Explainable AI framework through Multi-Context Multi-Dimensional Graph Neural Network
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Subject
Abstract
In this research, we explored the multifaceted realm of digital communication, emphasizing social media channels such as Twitter and Reddit, complemented by conventional data-gathering techniques like focus group discussions and structured digital sources like ancient Greek literature. By fusing these data types, we accessed a vast array of user-generated content, amalgamating sentiments, themes, and contextual cues deeply rooted within online communities. Advanced computational methods, including sentiment analysis and topic modeling, were employed to interpret this vast data ocean. However, traditional algorithms struggled to encapsulate the intricate interconnections in social media ecosystems, focus group interactions, and classic literature, primarily due to their inherent linearity. We adopted Graph Neural Networks (GNNs), a sophisticated machine learning paradigm adept at handling graph-based data to surmount these obstacles. GNNs displayed a flair for harnessing the relational dynamics inherent in complex systems such as social media, focus groups, and literature explaining the symbiosis between sentiment, context, and digital community. These relationships were converted into dense vector representations, capturing the nuanced interrelations of the graph constituents. This venture provided profound insights into the fabric of digital communication, steering in a groundbreaking method to identify community structures within digital platforms. The efficacy of this innovative approach was evident, presenting enhanced and more organic methods to model the intricate interplay within these networks. Expanding on this foundational work, our objective became harnessing GNNs to extract richer contextual portrayals of user interactions in the digital space, aiming for transformative effects on digital communication. Capturing the multiple contexts from the digital data, we can provide more insights into the GNN models during learning. Muti-context embedding can explain why a GNN model worked best for a data set. Confronted with the intricacy of disparate and unstructured digital data, we pioneered the Explainable AI framework via Multi-Context Multi-Dimensional Graph Neural Networks (MMGNN). This holistic solution facilitated the preservation of conversational nuance, streamlined dialogues, and pinpointed shared views or discord in conversations. The clarity our graph-based representation provides ensures a more trusted data comprehension. In our research, we engaged deeply with data from focus group meetings, ancient Greek dramas, and Reddit to unpack the potential of MMGNN. Utilizing MMGNN, we outstripped benchmark performances in classification tasks compared to the TextGCN and BertGCN. This methodology enabled superior management of diverse and unstructured data, paving the way for immediate MMGNN applications.
Table of Contents
Introduction -- Social contextual influences on healthy eating -- Contextual word embeddings in healthy dieting -- Interpretation of sentiment analysis in digital literature -- Sentiment analysis with human-in-the-loop -- Understanding emotion and topic patterns in digital data -- A graph-based system for conversation modeling -- Multi-context multi-dimensional graph neural network -- Conclusion and future work
DOI
PubMed ID
Degree
Ph.D. (Doctor of Philosophy)