We propose a method that sequentially acquires features and selects a classifier for label assignment in data instances of related variables. The objective is to accurately infer the values of such variables (labels), ensuring at the same time that the expected total feature acquisition cost is minimum. To this end, building upon our prior work, our proposed method selects to employ one out of a set of available classifiers after completing the feature acquisition stage. The resulting labels are propagated through the known Bayesian network, which captures the relationships between the variables, and used during the label assignment of the remaining variables. We assess the performance of our method using five datasets, and observe that using classifiers in an instance–wise fashion improves accuracy, but also leads to acquiring less features on average.

We present a sequential acquisition of features and experts framework for datum–wise classification. The goal is to accurately assign labels for each instance, minimizing the acquisition cost of features and experts. An expert uses domain knowledge to make decisions. Starting from a prior belief, features are sequentially acquired in a feature acquisition stage. When this stage terminates, the acquired subset of features is forwarded to an expert acquisition stage, where each expert provides their decision one at a time. At that time, contrary to prior work, the label assignment is reached based on the acquired experts’ decisions thus far. We evaluate the framework’s performance using six real–world datasets and compare it with existing methods. Experiments reveal that the proposed framework increases accuracy up to 56% compared to existing ensemble methods while acquiring 88% fewer features and, more importantly, 80% fewer experts on average.

We introduce a supervised machine learning framework for sequential datum–wise joint feature selection and classification. Our proposed approach sequentially acquires features one at a time during testing until it decides that acquiring more features will not improve label assignment. At that point, and in contrast to prior art, it assigns a label to the example under consideration by selecting between a simple internal and a more powerful external classifier. Easy–to–classify examples are handled by the internal classifier, which assigns labels based on the lowest expected misclassification cost. On the other hand, difficult–to–classify examples are forwarded to the external classifier to be assigned a label based on the acquired features. We demonstrate the performance of the proposed approach compared to existing methods using six publicly available datasets. Experiments indicate that the proposed approach improves accuracy up to 50% with respect to existing sequential methods, while acquiring up to 85% less number of features on average.

Despite the popularity of complex machine learning models, domain experts often struggle to understand and are reluctant to trust them due to lack of intuition and explanation of their predictions. Moreover, these cannot be used in many real–world applications, where features are not readily available but acquired at a cost. To address the latter challenge, dynamic instance–wise joint feature selection and classification selects both the order and the number of features to individually classify each data instance when features are sequentially acquired one at a time. Herein, its model–based and post hoc interpretability is demonstrated validating its utility in high–stakes applications. As a case study, predicting the credit risk of an individual based on financial and other data is considered. Experimental results show that the proposed method is indeed interpretable without sacrificing prediction accuracy.

Administrative data collected by homeless service providers offer a unique opportunity to understand how homeless individuals navigate the homeless system towards securing stable housing. However, the literature on predictive models in the context of homeless service provision has neglected the sequential nature of services that an individual receives over time. Our work addresses this gap by learning, from administrative data, a Bayesian network, which in turn can be used to accurately predict whether an individual will exit the system, or alternatively, the service she would be assigned to the next time she experiences homelessness. Experimental evaluation shows that the proposed approach outperforms prior art not only at predicting exit, but also the less frequent services (and thus more challenging to predict).

We present metamer identification plus (metaID+), an algorithm that enhances the performance of brain-computer interface (BCI)-based color vision assessment. BCI-based color vision assessment uses steady-state visual evoked potentials (SSVEPs) elicited during a grid search of colors to identify metamers—light sources with different spectral distributions that appear to be the same color. Present BCI-based color vision assessment methods are slow; they require extensive data collection for each color in the grid search to reduce measurement noise. metaID+ suppresses measurement noise using Gaussian process regression (i.e., a covariance function is used to replace each measurement with the weighted sum of all of the measurements). Thus, metaID+ reduces the amount of data required for each measurement. We evaluated metaID+ using data collected from ten participants and compared the sum-of-squared errors (SSE; relative to the average grid of each participant) between our algorithm and metaID (an existing algorithm). metaID+ significantly reduced the SSE. In addition, metaID+ achieved metaID’s minimum SSE while using 61.3% less data. By using less data to achieve the same level of error, metaID+ improves the performance of BCI-based color vision assessment.

