Durga Toshniwal
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Durga Toshniwal has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
- Ashish Kumar
Indian Institute of Technology Roorkee Roorkee India
5 shared publications - Bhola Ram Gurjar
Indian Institute of Technology Roorkee
3 shared publications Manoranjan Parida
Civil Engineering, Indian Institute of Technology Roorkee, Roorkee, India
2 shared publications- Neha Pant
Computer Science and Engineering, Indian Institute of Technology Roorkee, Roorkee, India
2 shared publications - Prashant Kumar
Department of Computer Science & Engineering, Indian Institute of Technology, Roorkee, India
2 shared publications - Tripti Garg
Computer Science and Engineering, Indian Institute of Technology Roorkee, Roorkee, India
2 shared publications
Latest Publications
- Open Access
- 20 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Intelligent Systems and Technology
- Vol. 17 (3) , 1-25
Frequency-Modulated Transformer Self-Attention for Advanced Infectious Disease Prediction
Time series forecasting of infectious diseases is crucial to addressing significant global health challenges. The application of AI, particularly Deep Learning (DL) algorithms, has demonstrated substantial success in sequence modeling; however, their performance in epidemiological forecasting remains constrained by the non-stationary nature and complex frequency dynamics of disease transmission. This research introduces a novel Modulated Transformer (FMT) framework to address these challenges. The FMT framework decomposes time series data into distinct frequency-modulated signals, utilizing self-attention mechanisms to capture temporal frequencies. A transformer encoder–decoder architecture then predicts and captures multi-scale temporal dependencies and makes accurate predictions. The novelty of the proposed approach lies in decomposing the input time series into Intrinsic Mode Functions (IMFs) and integrating the frequency-specific components into a Transformer architecture via entropy-based feature selection. The FMT framework significantly reduces Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) by approximately 50% and 65%, respectively, compared to conventional methods. Additionally, it achieves an 8% increase in the R2 score, demonstrating enhanced predictive accuracy. The proposed methodology is evaluated using COVID-19 datasets from multiple countries along with an influenza dataset and benchmarked against statistical, machine learning, and state-of-the-art DL baselines. The contribution of entropy-based IMF integration is systematically examined by comparing results with and without this component, underscoring its importance in improving predictive accuracy. This work highlights substantial improvements in predictive accuracy and computational efficiency, advancing epidemiological forecasting and supporting real-time public health decision-making and AI-driven disease surveillance systems.
- 20 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Intelligent Systems and Technology
- Vol. 17 (3) , 1-25
Open AccessFrequency-Modulated Transformer Self-Attention for Advanced Infectious Disease Prediction
Time series forecasting of infectious diseases is crucial to addressing significant global health challenges. The application of AI, particularly Deep Learning (DL) algorithms, has demonstrated substantial success in sequence modeling; however, their performance in epidemiological forecasting remains constrained by the non-stationary nature and complex frequency dynamics of disease transmission. This research introduces a novel Modulated Transformer (FMT) framework to address these challenges. The FMT framework decomposes time series data into distinct frequency-modulated signals, utilizing self-attention mechanisms to capture temporal frequencies. A transformer encoder–decoder architecture then predicts and captures multi-scale temporal dependencies and makes accurate predictions. The novelty of the proposed approach lies in decomposing the input time series into Intrinsic Mode Functions (IMFs) and integrating the frequency-specific components into a Transformer architecture via entropy-based feature selection. The FMT framework significantly reduces Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) by approximately 50% and 65%, respectively, compared to conventional methods. Additionally, it achieves an 8% increase in the R2 score, demonstrating enhanced predictive accuracy. The proposed methodology is evaluated using COVID-19 datasets from multiple countries along with an influenza dataset and benchmarked against statistical, machine learning, and state-of-the-art DL baselines. The contribution of entropy-based IMF integration is systematically examined by comparing results with and without this component, underscoring its importance in improving predictive accuracy. This work highlights substantial improvements in predictive accuracy and computational efficiency, advancing epidemiological forecasting and supporting real-time public health decision-making and AI-driven disease surveillance systems.
