Minh Nhut Nguyen

Positive unlabeled learning for time series classification

In many real-world applications of the time series classification problem, not only could the negative training instances be missing, the number of positive instances available for learning may also be rather limited. This has motivated the development of new classification algorithms that can learn from a small set P of labeled seed positive instances augmented with a set U of unlabeled instances (i.e. PU learning algorithms). However, existing PU learning algorithms for time series classification have less than satisfactory performance as they are unable to identify the class boundary between positive and negative instances accurately. In this paper, we propose a novel PU learning algorithm LCLC (Learning from Common Local Clusters) for time series classification. LCLC is designed to effectively identify the ground truths positive and negative boundaries, resulting in more accurate classifiers than those constructed using existing methods. We have applied LCLC to classify time series data from different application domains; the experimental results demonstrate that LCLC out-performs existing methods significantly.

An integrated framework for human activity classification

This paper presents an integrated framework to enable using standard non-sequential machine learning tools for accurate multi-modal activity recognition. We develop a novel framework that contains simple pre- and post-classification strategies to improve the overall performance. We achieve this through class-imbalance correction on the learning data using structure preserving oversampling (SPO), leveraging the sequential nature of sensory data using smoothing of the predicted label sequence and classifier fusion, respectively. Through evaluation on recent publicly available activity datasets comprising of a large amount of multi-dimensional sensory data, we demonstrate that our proposed strategies are effective in improving classification performance over common techniques such as One Nearest Neighbor (1NN) and Support Vector Machines (SVM). Our framework also shows better performance over sequential probabilistic models, such as Conditional Random Field (CRF) and Hidden Markov Model (HMM) and when these models are used as meta-learners.

Ensemble based positive unlabeled learning for time series classification

Many real-world applications in time series classification fall into the class of positive and unlabeled (PU) learning. Furthermore, in many of these applications, not only are the negative examples absent, the positive examples available for learning can also be rather limited. As such, several PU learning algorithms for time series classification have recently been developed to learn from a small set P of labeled seed positive examples augmented with a set U of unlabeled examples. The key to these algorithms is to accurately identify the likely positive and negative examples from U, but it has remained a challenge, especially for those uncertain examples located near the class boundary. This paper presents a novel ensemble based approach that restarts the detection phase several times to probabilistically label these uncertain examples more robustly so that a reliable classifier can be built from the limited positive training examples. Experimental results on time series data from different domains demonstrate that the new method outperforms existing state-of-the art methods significantly.

Home and Work Place Prediction for Urban Planning Using Mobile Network Data

We present methods to predict and validate home and work places of anonymized users using their mobile network data. Knowledge of home and work place of a user is essential in order to find his (and overall population) mobility profiles. There are many methods that predict home and work places using GPS data. But unlike GPS data, mobile network data using GSM do not provide the exact location of a phone event. We use a novel criterion that combines an extracted feature from mobile data (i.e., Inactivity - no phone event for a given period of time) with open source data about location category % (i.e., Streetdirectory.com) to predict home location. Results show that the new criterion gives better prediction accuracy than inactivity alone. We predict work place using the idea that one goes to her work place on most of the weekdays but rarely on weekends. We validate our methods by comparing against the ground truth obtained from open source data. Validation results show that our proposed methods are about 25% more accurate than existing methods both for home and work place predictions.

Ensemble Based Real-Time Adaptive Classification System for Intelligent Sensing Machine Diagnostics

The deployment of a sensor node to manage a group of sensors and collate their readings for system health monitoring is gaining popularity within the manufacturing industry. Such a sensor node is able to perform real-time configurations of the individual sensors that are attached to it. Sensors are capable of acquiring data at different sampling frequencies based on the sensing requirements. The different sampling rates affect power consumption, sensor lifespan, and the resultant network bandwidth usage due to the data transfer incurred. These settings also have an immediate impact on the accuracy of the diagnostics and prognostics models that are employed for system health monitoring. In this paper, we propose a novel adaptive classification system architecture for system health monitoring that is well suited to accommodate and take advantage of the variable sampling rate of sensors. As such, our proposed system is able to yield a more effective health monitoring system by reducing the power consumption of the sensors, extending the sensors lifespan, as well as reducing the resultant network traffic and data logging requirements. We also propose an ensemble based learning method to integrate multiple existing classifiers with different feature representations, which can achieve significantly better, stable results compared with the individual state-of-the-art techniques, especially in the scenario when we have very limited training data. This result is extremely important in many real-world applications because it is often impractical, if not impossible, to hand-label large amounts of training data.

Discovery of Rab1 binding sites using an ensemble of clustering methods: Clustering for Finding Rab1 Binding Sites

Targeting non‐native‐ligand binding sites for potential investigative and therapeutic applications is an attractive strategy in proteins that share common native ligands, as in Rab1 protein. Rab1 is a subfamily member of Rab proteins, which are members of Ras GTPase superfamily. All Ras GTPase superfamily members bind to native ligands GTP and GDP, that switch on and off the proteins, respectively. Rab1 is physiologically essential for autophagy and transport between endoplasmic reticulum and Golgi apparatus. Pathologically, Rab1 is implicated in human cancers, a neurodegenerative disease, cardiomyopathy, and bacteria‐caused infectious diseases. We have performed structural analyses on Rab1 protein using a unique ensemble of clustering methods, including multi‐step principal component analysis, non‐negative matrix factorization, and independent component analysis, to better identify representative Rab1 proteins than the application of a single clustering method alone does. We then used the identified representative Rab1 structures, resolved in multiple ligand states, to map their known and novel binding sites. We report here at least a novel binding site on Rab1, involving Rab1‐specific residues that could be further explored for the rational design and development of investigative probes and/or therapeutic small molecules against the Rab1 protein. Proteins 2017; 85:859–871.