In this paper, the problem of detecting accidents using speed sensors distributed spatially on a freeway is considered. Due to the significant impact of road accidents on health and development, early and accurate detection of accidents is crucial. To address this issue, a novel Bayesian quickest change detection formulation is introduced, which considers both average detection delay and false alarm rate. The optimum strategy is derived via dynamic programming and shown to compare a recursively computed statistic with a function of false alarm to identify accidents as they happen. Considering that post–accident conditions are typically not known, two methods are proposed that recursively estimate multiple unknown parameters during the accident detection process. Further, four aggregation methods are proposed to improve performance exploiting spatial correlations between sensors. Extensive evaluation results demonstrate improvement up to 65.2% and 87.2% in average detection delay and false alarm rate, respectively, against prior work.

We present CCACUSUM, a classifier for steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) that determines whether a user is attending to a flickering stimulus or is at rest. Correct classification of these two states establishes cause and effect between the BCI and its user, which is essential for helping individuals with complex communication disorders (CCDs) communicate. However, the low signal-to-noise ratio SSVEPs of individuals with CCDs often leads to the incorrect detection of transitions between states, undermining the relationship between cause and effect. To address this challenge, CCACUSUMdetects transitions between states and then infers the user’s state from these transitions. To test our algorithm, we compared CCACUSUMto a traditional canonical correlation analysis classifier using electroencephalography (EEG) data collected from four participants using MusicBox—a simple SSVEP-based BCI. Compared with the traditional classifier, CCACUSUMcorrectly identified true transitions in user state while reducing the number of falsely detected transitions. This improves the cause and effect relationship between the BCI and the user and, with further development, should improve the ability of people with CCDs to use SSVEP-based BCIs to communicate.

We introduce a supervised machine learning framework to perform joint feature selection and classification individually for each data instance during testing. In contrast to our prior work, we decide both the order and the number of features for each data instance. Specifically, our proposed solution dynamically selects the feature to review at each stage based on the already observed features and stops the selection process to make a prediction once it determines no classification improvement can be achieved. To gain insights, we analyze the properties of the proposed solution. Based on these properties, we propose a fast algorithm and demonstrate its effectiveness compared to the state–of–the–art using 4 publicly available datasets.

The potentially detrimental effects of cyberbullying have led to the development of numerous automated, data-driven approaches, with emphasis on classification accuracy. Cyberbullying, as a form of abusive online behavior, although not well-defined, is a repetitive process, i.e., a sequence of aggressive messages sent from a bully to a victim over a period of time with the intent to harm the victim. Existing work has focused on harassment (i.e., using profanity to classify toxic comments independently) as an indicator of cyberbullying, disregarding the repetitive nature of this harassing process. However, raising a cyberbullying alert immediately after an aggressive comment is detected can lead to a high number of false positives. At the same time, two key practical challenges remain unaddressed: (i) detection timeliness, which is necessary to support victims as early as possible, and (ii) scalability to the staggering rates at which content is generated in online social networks. In this work, we introduce CONcISE , a novel approach for timely and accurate Cyberbullying detectiON in online social media SEssions. CONcISE is a two-stage online approach designed to reduce the time to raise a cyberbullying alert by sequentially examining comments as they become available over time, and minimizing the number of feature evaluations necessary for a decision to be made for each comment. Extensive experiments on a real-world Instagram dataset with users and comments demonstrate the effectiveness, scalability, and timeliness of our approach and its benefits over existing methods. Additional experiments using a Twitter dataset offer evidence in support of the potential generalizability of CONcISE to other social media platforms.

Online feature selection and classification is crucial for time sensitive decision making. Existing work however either assumes that features are independent or produces a fixed number of features for classification. Instead, we propose an optimal framework to perform joint feature selection and classification on-the-fly while relaxing the assumption on feature independence. The effectiveness of the proposed approach is showed by classifying urban issue reports on the SeeClickFix civic engagement platform. A significant reduction in the average number of features used is observed without a drop in the classification accuracy.

The relationship between local governments and the general public is being redefined by the increasing use of online platforms that enable participatory reporting of non-emergency urban issues, such as potholes and illegal graffiti by concerned citizens to their local authorities. In this work, we study, for the first time, participatory reporting data together with neighborhood-level demographics, socioeconomic indicators, and pedestrian friendliness and transit and bike scores, across multiple neighborhoods in the Capital District of the New York State. Our data-driven approach offers a large-scale, low-cost alternative to traditional survey methods, and provides insights on citizen participation and satisfaction, and public value creation on such platforms. Our findings can be used to guide government service departments to work more closely with each neighborhood to improve the offline and online communication channels through which citizens can report urban issues.

