A Sparse-Attention Framework for High-Dimensional Environmental Cost Prediction

Matta, Tejaswini (2025) A Sparse-Attention Framework for High-Dimensional Environmental Cost Prediction. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (866kB) | Preview
Preview	PDF (Configuration Manual) Download (598kB) | Preview

Abstract

Environmental sustainability assessment increasingly depends on data-driven models that can analyse complex, high-dimensional environmental indicators for accurate environmental cost prediction. However, capturing nonlinear relationships across heterogeneous sustainability features remains challenging, particularly for large-scale corporate reporting and policy-facing decision-making. Traditional analytical and statistical approaches often fail to represent complex interactions in such data and tend to underperform when variables are numerous, correlated, and nonlinearly related. Moreover, many conventional machine learning methods do not incorporate attention mechanisms to focus on the most informative features, while transformer-based models support attention-driven learning but are often computationally intensive, limiting their suitability in resource-constrained environments. To address these limitations, this study leverages a Sparse-Attention Tab Net model that combines attention-based feature selection with improved computational efficiency for environmental cost estimation. A unified evaluation framework is developed to compare classical machine learning baselines, recurrent deep learning models (LSTM and GRU), and Tab Net using consistent preprocessing, train–test splitting, and regression metrics (R², MSE, RMSE). Tab Net applies sparse attention to perform adaptive feature selection and stable learning on tabular environmental data, while also enabling clearer identification of influential predictors through attentive feature masks. Experimental results show that deep learning models outperform traditional machine learning baselines, with GRU and LSTM achieving strong predictive performance. The Sparse-Attention Tab Net model delivers the best overall results, achieving an R² of 0.98 and the lowest MSE and RMSE among all evaluated models. These findings indicate that sparsity-driven attention can improve both accuracy and feature-level clarity for robust environmental cost prediction in sustainability assessment.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Rustam, Furqan UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning G Geography. Anthropology. Recreation > GF Human ecology. Anthropogeography > Sustainability
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	08 Sep 2026 10:18
Last Modified:	08 Sep 2026 10:18
URI:	https://norma.ncirl.ie/id/eprint/9882

Actions (login required)

View Item