- Open Access
- 9 December 2025
- Published by Springer Nature in Scientific Reports
- Vol. 16 (1)
A meta-learning ensemble framework for robust and interpretable prediction of emergency medical services demand
Accurate and robust forecasting of Emergency Medical Services (EMS) demand is crucial for ensuring timely ambulance dispatch and efficient resource allocation, particularly in low-resource public health systems, such as those in India. While most prior EMS forecasting studies have focused on urban settings in developed countries with rich, granular data, limited research has explored district-level forecasting using real-world ambulance dispatch data from India. Moreover, existing models often trade off robustness for accuracy or rely on complex black-box architectures, limiting their interpretability and real-world deployment. This study examines whether a heterogeneous ensemble of interpretable and complementary learners can outperform traditional and state-of-the-art regressors for district-level EMS forecasting, utilizing limited real-world features. To address this challenge, we propose EM-LR (Ensembled Meta-Learner with Linear Regression), a meta-learning framework that integrates four diverse base models−Lasso Regression, Support Vector Regression (SVR), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGB)−via a linear regression meta-learner. Unlike prior meta-learners that stack similar tree-based or linear models, EM-LR combines low-variance, diverse learners to enhance robustness while maintaining model interpretability through SHAP-based feature analysis and transparent ensemble weights. Using only temporal and meteorological inputs, EM-LR forecasts daily EMS call volumes across five diverse districts in the state of Uttar Pradesh. We benchmark EM-LR against traditional models and recent advanced variants, including Twin Bounded Least Squares Support Vector Regression (TBLSSVR), Asymmetric-Huber based Extreme Learning Machine (AHELM), and Mexican-Hat Kernelized Large Margin Distribution Machine-based Regression (MHKLDMR), demonstrating superior accuracy and reduced prediction variance. Experimental results show up to 9.5% reduction in RMSE and over 40% variance reduction. EM-LR thus offers a scalable and interpretable forecasting solution tailored to the operational constraints of developing public health systems, supporting data-driven emergency planning and equitable healthcare delivery.
- 9 December 2025
- Published by Springer Nature in Scientific Reports
- Vol. 16 (1)
Open AccessA meta-learning ensemble framework for robust and interpretable prediction of emergency medical services demand
Accurate and robust forecasting of Emergency Medical Services (EMS) demand is crucial for ensuring timely ambulance dispatch and efficient resource allocation, particularly in low-resource public health systems, such as those in India. While most prior EMS forecasting studies have focused on urban settings in developed countries with rich, granular data, limited research has explored district-level forecasting using real-world ambulance dispatch data from India. Moreover, existing models often trade off robustness for accuracy or rely on complex black-box architectures, limiting their interpretability and real-world deployment. This study examines whether a heterogeneous ensemble of interpretable and complementary learners can outperform traditional and state-of-the-art regressors for district-level EMS forecasting, utilizing limited real-world features. To address this challenge, we propose EM-LR (Ensembled Meta-Learner with Linear Regression), a meta-learning framework that integrates four diverse base models−Lasso Regression, Support Vector Regression (SVR), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGB)−via a linear regression meta-learner. Unlike prior meta-learners that stack similar tree-based or linear models, EM-LR combines low-variance, diverse learners to enhance robustness while maintaining model interpretability through SHAP-based feature analysis and transparent ensemble weights. Using only temporal and meteorological inputs, EM-LR forecasts daily EMS call volumes across five diverse districts in the state of Uttar Pradesh. We benchmark EM-LR against traditional models and recent advanced variants, including Twin Bounded Least Squares Support Vector Regression (TBLSSVR), Asymmetric-Huber based Extreme Learning Machine (AHELM), and Mexican-Hat Kernelized Large Margin Distribution Machine-based Regression (MHKLDMR), demonstrating superior accuracy and reduced prediction variance. Experimental results show up to 9.5% reduction in RMSE and over 40% variance reduction. EM-LR thus offers a scalable and interpretable forecasting solution tailored to the operational constraints of developing public health systems, supporting data-driven emergency planning and equitable healthcare delivery.
- Open Access
- 26 February 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-32
Weather-driven risk assessment model for two-wheeler road crashes in Uttar Pradesh, India
This study investigates the relationship between weather conditions and two-wheeler road crashes in Uttar Pradesh, India, which experiences diverse climatic conditions. A novel framework, the Weather-Influenced Clustering and Random Sampling (WICRS) model, is proposed for Relative Accident (crash) Risk (RAR) analysis. Initially, a preliminary analysis of crash data based on location, human, and environmental factors provides insights into contributing factors. Building on these findings, the WICRS model categorizes weather patterns using highly randomized sampling-based clustering, a departure from traditional matched pair analysis (MPA). The study also conducts a stratified RAR analysis, considering variables such as gender, road type, and time of day. The effectiveness of the WICRS model is validated by comparing its impact with MPA, specifically examining risk analysis for wet and non-wet days. The dataset includes over 954,000 two-wheeler crash incidents, combined with historical weather data over six years. The findings highlight the significance of weather conditions in two-wheeler crashes and support the use of the WICRS model for detailed RAR analysis and road safety policy formulation.