In this paper, the problem of detecting freeway accidents in real-time based on speed readings from spatially distributed road sensors of variable accuracy is addressed. To ensure robust decision-making, a novel two-stage approach is proposed. Specifically, in the first stage, each sensor generates decisions using a Bayesian quickest change detection framework. In the second stage, individual sensor decisions are aggregated via an optimal stopping approach that optimizes the tradeoff between the costs of aggregation and misclassification. Evaluation of the proposed two-stage approach on a real-world traffic dataset collected from the I405 freeway that passes through the Los Angeles County demonstrates improvements up to 65.2% and 87.2% in average detection delay and probability of false alarm, respectively, as compared to the state-of-the-art.

Empowering citizens to interact directly with their local governments through civic engagement platforms has emerged as an easy way to resolve urban issues. However, for authorities to manually process reported issues is both impractical and inefficient; accurate, online and near-real-time processing methods are necessary to maintain citizens’ satisfaction with their local governments. Herein, an optimal stopping framework is proposed to process urban issue requests quickly and accurately. The optimal classification and stopping rules are derived, and significant reduction in time-to-decision without sacrificing accuracy is demonstrated on a real-world dataset from SeeClickFix.

The potentially detrimental effects of cyberbullying have led to the development of numerous automated, data-driven approaches, with emphasis on classification accuracy. Cyberbullying, as a form of abusive online behavior, although not well-defined, is a repetitive process, i.e., a sequence of aggressive messages sent from a bully to a victim over a period of time with the intent to harm the victim. Existing work has focused on harassment (i.e., using profanity to classify toxic comments independently) as an indicator of cyberbullying, disregarding the repetitive nature of this harassing process. However, raising a cyberbullying alert immediately after an aggressive comment is detected can lead to a high number of false positives. At the same time, two key practical challenges remain unaddressed: (i) detection timeliness, which is necessary to support victims as early as possible, and (ii) scalability to the staggering rates at which content is generated in online social networks.

Cyberbullying has emerged as a large-scale societal problem that demands accurate methods for its detection in an effort to mitigate its detrimental consequences. While automated, data-driven techniques for analyzing and detecting cyberbullying incidents have been developed, the scalability of existing approaches has largely been ignored. At the same time, the complexities underlying cyberbullying behavior (e.g., social context and changing language) make the automatic identification of “the best subset of features” to use challenging. We address this gap by formulating cyberbullying detection as a sequential hypothesis testing problem. Based on this formulation, we propose a novel algorithm to drastically reduce the number of features used in classification. We demonstrate the utility, scalability and responsiveness of our approach using a real-world dataset from Instagram, the online social media platform with the highest percentage of users reporting experiencing cyberbullying. Our approach improves recall by a staggering 700%, while at the same time reducing the average number of features by up to 99.82% compared to state-of-the-art supervised cyberbullying detection methods, learning approaches that require weak supervision, and traditional offline feature selection and dimensionality reduction techniques.

The increase of motor vehicle accidents over the past decades has made their early and accurate detection crucial not only for saving human lives, but also for decreasing time and energy wasted due to congestion. Unlike existing work that has thus far focused on accident detection, this work recognizes the problem of accident localization as an equally important challenge to be addressed. A novel algorithm is proposed to optimally detect the time of an accident and subsequently estimate its geographic location in near-real-time based on speed sensor readings. Evaluation on a large-scale real-world dataset demonstrates that the proposed approach achieves significant gains compared to state-of-the-art with respect to false alarm rate and average detection delay, while at the same time being capable of pinpointing the exact location of accidents within less than 2 miles on average.

Cyberbullying has emerged as a serious societal and public health problem that demands accurate methods for the detection of cyber-bullying instances in an effort to mitigate the consequences. While techniques to automatically detect cyberbullying incidents have been developed, the scalability and timeliness of existing cyberbullying detection approaches have largely been ignored. We address this gap by formulating cyberbullying detection as a sequential hypothesis testing problem. Based on this formulation, we propose a novel algorithm designed to reduce the time to raise a cyberbullying alert by drastically reducing the number of feature evaluations necessary for a decision to be made. We demonstrate the effectiveness of our approach using a real-world dataset from Twitter, one of the top five networks with the highest percentage of users reporting cyberbullying instances. We show that our approach is highly scalable while not sacrificing accuracy for scalability.

