Most existing evaluations of explainable machine learning (ML) methods rely on simplifying assumptions or proxies that do not reflect real-world use cases; the handful of more robust evaluations on real-world settings have shortcomings in their design, generally leading to overestimation of methods' real-world utility. In this work, we seek to address this by conducting a study that evaluates post-hoc explainable ML methods in a setting consistent with the application context and provide a template for future evaluation studies. We modify and improve a prior study on e-commerce fraud detection by relaxing the original work's simplifying assumptions that departed from the deployment context. Our study finds no evidence for the utility of the tested explainable ML methods in the context, which is a drastically different conclusion from the earlier work. This highlights how seemingly trivial experimental design choices can yield misleading conclusions about method utility. In addition, our work carries lessons about the necessity of not only evaluating explainable ML methods using tasks, data, users, and metrics grounded in the intended application context but also developing methods tailored to specific applications, moving beyond general-purpose explainable ML methods.

Rental assistance programs provide individuals with financial assistance to prevent housing instabilities caused by evictions and avert homelessness. Since these programs operate under resource constraints, they must decide who to prioritize. Typically, funding is distributed by a reactive allocation process that does not systematically consider risk of future homelessness. We partnered with Anonymous County (PA) to explore a proactive and preventative allocation approach that prioritizes individuals facing eviction based on their risk of future homelessness. Our ML models, trained on state and county administrative data accurately identify at-risk individuals, outperforming simpler prioritization approaches by at least 20% while meeting our equity and fairness goals across race and gender. Furthermore, our approach would reach 28% of individuals who are overlooked by the current process and end up homeless. Beyond improvements to the rental assistance program in Anonymous County, this study can inform the development of evidence-based decision support tools in similar contexts, including lessons about data needs, model design, evaluation, and field validation.

Deep Neural Networks (DNNs) have limited ability to explain their acquired knowledge or decision rationale. As a result, end-users perceive DNNs as black-boxes and are hesitant to fully adopt them in safety-critical applications. Therefore, developing explainable DNNs has become a prime interest in neural network research. This paper presents a methodology for linguistically explaining the knowledge a DNN classifier has acquired in training. The main objective is to help users understand what the DNN has learned about each class. The presented methodology is fuzzy logic based and involves end-users of the system in the explanation process, enabling users to customize the explanations to match their requirements. This paper presents the explanation methodology, metrics of explanation quality, validation steps, and a discussion of advantages and limitations. The explanation methodology was implemented on a benchmark classification problem. Experimental results demonstrated the method's capability to explain the DNN-knowledge and validated the explanations.

A viable solution for increasing fuel efficiency in vehicles is optimizing driver behavior. In our previous work, we proposed a data-driven Intelligent Driver System (IDS), which calculated an optimal driver behavior profile for a fixed route. During operation, the optimal behavior was prompted to the drivers to guide their behavior toward improving fuel efficiency. This system was proposed for fleet vehicles mainly because a small increase in fuel efficiency of fleet vehicles has a significant impact on the economy. The system was tested on a portion of the fleet's route (12km) and achieved 9-20% of fuel saving. One limitation of the IDS was that the prompted behavior profile was the same for all drivers. However, the approach of driving is significantly different from driver to driver. Therefore, it is important to capture those differences in the optimal behavior profile creation and prompting. This paper presents the first steps of a modified IDS that incorporates different approaches of drivers in optimal behavior profile creation. This work has three main components: 1) analyzing the capability of scaling our previously proposed IDS to the complete route of the fleet, 2) assessing the capability of identifying different types of driver behavior from data, and 3) proposing an IDS framework for integrating different driver behavior in optimizing driver behavior. Experimental results showed that the existing IDS was able to achieve 26-37% estimated fuel savings on the complete route. Conclusions of the paper are: 1)the existing IDS scaled to longer routes, and 2) It is possible to identify different driver behavior using data.

