Predictive modeling using heterogeneous data sources

Abstract

The increasing availability of heterogeneous data has transformed the scope and capabilities of predictive modeling across scientific, social, and environmental domains. Modern predictive systems no longer rely on a single structured dataset collected under controlled conditions. Instead, relevant information is distributed across multiple modalities, platforms, temporal scales, and relational structures. Social media interactions generate large volumes of unstructured textual and network data that reflect human behavior and collective attention. Environmental monitoring systems produce high-dimensional geospatial and temporal measurements describing weather patterns, soil properties, and ecological conditions. Scientific repositories and digital libraries contain rapidly growing collections of publications, metadata, and domain-specific representations that require automated extraction and organization. These heterogeneous sources provide substantial predictive potential, but they also introduce significant methodological challenges related to data integration, representation learning, scalability, and generalization. Machine learning has emerged as a powerful framework for extracting predictive signals from such complex data environments. However, conventional approaches often assume homogeneous feature spaces, abundant labeled data, and static distributions. These assumptions rarely hold in real-world predictive settings. Data collected from different sources frequently exhibit variations in semantics, structure, quality, and temporal dynamics. Labels are often scarce, incomplete, or expensive to obtain. Moreover, predictive systems deployed in operational settings must generalize across domains, evolving environments, and previously unseen data distributions. This dissertation investigates predictive modeling under heterogeneous data conditions through a collection of studies spanning social media analytics, environmental hazard prediction, graph representation learning, and scientific knowledge extraction. Although these applications originate from different domains, they share a common objective: transforming fragmented and structurally diverse data into meaningful predictive representations that support inference and decision-making. The first component of this work focuses on predictive modeling in social media environments. Online platforms contain large volumes of user-generated content that can support the analysis of public discourse, user engagement, and community behavior. However, differences in platform characteristics and annotation availability complicate cross-platform prediction tasks. One study addresses this challenge through a transfer learning framework that predicts topic relevance in Reddit using annotated Twitter data and weak supervision derived from Wikipedia. By incorporating domain-adversarial adaptation, the framework improves generalization across heterogeneous social media platforms. A second study models Reddit user interaction and experience using heterogeneous graph representations that integrate user behavior, subreddit relationships, and semantic textual embeddings. These approaches demonstrate the importance of representation learning and relational modeling for predictive analysis in socially generated data. The second component addresses predictive modeling for environmental hazards, with a particular emphasis on landslide susceptibility prediction. Environmental prediction problems require the integration of multimodal geospatial and temporal information derived from weather systems, vegetation patterns, topographical characteristics, and soil conditions. One study develops a machine learning framework that combines heterogeneous environmental variables to estimate landslide susceptibility while addressing challenges associated with class imbalance and high-dimensional data. A subsequent study extends this framework by incorporating Physics-Informed Neural Networks to model soil moisture dynamics and rainfall-induced landslide behavior. By integrating physical constraints with data-driven learning, this work demonstrates how domain knowledge can improve predictive reliability in complex environmental systems. The dissertation also includes two ongoing projects that further extend predictive modeling capabilities for heterogeneous data environments. The first introduces TML-Mixer, a framework for temporal multilayer graph representation learning that captures evolving dependencies across interconnected graph layers. This work addresses the growing need for predictive models capable of representing dynamic relational systems over time. The second project focuses on scientific knowledge extraction through the development of a containerized large language model framework for identifying publications containing phylogenetic tree representations. This work explores the use of large language models and metadata-driven classification for scalable scientific data curation. Collectively, these studies contribute to a broader research agenda centered on predictive modeling with heterogeneous data sources. Across domains, the dissertation investigates how machine learning models can integrate multimodal information, learn robust representations, adapt across distributions, and incorporate structural or scientific knowledge to improve predictive performance in complex real-world environments.