Dipanwita Thakur
@unical.it
Assistant Professor
University of Calabria
RESEARCH, TEACHING, or OTHER INTERESTS
Computer Engineering, Artificial Intelligence, Human-Computer Interaction, Energy
Scopus Publications
Scholar Citations
Scholar h-index
Scholar i10-index
Scopus Publications
Dipanwita Thakur and Suparna Biswas
Elsevier BV
Dipanwita Thakur and Arindam Pal
Association for Computing Machinery (ACM)
Extreme Learning Machine (ELM) is becoming a popular learning algorithm due to its diverse applications, including Human Activity Recognition (HAR). In ELM, the hidden node parameters are generated at random, and the output weights are computed analytically. However, even with a large number of hidden nodes, feature learning using ELM may not be efficient for natural signals due to its shallow architecture. Due to noisy signals of the smartphone sensors and high dimensional data, substantial feature engineering is required to obtain discriminant features and address the “curse-of-dimensionality”. In traditional ML approaches, dimensionality reduction and classification are two separate and independent tasks, increasing the system’s computational complexity. This research proposes a new ELM-based ensemble learning framework for human activity recognition to overcome this problem. The proposed architecture consists of two key parts: (1) Self-taught dimensionality reduction followed by classification. (2) they are bridged by “Subsampled Randomized Hadamard Transformation” (SRHT). Two different HAR datasets are used to establish the feasibility of the proposed framework. The experimental results clearly demonstrate the superiority of our method over the current state-of-the-art methods.
Dipanwita Thakur, Antonella Guzzo, and Giancarlo Fortino
Institute of Electrical and Electronics Engineers (IEEE)
Dipanwita Thakur and Suparna Biswas
Springer Science and Business Media LLC
Dipanwita Thakur, Antonella Guzzo, and Giancarlo Fortino
IEEE
Human activity recognition (HAR) based on smart devices is gaining increasing attention from the pervasive computing research community due to its wide application in smart healthcare. Modern deep learning models for recognizing human activities rely heavily on sensor data to attain great accuracy. However, leveraging data gathered from smart devices to train such models in a data center results in significant energy consumption and potential privacy violations. By merging numerous local models that are trained on data coming from various clients, federated learning can be used to address the aforementioned problems. By creating a federated learning-based convolutional neural network (CNN) model, which is trained on both public and real-world datasets, we analyze federated learning (FL) to train a human activity identification classifier and compare its performance to centralized learning. The global model achieves accuracy comparable to centralized learning when trained using federated learning on skewed datasets. Additionally, we discover that the selection of clients is a significant problem and suggest a federated learning algorithm that selects the maximum number of clients to increase the convergence rate and model correctness of FL.
Dipanwita Thakur, Sandip Roy, Suparna Biswas, Edmond S. L. Ho, Samiran Chattopadhyay, and Sachin Shetty
IEEE
In smart and intelligent health care, smartphone sensor-based automatic recognition of human activities has evolved as an emerging field of research. In many application domains, deep learning (DL) strategies are more effective than conventional machine learning (ML) models, and human activity recognition (HAR) is no exception. In this paper, we propose a novel framework (CAEL-HAR), that combines CNN, Autoencoder and LSTM architectures for efficient smartphone-based HAR operation. There is a natural synergy between the modeling abilities of LSTMs, autoencoders, and CNNs. While AEs are used for dimensionality reduction and CNNs are the best at automating feature extraction, LSTMs excel at modeling time series. Taking advantage of their complementarity, the proposed methodology combines CNNs, AEs, and LSTMs into a single architecture. We evaluated the proposed architecture using the UCI, WISDM public benchmark datasets. The simulation and experimental results certify the merits of the proposed method and indicate that it outperforms computing time, F1-score, precision, accuracy, and recall in comparison to the current state-of-the-art methods.
Dipanwita Thakur and Suparna Biswas
Institute of Electrical and Electronics Engineers (IEEE)
Transition-aware activity recognition is an inherent component of online health monitoring and ambient assisted living. An explosion of technology breakthroughs in wireless sensor networks, wearable computing, and mobile computing has facilitated this. However, real time, dynamic activity recognition is still challenging in practice. As reported in the existing literature, machine learning techniques are successfully used on the presegmented data to deliver transition-aware activity recognition systems. However, these strategies are frequently ineffective when used in a near-real-time context. This article presents an online change point detection (OCPD) strategy to segment the continuous multivariate time-series smartphone sensor data and its application in a transition-aware activity recognition framework. The proposed OCPD strategy is based on the hypothesis-and-verification principle. After the online data stream segmentation using the proposed OCPD strategy, feature engineering is performed to retain the essential features. Then, synthetic minority oversampling technique (SMOTE) is applied to balance the dataset. Finally, practical experiments are carried out to verify the suggested frameworks’ efficiency and reliability. The results reveal that the proposed OCPD strategy with ensemble classifier achieves a greater recognition rate (F-Measure: 99.80%) compared to methods stated in the literature.
Dipanwita Thakur and Suparna Biswas
Elsevier BV
Dipanwita Thakur and Suparna Biswas
Institute of Electrical and Electronics Engineers (IEEE)
This research revealed a reliable hemiplegia gait monitoring strategy to help medical practitioners in keeping track of a patient’s status. Although numerous technologies have been utilized in the past to collect motion data from patients, the high costs and huge spaces required make them challenging to use in a home setting for rehabilitation. A telemedicine protocol requires a reliable patient monitoring technique that can automatically record and classify patient movements. To achieve this, we propose an attention-based deep learning framework for hemiplegia gait prediction with smartphone-based sensory data, i.e., accelerometer and gyroscope. Firstly, convolutional neural network long short-term memory (CNN-LSTM) architecture is proposed to automatically learn potential features from the high-frequency sensory data. Moreover, considering the effectiveness of the domain expert knowledge-based hand-engineered features for gait analysis, we combine the automatically learned features and the extracted hand-engineered features from sensory data. Secondly, an attention network is proposed to tune the significance of two different features, considering these two different sourced features may be complementary to each other. Finally, extensive experiments are carried out to establish the effectiveness of the suggested hemiplegia gait prediction method in the evaluation of $5\\times $ 2 fold cross-validation and leave-one-subject-out (LOSO) cross-validation, which is more difficult and practical.