- 26 February 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-32
Open AccessWeather-driven risk assessment model for two-wheeler road crashes in Uttar Pradesh, India
This study investigates the relationship between weather conditions and two-wheeler road crashes in Uttar Pradesh, India, which experiences diverse climatic conditions. A novel framework, the Weather-Influenced Clustering and Random Sampling (WICRS) model, is proposed for Relative Accident (crash) Risk (RAR) analysis. Initially, a preliminary analysis of crash data based on location, human, and environmental factors provides insights into contributing factors. Building on these findings, the WICRS model categorizes weather patterns using highly randomized sampling-based clustering, a departure from traditional matched pair analysis (MPA). The study also conducts a stratified RAR analysis, considering variables such as gender, road type, and time of day. The effectiveness of the WICRS model is validated by comparing its impact with MPA, specifically examining risk analysis for wet and non-wet days. The dataset includes over 954,000 two-wheeler crash incidents, combined with historical weather data over six years. The findings highlight the significance of weather conditions in two-wheeler crashes and support the use of the WICRS model for detailed RAR analysis and road safety policy formulation.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 806-813
Semantic alignment in hyperbolic space for fine-grained emotion classification
Ashish Kumar, Durga Toshniwal. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop). 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 806-813
Open AccessSemantic alignment in hyperbolic space for fine-grained emotion classification
Ashish Kumar, Durga Toshniwal. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop). 2025.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 872-883
HyILR: Hyperbolic Instance-Specific Local Relationships for Hierarchical Text Classification
Ashish Kumar, Durga Toshniwal. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop). 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 872-883
Open AccessHyILR: Hyperbolic Instance-Specific Local Relationships for Hierarchical Text Classification
Ashish Kumar, Durga Toshniwal. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop). 2025.
- Open Access
Improving Multi-Domain Task-Oriented Dialogue System with Offline Reinforcement Learning
Task-oriented dialogue (TOD) system is designed to accomplish user-defined tasks through dialogues. The TOD system has progressed towards end-to-end modeling by leveraging pre-trained large language models. Fine-tuning the pre-trained language models using only supervised learning leads to the exposure bias and token loss problem and it deviates the models from completing the user's task. To address these issues, we propose a TOD system that leverages a unified pre-trained language model, GPT2, as a base model. It is optimized using supervised learning and reinforcement learning (RL). The issues in the TOD system are mitigated using a non-differentiable reward function. The reward is calculated using the weighted sum of the success rate and BLEU evaluation metrics. The success rate and BLEU metrics in reward calculation guide the language model for user task completion while ensuring a coherent and fluent response. Our model is acquired by fine-tuning a pre-trained model on the dialogue-session level which comprises user utterance, belief state, system act, and system response. Experimental results on MultiWOZ2.1 demonstrate that our model increases the inform rate by 1.60% and the success rate by 3.17% compared to the baseline.
Open AccessImproving Multi-Domain Task-Oriented Dialogue System with Offline Reinforcement Learning
Task-oriented dialogue (TOD) system is designed to accomplish user-defined tasks through dialogues. The TOD system has progressed towards end-to-end modeling by leveraging pre-trained large language models. Fine-tuning the pre-trained language models using only supervised learning leads to the exposure bias and token loss problem and it deviates the models from completing the user's task. To address these issues, we propose a TOD system that leverages a unified pre-trained language model, GPT2, as a base model. It is optimized using supervised learning and reinforcement learning (RL). The issues in the TOD system are mitigated using a non-differentiable reward function. The reward is calculated using the weighted sum of the success rate and BLEU evaluation metrics. The success rate and BLEU metrics in reward calculation guide the language model for user task completion while ensuring a coherent and fluent response. Our model is acquired by fine-tuning a pre-trained model on the dialogue-session level which comprises user utterance, belief state, system act, and system response. Experimental results on MultiWOZ2.1 demonstrate that our model increases the inform rate by 1.60% and the success rate by 3.17% compared to the baseline.
- Open Access
Modeling Text-Label Alignment for Hierarchical Text Classification
Hierarchical Text Classification (HTC) aims to categorize text data based on a structured label hierarchy, resulting in predicted labels forming a sub-hierarchy tree. The semantics of the text should align with the semantics of the labels in this sub-hierarchy. With the sub-hierarchy changing for each sample, the dynamic nature of text-label alignment poses challenges for existing methods, which typically process text and labels independently. To overcome this limitation, we propose a Text-Label Alignment (TLA) loss specifically designed to model the alignment between text and labels. We obtain a set of negative labels for a given text and its positive label set. By leveraging contrastive learning, the TLA loss pulls the text closer to its positive label and pushes it away from its negative label in the embedding space. This process aligns text representations with related labels while distancing them from unrelated ones. Building upon this framework, we introduce the Hierarchical Text-Label Alignment (HTLA) model, which leverages BERT as the text encoder and GPTrans as the graph encoder and integrates text-label embeddings to generate hierarchy-aware representations. Experimental results on benchmark datasets and comparison with existing baselines demonstrate the effectiveness of HTLA for HTC.
