Forecasting topic trends of blockchain utilizing topic modeling and deep learning-based time-series prediction on different document types

Abstract

Topic trends in rapidly evolving domains like blockchain are dynamic and pose prediction challenges. To address this, we propose a novel framework that integrates topic modeling, clustering, and time-series deep learning models. These models include both non-graph-based and graph-based approaches. Blockchain-related documents of three types—academic papers, patents, and news articles—are collected and preprocessed. Random and topic subgraphs are constructed as inputs for model training and forecasting across various time epochs. The four models (LSTM, GRU, AGCRN, and A3T-GCN) are trained on random subgraphs, and the trained models forecast topic trends using topic subgraphs. We also analyze the distinctive characteristics of each document type and investigate the causal relationships between them. The results indicate that non-graph-based models, such as LSTM, perform better on periodic data like academic papers, whereas graph-based models, such as AGCRN and A3T-GCN, excel at capturing non-periodic patterns in patents and news articles. Our framework demonstrates robust performance, offering a versatile tool for blockchain-related trend analysis and forecasting. The code and environments are available at https://github.com/textmining-org/topic-forecasting.