Early identification of experts in online question answering communities using neural networks

Abstract

Community Question Answering (CQA) platforms designed to enable users to ask questions and seek answers from the community have gained an increased popularity over the years. Prior research has shown that these communities thrive largely due to a small subset of expert users who provide comprehensive and often accurate answers. Identification of such experts in community question answering (CQA) platforms is an important, yet challenging task. Many prior studies employ variants of classical machine learning and link analysis techniques to tackle this challenging task. However, studies leveraging novel advancements in neural networks to tackle this challenging problem are quite sparse. This thesis work models community question answering forums as heterogeneous graphs and leverages novel Graph Neural Networks to learn rich representations of users. Next, this thesis work adapts a neural network model to leverage learned user representations and embeddings of questions extracted from a pre-trained large language model (LLM) to identify expert users who are best suited to answer a given question.Through experiments conducted on a real-world CQA dataset, the author demonstrates the effectiveness of the proposed approach. Furthermore, existing studies have not sufficiently explored the effects of considering different degrees of relevance among neighbors in CQA interaction graph, and largely assume all neighbors to be of equal relevance. This thesis work explores this under-explored area, and experiments with Graph Attention Networks (GAT) which makes use of attention mechanisms to weigh the messages passed from neighboring nodes. Furthermore, this thesis work proposes a novel edge-weighting scheme based on temporal decay to explicitly assign relevance scores to interactions among nodes. Through experiments, the author demonstrates how capturing different degrees of relevance among neighbors helps improve the identification of community experts.