Open AccessModeling Text-Label Alignment for Hierarchical Text Classification
Hierarchical Text Classification (HTC) aims to categorize text data based on a structured label hierarchy, resulting in predicted labels forming a sub-hierarchy tree. The semantics of the text should align with the semantics of the labels in this sub-hierarchy. With the sub-hierarchy changing for each sample, the dynamic nature of text-label alignment poses challenges for existing methods, which typically process text and labels independently. To overcome this limitation, we propose a Text-Label Alignment (TLA) loss specifically designed to model the alignment between text and labels. We obtain a set of negative labels for a given text and its positive label set. By leveraging contrastive learning, the TLA loss pulls the text closer to its positive label and pushes it away from its negative label in the embedding space. This process aligns text representations with related labels while distancing them from unrelated ones. Building upon this framework, we introduce the Hierarchical Text-Label Alignment (HTLA) model, which leverages BERT as the text encoder and GPTrans as the graph encoder and integrates text-label embeddings to generate hierarchy-aware representations. Experimental results on benchmark datasets and comparison with existing baselines demonstrate the effectiveness of HTLA for HTC.
- 22 August 2024
- Published by Springer Nature
- p. 163-179
Modeling Text-Label Alignment for Hierarchical Text Classification
- 22 August 2024
- Published by Springer Nature
- p. 163-179
Modeling Text-Label Alignment for Hierarchical Text Classification
- 18 June 2024
- Published by Elsevier in Computers and Electrical Engineering
- Vol. 118
FedNISP: Neuron Importance Scope Propagation pruning for communication efficient federated learning
- 18 June 2024
- Published by Elsevier in Computers and Electrical Engineering
- Vol. 118
FedNISP: Neuron Importance Scope Propagation pruning for communication efficient federated learning
- Open Access
- 16 May 2024
- Published by Springer Nature in Scientific Reports
- Vol. 14 (1) , 1-12
Multi-step forecasting of dissolved oxygen in River Ganga based on CEEMDAN-AdaBoost-BiLSTM-LSTM model
Accurate prediction of Dissolved Oxygen (DO) is an integral part of water resource management. This study proposes a novel approach combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) with AdaBoost and deep learning for multi-step forecasting of DO. CEEMDAN generates Intrinsic Mode Functions (IMFs) with different frequencies, capturing non-linear and non-stationary characteristics of the data. The high-frequency and medium-frequency IMFs, characterized by complex patterns and frequent changes over time, are predicted using Adaboost with Bidirectional Long Short-Term Memory (BiLSTM) as the base estimator. The low-frequency IMFs, characterized by relatively simple patterns, are predicted using standalone Long Short-Term Memory (LSTM). The proposed CEEMDAN-AdaBoost-BiLSTM-LSTM model is tested on data from ten stations of river Ganga. We compare the results with six models without decomposition and four models utilizing decomposition. Experimental results show that using a tailored prediction technique based on each IMF’s distinctive features leads to more accurate forecasts. CEEMDAN-AdaBoost-BiLSTM-LSTM outperforms CEEMDAN-BiLSTM with an average improvement of 25.458% for RMSE and 37.390% for MAE. Compared with CEEMDAN-AdaBoost-BiLSTM, an average improvement of 20.779% for RMSE and 28.921% for MAE is observed. Diebold-Mariano test and t-test suggest a statistically significant difference in performance between the proposed and compared models.
- 16 May 2024
- Published by Springer Nature in Scientific Reports
- Vol. 14 (1) , 1-12
Open AccessMulti-step forecasting of dissolved oxygen in River Ganga based on CEEMDAN-AdaBoost-BiLSTM-LSTM model
Accurate prediction of Dissolved Oxygen (DO) is an integral part of water resource management. This study proposes a novel approach combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) with AdaBoost and deep learning for multi-step forecasting of DO. CEEMDAN generates Intrinsic Mode Functions (IMFs) with different frequencies, capturing non-linear and non-stationary characteristics of the data. The high-frequency and medium-frequency IMFs, characterized by complex patterns and frequent changes over time, are predicted using Adaboost with Bidirectional Long Short-Term Memory (BiLSTM) as the base estimator. The low-frequency IMFs, characterized by relatively simple patterns, are predicted using standalone Long Short-Term Memory (LSTM). The proposed CEEMDAN-AdaBoost-BiLSTM-LSTM model is tested on data from ten stations of river Ganga. We compare the results with six models without decomposition and four models utilizing decomposition. Experimental results show that using a tailored prediction technique based on each IMF’s distinctive features leads to more accurate forecasts. CEEMDAN-AdaBoost-BiLSTM-LSTM outperforms CEEMDAN-BiLSTM with an average improvement of 25.458% for RMSE and 37.390% for MAE. Compared with CEEMDAN-AdaBoost-BiLSTM, an average improvement of 20.779% for RMSE and 28.921% for MAE is observed. Diebold-Mariano test and t-test suggest a statistically significant difference in performance between the proposed and compared models.