Francis Nji Successfully Defends his PhD Dissertation

Francis successfully defended on Thursday, July 16, 2026

Francis successfully defended his dissertation on Thursday, July 16, 2026. Francis's dedicated contributions have made him an important member of the iHARP and UMBC communities. This achievement is a true testament to his scholarly rigor and hard work.

Congratulations, Francis, on this incredible milestone! We have been honored to witness your growth and look forward to your future successes. The iHARP community wishes you the very best as you embark on the next chapter of your journey!

_____________________________

Dissertation Title: Accurate Clustering of High-Dimensional Multivariate Spatiotemporal Data

Committee

• Dr. Jianwu Wang, Chair/Advisor, UMBC
• Dr. Vandana P. Janeja, Co-Advisor, UMBC
• Dr. James Foulds, UMBC
• Dr. Yiqun Xie,  University of Maryland, College Park.
• Dr. Aneesh Subramanian, University of Colorado Boulder

Abstract

The rapid growth of multidimensional multivariate spatiotemporal data from Earth observation systems, sensor networks, climate reanalysis products, and large- scale monitoring platforms has created unprecedented opportunities for understanding complex natural and human-driven processes. These datasets simultaneously vary across space, time, and multiple variables, enabling the discovery of hidden patterns, climate regimes, anomalies, and evolving system behaviors. However, clustering such data remains challenging due to high dimensionality, nonlinear interactions, spatial autocorrelation, long-range temporal dependencies, noise, missing values, nonstationarity, and multiscale structures. Traditional clustering approaches
primarily rely on distance-based or correlation-based similarities and often fail to capture the complex spatial, temporal, and causal mechanisms underlying these systems.

To address these challenges, this dissertation develops three novel deep clustering frameworks for multivariate spatiotemporal data. First, Hybrid Ensemble Deep Graph Temporal Clustering (HEDGTC) integrates homogeneous and heterogeneous ensemble clustering with a dual-consensus strategy and a Deep Graph Attention Autoencoder to improve clustering robustness, stability, and accuracy. By combining object co-occurrence consensus and non-negative matrix factorization consensus, HEDGTC effectively reduces noise and misclassification while preserving temporal and relational structures. 

Building on the foundation established by HEDGTC, we introduce the Bi-directional Temporal Graph Attention Transformer (B-TGAT), an end-to-end deep clustering framework for multivariate spatiotemporal data. The proposed architecture combines a ConvLSTM-based U-Net autoencoder for learning rich spatiotemporal representations, graph attention mechanisms for modeling spatial relationships, and a bi-directional transformer for capturing long-range temporal dependencies in both forward and backward directions. This integrated design enables B-TGAT to effectively learn complex spatial and temporal patterns from climate data. By jointly modeling local spatial interactions and long-term temporal dynamics, B-TGAT un- covers meaningful climate regimes, large-scale teleconnections, abrupt climate transitions, and extreme weather events. The learned latent representations improve cluster compactness, stability, and separation, leading to more accurate, robust, and interpretable clustering of multivariate spatiotemporal climate datasets. This unified framework provides a powerful solution for discovering hidden climate patterns that are difficult to identify using conventional clustering approaches. 

Recognizing that correlation-based representations alone may not reveal the true drivers of system evolution, this dissertation further proposes Causal Adversarial Subspace Clustering (CASC), a causality-aided deep clustering framework that transforms spatiotemporal regime discovery from a correlation-driven process into a causal-temporal learning paradigm. CASC integrates a U-Net-inspired adversarial autoencoder, stacked FAConvLSTM layers, graph attention-based self-expressive learning, and two novel objectives: Causal Subspace Preservation (CSP) Loss and Dynamic Temporal Subspace Evolution (DTSE) Loss. These mechanisms enable the model to discover clusters that are not only geometrically compact but also causally meaningful and temporally coherent. A Subspace-Aware Energy-Based TemporalDiscriminator further enhances clustering stability, interpretability, and robustness by evaluating latent representations according to cluster-specific subspace structures.
Extensive experiments on multiple real-world climate and environmental datasets demonstrate  that the proposed frameworks consistently outperform state-of-the-art traditional and deep clustering approaches across clustering quality, robustness, stability, and interpretability metrics. Collectively, the proposed methods establish a unified framework for learning hierarchical spatial, temporal, and causal representations from complex spatiotemporal data. The resulting models provide powerful tools for climate regime discovery, environmental monitoring, anomaly detection, and scientific understanding of dynamic Earth systems, while advancing the state of the art in deep unsupervised spatiotemporal clustering.