Published Papers
HIERVAR: A Hierarchical Feature Selection Method for Time Series Analysis
IEEE International Workshop on Machine Learning for Signal Processing (MLSP) 2024
Random representation methods like ROCKET and MiniROCKET generate powerful features but produce thousands of redundant ones. HIERVAR addresses this through a hierarchical feature selection approach combining ANOVA variance analysis with E-ROCKET. Our method substantially reduces features by over 94% while preserving classification accuracy, marking a significant advancement in time series feature selection.
Key Contributions: Two-phase selection pipeline integrating ridge coefficient ranking with ANOVA filtering. Demonstrates ~90% feature reduction without accuracy loss across 108 UCR datasets.
Time Series Classification Using Convolutional Kernel and Adaptive Dynamic Thresholding
IEEE International Conference on Communications (ICC) 2024
A novel classification approach combining the Metropolis-Hastings algorithm within an MCMC framework for optimized feature selection. The method integrates convolutional kernels with randomized threshold exceedance rates (rTER) for robust feature extraction, and employs Locality-Sensitive Hashing (LSH) for diverse feature selection. Achieves state-of-the-art efficiency on medical time-series datasets.
Key Contributions: Integration of MCMC-based adaptive sampling with LSH for feature diversification. Novel dynamic thresholding mechanism achieving 98.4% accuracy on CinCECGTorso dataset.
RASTER: Representation Learning for Time Series Classification using Scatter Score and Randomized Threshold Exceedance Rate
IEEE International Workshop on Machine Learning for Signal Processing (MLSP) 2023
RASTER introduces a novel memory-efficient time series representation learning approach. It uses a modified Swendsen-Wang algorithm to create diverse filter banks, proposes Scatter Score (SS) as a fast metric to evaluate feature informativeness, and introduces randomized Threshold Exceedance Rate (rTER) for temporal-aware downsampling. RASTER significantly outperforms ROCKET, MiniROCKET, ResNet, and InceptionTime across 30 datasets.
Key Contributions: Time Series Swendsen-Wang (TSSW) algorithm for filter generation. Scatter Score metric for dilation factor evaluation. rTER for localized temporal pattern detection.
Representation Learning of Clinical Multivariate Time Series with Random Filter Banks
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2023
The RandFreqButch method addresses the challenge of training ML models on limited medical time series data. By generating multiple random filters in the frequency domain with varied cutoff frequencies, the method creates enriched signal representations for classification. Experiments on EEG, ECG, and respiratory oscillometry datasets demonstrate significant performance improvements over baseline methods.
Key Contributions: Frequency-domain augmentation strategy for limited healthcare data. Random lowpass/highpass/bandpass filter banks for feature enrichment. Up to 13% accuracy improvement on medical datasets.
Modified Deep Residual Network Architecture Deployed on Serverless IoT Platform Based on Human Activity Recognition Application
Future Generation Computer Systems (Elsevier)
This paper presents deep learning pipelines for smartphone sensor time-series classification in human activity recognition. The approach scales training and inference using distributed processing on Apache Spark, emphasizing efficient data handling and reproducible experimentation for IoT applications.
Key Contributions: Modified ResNet architecture optimized for HAR. Serverless deployment architecture for IoT edge computing. Distributed training pipeline using Apache Spark.
Research Themes
Random Representation Learning
Methods like ROCKET, MiniROCKET, and RASTER that use randomized convolutional kernels to efficiently extract features from time series without expensive training procedures.
Feature Selection
Techniques including HIERVAR, CoMIND, and CURVE for selecting compact, discriminative feature subsets while maintaining classification performance on high-dimensional data.
Healthcare ML
Clinical applications including oscillometry analysis, EEG/ECG classification, and uncertainty-aware models for safety-critical medical decision support.