11/03/2026
By Syed Mahmood

The Kennedy College of Sciences, Miner School of Computer & Information Sciences, invites you to attend a doctoral dissertation defense by Syed Abrab Mahmood, titled "Integrating Topological Data Analysis and Deep Sequence Models for Robust Multivariate Time Series Representation Learning: Applications in Space Weather Forecasting."

Details

  • Date: Thursday, Nov. 19, 2026
  • Time: 2 p.m. EST
  • Location:Zoom meeting

Committee Members

  • Dr. Ruizhe Ma (advisor), assistant professor, Miner School of Computer & Information Sciences, University of Massachusetts Lowell
  • Dr. Hong Yu, professor, Miner School of Computer & Information Sciences, director, CHORDS, University of Massachusetts Lowell
  • Dr. Sashank Narain, assistant professor, Miner School of Computer & Information Sciences, University of Massachusetts Lowell
  • Dr. Timothy Cook, Kennedy College of Sciences, Department of Physics & Applied Physics, University of Massachusetts Lowell

Abstract

Multivariate time-series solar-flare forecasting requires representations that capture magnetic-field evolution while accommodating missing observations, severe class imbalance, and changes across observation periods. This dissertation investigates the integration of topological data analysis with deep sequence models using the SWAN-SF benchmark. Temporal encoders process ordered magnetic-feature sequences, while a complementary branch processes persistence images derived from the geometry of the same observations. The studies examine recurrent architectures, BiLSTM-topology fusion, Transformer-topology fusion, expanding source temporal evaluation, and a revised preprocessing reference for uncertainty analysis.

A recurrent-model comparison shows that architecture and class-balancing choices affect true skill statistic (TSS) and Heidke skill score (HSS2) differently. The BiLSTM-based fusion study reports mean TSS of 0.785 and mean HSS2 of 0.620, compared with 0.755 and 0.603 for its sequence-only ablation. Transformer fusion achieves mean TSS of 0.8149 and mean HSS2 of 0.5795 across four adjacent-partition evaluations. In its first-partition baseline comparison, fusion increases TSS from 0.7746 to 0.8171, alongside a reported trade-off in false alarms and precision. A substantial parameter-count difference limits attribution of the improvement solely to topology.

Expanding-source evaluations report mean TSS and HSS2 of 0.8947 and 0.5129 for TDA-BiLSTM and 0.9022 and 0.4331 for TDA-Transformer. The results illustrate why several measures are needed to assess forecasting behavior. The differences in label cohorts and preprocessing prevent treating all study averages as a common model ranking. This revised protocol, excluding synthetic oversampling and random class-balancing undersampling, while retaining boundary cleaning, supplies nine-fold-seed results with mean TSS of 0.866572 and mean HSS2 of 0.562540. These establish a point-prediction reference and calibration and conformal-coverage conclusions.

The collective evidence supports studying topological descriptors as complementary inputs to temporal models and emphasizes careful separation of architectural effects, sample-selection effects, and temporal transfer. The dissertation develops a framework for interpreting discrimination, probability quality, and prediction-set usefulness, while identifying the matched ablations and experiment records required to strengthen causal and reliability claims.