Dipanwita Thakur and Suparna Biswas
Chapman and Hall/CRC
Dipanwita Thakur and Suparna Biswas
Springer Nature Singapore
Dipanwita Thakur, Suparna Biswas, Edmond S. L. Ho, and Samiran Chattopadhyay
Institute of Electrical and Electronics Engineers (IEEE)
The self-regulated recognition of human activities from time-series smartphone sensor data is a growing research area in smart and intelligent health care. Deep learning (DL) approaches have exhibited improvements over traditional machine learning (ML) models in various domains, including human activity recognition (HAR). Several issues are involved with traditional ML approaches; these include handcrafted feature extraction, which is a tedious and complex task involving expert domain knowledge, and the use of a separate dimensionality reduction module to overcome overfitting problems and hence provide model generalization. In this article, we propose a DL-based approach for activity recognition with smartphone sensor data, i.e., accelerometer and gyroscope data. Convolutional neural networks (CNNs), autoencoders (AEs), and long short-term memory (LSTM) possess complementary modeling capabilities, as CNNs are good at automatic feature extraction, AEs are used for dimensionality reduction and LSTMs are adept at temporal modeling. In this study, we take advantage of the complementarity of CNNs, AEs, and LSTMs by combining them into a unified architecture. We explore the proposed architecture, namely, “ConvAE-LSTM”, on four different standard public datasets (WISDM, UCI, PAMAP2, and OPPORTUNITY). The experimental results indicate that our novel approach is practical and provides relative smartphone-based HAR solution performance improvements in terms of computational time, accuracy, F1-score, precision, and recall over existing state-of-the-art methods.
Dipanwita Thakur and Suparna Biswas
Springer Singapore
Dipanwita Thakur, Antonella Guzzo, and Giancarlo Fortino
IEEE
Smartwatch based Human Activity Recognition (HAR) is gaining popularity due to habitual unhealthy behavior of the population and the rich in-built sensors of smartwatch. Raw sensor data is not well suited for the classifiers to identify similar activity patterns. According to the HAR literature handcrafted features are beneficial to properly identify the activities, which is time consuming and need expert domain knowledge. Automatic feature extraction libraries give high-dimensional feature sets that increase the computation and memory cost. In this work, we present an Ensemble Learning framework that exploit dimensional reduction and visualization to improve performance specification. Specifically, using Time Series Feature Extraction Library (TSFEL), the high dimensional features are extracted automatically. Then, to reduce the dimension of the feature set and proper visualization, Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE) are used respectively. The relevant extracted features using PCA are fed to an ensemble of three different Machine Learning (ML) classifiers to identify six different human physical activities. We also compare the proposed method with three popularly used shallow ML methods. Self collected human activity smartwatch sensor signal is used to establish the feasibility of the proposed framework. We observe that the proposed framework outper-forms the existing state-of-the-art benchmark frameworks, with an accuracy of 96%.
Dipanwita Thakur and Suparna Biswas
Springer Science and Business Media LLC
Dipanwita Thakur and Suparna Biswas
Springer Science and Business Media LLC
Dipanwita Thakur and Manas Khatua
Elsevier BV
Dipanwita Thakur
IEEE
With the onset of new network technology, the praxis of internet is growing eminently. This leads to tremendous demand on the underlying infrastructure. So load balancing is a technique to efficiently disseminate the load across the various entities to utilize the resources, maximize throughput and scale down the response time to avoid the single entity from being overloaded. Software Defined Network (SDN) is seen as a promising toolset to expedite network programmability to develop network application. Unlike traditional network paradigm, it dissevers the control (brain) and forwarding plane, which benefits dynamic computing. SDN controller is accountable for managing the complete network, it has the global view and knowledge about the load on every switch along the routing path, which makes the practice of load balancing easy and cost-effective. To allay the problem of availability and scalability faced due to the centralized controller (i.e., Single point of failure), multiple distributed SDN controllers are used. Distributing the load among these multiple controllers is itself a significant challenge. This paper proposes an algorithm for multiple distributed SDN controller load balancing using clustering and bully election algorithm.
Richa Kesharwani and Dipanwita Thakur
IEEE
The geospatial objects laying on the earth surfaces are changing rapidly due to various reasons. The multi-temporal remote sensing images are widely used data sources to capture changed occurred at a particular location. Change detection is the process to determine disagreement in the nature of geospatial objects of the same location from different time's instances. Unsupervised change detection is an approach to determine the changed and unchanged state of a location without using any label information. The remote sensing images are captured in multiple spectral bands and the extraction of spatial, texture and shape features in each band increases the dimensionality of the data sets. The conventional change detection techniques may perform poor for the high dimensional remote sensing data sets. In this work, a multi-view learning based unsupervised technique has been developed for the unsupervised change detection in temporal remote sensing images. The experimental results on the two benchmark data sets show that the proposed unsupervised multi-view approach has performed better than single view unsupervised change detection techniques.
Dipanwita Thakur and Manas Khatua
Springer Singapore