This paper considers the problem of state tracking with observation control for a particular class of dynamical systems. The system state evolution is described by a discrete-time, finite-state Markov chain, while the measurement process is characterized by a controlled multi-variate Gaussian observation model. The computational complexity of the optimal control strategy proposed in our prior work proves to be prohibitive. A suboptimal, lower complexity algorithm based on the Fisher information measure is proposed. Toward this end, the preceding measure is generalized to account for multi-valued discrete parameters and control inputs. A closed-form formula for our system model is also derived. Numerical simulations are provided for a physical activity tracking application showing the near-optimal performance of the proposed algorithm.

The problem of sensing strategy design for active state tracking is considered. The system state is modeled by a discrete-time, finite-state Markov chain, which is observed through Gaussian measurement vectors that are dynamically selected by a controller. To overcome the computational complexity associated with the optimal sensing strategy derived in our prior work, a sensing strategy based on the sequential Weiss-Weinstein lower bound (WWLB) is proposed. To this end, closed-form WWLB formulae for our system model are obtained, while accommodating for multi-valued discrete parameters and control inputs. Numerical results validating the success of the proposed strategy on real data from a physical activity tracking application are provided.

In this paper, the problem of state tracking with controlled observations is considered for a system modeled by a discrete-time, finite-state Markov chain. The system state is `hidden' and observed via conditionally Gaussian measurements that are shaped by the underlying state and an exogenous control input. Following an innovations approach, a Kalman-like filter is derived to estimate the Markov chain system state. To optimize the control strategy, the associated mean-squared error is used as an optimization criterion for a partially observable Markov Decision Process (POMDP). The optimal solution is determined via stochastic dynamic programming. Numerical results are presented for the application of physical activity detection in heterogeneous, wireless body area networks.

The problem of enhancing the quality of system state estimates is considered for a special class of dynamical systems. Specifically, a system characterized by a discrete-time, finite-state Markov chain state and observed via conditionally Gaussian measurements is assumed. The associated mean vectors and covariance matrices are tightly intertwined with the system state and a control input selected by a controller. Exploiting an innovations approach, finite-dimensional, non-linear approximate MMSE smoothing estimators are derived for the Markov chain system state. The resulting smoothers are driven by a control policy determined by a stochastic dynamic programming algorithm, which minimizes the MSE filtering error, and was proposed in our earlier work. An application of the smoothers derived in this paper is presented for the problem of physical activity detection in wireless body sensing networks, which illustrates the performance enhancement due to smoothing.

Wireless Body Area Networks (WBANs) refer to a class of wireless sensor networks that are expected to support a wide variety of applications ranging from healthcare and emergency response to entertainment and sports. A WBAN can be characterized by a small number of heterogeneous sensors and an energy-constrained fusion center e.g. cellphone. A key goal is to maximize the lifetime of such a unique sensor network and based on an actual implementation of a prototype WBAN, the limited energy budget of the fusion center is a critical impediment. To overcome this issue, the stochastic control framework introduced in our earlier work is extended to account for less number of samples and sensor heterogeneity is redefined in terms of worst-case detection error probability. A Maximum Likelihood detector based on the belief state is also introduced to increase the detection performance. To account for the energy-constrained fusion center, our initial optimization problem is reformulated to two alternative, but distinct constrained versions and two completely new algorithms, E2MBADP and GME2PS2, are devised. Simulations on real-world data are provided to validate the schemes' performance. Energy gains on the order of 64% while achieving the same detection accuracy (99%) as an equal allocation scheme across sensors are observed.

Wireless body area networks (WBANs) are emerging as a powerful tool for health management, emergency response, military personnel wellness as well as sports and entertainment. In contrast to traditional sensor networks for, say, environmental sensing, WBANs are often characterized by a modest number of heterogeneous sensors wirelessly coupled to a fusion center such as a mobile phone. Based on an actual implementation of a prototype WBAN, energy efficiency at the fusion center has proven to be one of the critical roadblocks to long-term deployment of WBANs. To this end, a novel formulation based on stochastic control tools is devised to model the sensor selection process. Sensors are heterogeneous both in their discrimination capabilities as well as their energy cost, further challenging sensor selection. The goal is to maximize the WBAN's lifetime while optimizing the performance of a physical state detection application. To this end, an optimal dynamic programming algorithm is derived. However, due to the prohibitive complexity of the optimal method, a low-cost approximation scheme, T3S, is designed. The low complexity design is based on several key properties of the cost functional. The proposed T3S scheme is evaluated on real-world data collected from an implemented WBAN and observed to offer near optimal performance with significantly lower complexity.