects

FaNDaE (Fake News Detection and Explanation using Knowledge Graphs)

Detecting fake news is of utmost importance in today’s social-media-centric society, where people often quickly form shallow or wrong opinions about complex facts. For example, the World Health Organization (WHO) observed that besides the coronavirus pandemics, there is also an infodemic caused by fake coronavirus claims. There is a plethora of domain-agnostic algorithms developed to detect fake news. However, applying these techniques to specific domains like health, where background activable knowledge plays a crucial role, can lead to unsatisfactory results. Besides, in today’s era dominated by black box data- driven AI algorithms, corroborating predictions (i.e., algorithmic verdicts about news) with some form of (visual) explanation becomes crucial. This project aims at developing a fake news detection framework called FaNDaE (Fake News Detection and Explanation using Knowledge Graphs) to tackle these challenges. FaNDaE will be designed as a modular architecture composed of three layers. First, a customizable knowledge substrate combines textual information from news in a target domain with external domain- specific knowledge available from knowledge graphs (KGs). Second, a learning layer builds upon the previous layer and intertwines data-driven AI learning models with knowledge-driven inference rules to make predictions. Third, an explanation layer leverages the previous layers to generate both plausibility scores of the news articles and explanations for the scores. FaNDaE’s data-level explanations can be abstracted into complex patterns that enrich KGs with new domain knowledge. We plan to test FaNDaE in both general and specific domains.

SaGRL (Semantic-Aware Graph Representation Learning)

Knowledge Graph (KGs) are becoming an essential ingredient as a bridge between logic/symbolic knowledge and machine learning. Graph Neural Networks (GNNs) are the state-of-the-art for KG representation learning. However, we recognize that existing GNN methods primarily leverage the graph structure while the semantics in the KGs is almost discarded; indeed, neighbors are determined solely by relying on topological information, thus overlooking the pre-existing feature-based or schema-based node similarity. Moreover, these strategies mainly adopt a black-box approach and lack the interpretability of model predictions.

The goal of this project is to design, implement, and experimentally evaluate novel learning models for GNNs that, when incorporating the KG semantics, improve the model accuracy, scalability, and the interpretability.

We plan research work along the following lines:

-Semantic-based message passing focused on designing novel notions of node neighbors, and message passing strategies used by GNN to aggregate/refine node representations. Instead of recursively aggregating node neighbors, the idea is to leverage node/edge types, and the KG schema to (pre)build contextualized subgraphs and perform the aggregation over them, thus avoiding expensive recursive procedures. This will also pave the way toward semantic-driven node sampling strategies that will improve the scalability of GNNs.

-Model prediction explanations focused on generating model prediction explanations by intertwing schema information into the learning model to identify data in the KG that were activated during the generation of a model prediction. Explanations can have different levels of granularity, spanning from simple paths or fixed substructures (e.g., triangles) to complex subgraphs, thus allowing to obtain different explanation details and witnessing the formal reasoning process to some extent.