Structural analysis of protein tyrosine phosphatase 1B reveals potentially druggable allosteric binding sites

Catalytic proteins such as human protein tyrosine phosphatase 1B (PTP1B), with conserved and highly polar active sites, warrant the discovery of druggable nonactive sites, such as allosteric sites, and potentially, therapeutic small molecules that can bind to these sites. Catalyzing the dephosphorylation of numerous substrates, PTP1B is physiologically important in intracellular signal transduction pathways in diverse cell types and tissues. Aberrant PTP1B is associated with obesity, diabetes, cancers, and neurodegenerative disorders. Utilizing clustering methods (based on root mean square deviation, principal component analysis, nonnegative matrix factorization, and independent component analysis), we have examined multiple PTP1B structures. Using the resulting representative structures in different conformational states, we determined consensus clustroids and used them to identify both known and novel binding sites, some of which are potentially allosteric. We report several lead compounds that could potentially bind to the novel PTP1B binding sites and can be further optimized. Considering the possibility for drug repurposing, we discovered homologous binding sites in other proteins, with ligands that could potentially bind to the novel PTP1B binding sites.

Deep Learning for Human Activity Recognition

This chapter focuses on the problem of human activity recognition (HAR), in which inputs in the form of multichannel time series signals are acquired from a set of body-worn wearable sensors and outputs are predefined human activities. In this problem, extracting effective features for identifying activities is a critical but challenging task. Most existing work relies on heuristic hand-crafted feature design and shallow feature learning architectures, which cannot find very discriminative features to accurately classify different activities. In this chapter, we propose a systematic feature learning method for the HAR problem. This method adopts a deep convolutional neural network (CNN) to automate feature learning from the raw inputs in a systematic way. Through the deep architecture, higher level abstract representations of low level raw time series signals are learned as effective features without the need for hand-crafting features. By leveraging the labeled information via supervised learning, the learned features are endowed with more discriminative power. Such a unification of feature learning and classification results in mutual enhancements in both. These unique advantages of the CNN lead to a mutually enhanced outcome of HAR, as verified in the experiments on multiple HAR datasets and comparisons with several state-of-the-art techniques.

An integrated framework for human activity recognition

This poster presents an integrated framework to enable using standard non-sequential machine learning tools for accurate multi-modal activity recognition. Our framework contains simple pre- and post-classification strategies such as class-imbalance correction on the learning data using structure preserving oversampling, leveraging the sequential nature of sensory data using smoothing of the predicted label sequence and classifier fusion, respectively, for improved performance. Through evaluation on recent publicly-available OPPORTUNITY activity datasets comprising of a large amount of multi-dimensional, continuous-valued sensory data, we show that our proposed strategies are effective in improving the performance over common techniques such as One Nearest Neighbor (1NN) and Support Vector Machines (SVM). Our framework also shows better performance over sequential probabilistic models, such as Conditional Random Field (CRF) and Hidden Markov Models (HMM) and when these models are used as meta-learners.

In situ hydrogel formation for biomedical applications using acoustic cavitation from high intensity focused ultrasound

There is a growing interest in polymer mechanochemistry for their industrial applications. For example, stress-induced crosslinking gel formation from polymer networks is a rapidly growing field of study. Recent work utilizes a variety of different polymer structures and crosslinking mechanisms. However, these polymers are typically soluble in only organic solvents and require the use of a sonicating probe or bath at frequencies below 100 kHz. These requirements limit their use in biomedical applications that require in situ gel formation within a patient (e.g., blocking of varicose veins, internal wound healing, etc.). Here we report on the development of a water soluble block copolymer that forms a hydrogel in the presence of acoustic cavitation from high intensity focused ultrasound. These block copolymers are comprised of hydrophilic polyethylene glycol methyl methacrylate units and hydrophobic tridentate crosslinkers. The tridentate crosslinker forms bonds with free metal ions in solution only in the presence of acoustic cavitation induced mechanical stress. We show that the block copolymer is capable of forming a hydrogel in under 90 seconds and will also block a liquid channel formed in an agarose cylinder.There is a growing interest in polymer mechanochemistry for their industrial applications. For example, stress-induced crosslinking gel formation from polymer networks is a rapidly growing field of study. Recent work utilizes a variety of different polymer structures and crosslinking mechanisms. However, these polymers are typically soluble in only organic solvents and require the use of a sonicating probe or bath at frequencies below 100 kHz. These requirements limit their use in biomedical applications that require in situ gel formation within a patient (e.g., blocking of varicose veins, internal wound healing, etc.). Here we report on the development of a water soluble block copolymer that forms a hydrogel in the presence of acoustic cavitation from high intensity focused ultrasound. These block copolymers are comprised of hydrophilic polyethylene glycol methyl methacrylate units and hydrophobic tridentate crosslinkers. The tridentate crosslinker forms bonds with free metal ions in solution only in the...