The deep self-organizing map (DSOM) was introduced to embed hierarchical feature abstraction capability to self-organizing maps (SOMs). This paper presents an extended version of the original DSOM algorithm (E-DSOM). E-DSOM enhances the DSOM in two ways-learning algorithm is modified to be completely unsupervised, and architecture is modified to learn features of different resolution in hidden layers. E-DSOM has three main advantages over the original DSOM: 1) improved classification accuracy; 2) improved generalization capability; and 3) need of fewer sequential layers (reduced training time). E-DSOM was tested on benchmark and real-world datasets and was compared against DSOM, SOM, sStacked autoencoder (AE), and stacked convolutional autoencoder (CAE). Experimental results showed that the E-DSOM outperformed DSOM with improvements of classification accuracy up to 15% while saving training time up to 19% on all datasets. Moreover, E-DSOM evidenced better generalization capability compared to the DSOM by showing superior performance on all datasets with induced noise. Further, E-DSOM showed comparable performance to the AE and the CAE while outperforming them on two datasets.

Anomaly detection in industrial processes is crucial for general process monitoring and process health assessment. Deep Neural Networks (DNNs) based anomaly detection has received increased attention in recent work. Albeit their high accuracy, the black-box nature of DNNs is a drawback in practical deployment. Especially in industrial anomaly detection systems, explanations of DNN detected anomalies are crucial. This paper presents a framework for DNN based anomaly detection which provides explanations of detected anomalies. The framework answers the following questions during online processing: 1) “why is it an anomaly?” and 2) “what is the confidence?” Further, the framework can be used offline to evaluate the “knowledge” of the trained DNN. The framework reduces the opaqueness of the DNN based anomaly detector and thus improves human operators' trust in the algorithm. This paper implements the first steps of the presented framework on the benchmark KDD-NSL dataset for Denial of Service (DoS) attack detection. Offline DNN explanations showed that the DNN was detecting DoS attacks based on features indicating destination of connection, frequency and amount of data transferred while showing an accuracy around 97%.

Visual data mining facilitates the involvement of domain experts in the data mining processes. The effectiveness of visual data mining is especially dominant when paired with unsupervised methods due to the abundance of unlabeled data. Deep Self-Organizing Maps (DSOMs) are unsupervised learning architectures capable of high level feature abstraction. In this paper, we analyze the effectiveness of using DSOMs for visual data mining. DSOM's visual data mining capability was evaluated using the following visual data explorations methodologies: 1) U-Matrix, 2) hit maps and 3) data histograms. In comparison with traditional single layered SOM architectures, experimental results showed that DSOMs produced more accurate visual representations of the underlying data distributions. Therefore, DSOM is a viable method for generating easily understandable visual representations of high-dimensional complex datasets. These visual representations can be powerful tools in the real world, leading to better understanding of systems and thus enabling the design of better algorithms for control and monitoring.

Deep Neural Networks based intrusion detection systems (DNN-IDS) have proven to be effective. However, in domains like critical infrastructure security, user trust on the DNN-IDS is imperative and high accuracy isn't sufficient. The black-box nature of DNNs hinders transparency of the DNN-IDS, which is necessary for building trust. The main objective of this work is to improve user trust by improving transparency of the DNN-IDS by making it more communicative. This paper presents a methodology to generate offline and online feedback to the user on the decision making process of the DNN-IDS. Offline, the user is reported the input features that are most relevant in detecting each type of intrusion by the trained DNN-IDS. Online, for each detection, the user is reported the inputs features that contributed most to the detection. The presented method was implemented on the KDD-NSL dataset with a multi-layer perceptron (MLP) based DNN-IDS. Binary and multi-class classification was carried out on the dataset. Further, several DNN-IDS architectures with different depth were tested to study the factors that drive classification. It was observed that despite showing very similar accuracy results, the factors that drove the decisions were different across architectures. This evidences that the qualitative analysis that is enabled through reporting relevant input features is important for the user to make a more informed decision in choosing a DNN-IDS. This online and offline feedback leads to improving the transparency of the DNN-IDS and helps build trust prior to and during deployment.

Cyber-Physical Systems (CPSs)have become ubiquitous in recent years and has become the core of modern critical infrastructure and industrial applications. Therefore, ensuring security is a prime concern. Due to the success of Deep Learning (DL)in a multitude of domains, development of DL based CPS security applications have received increased interest in the past few years. Developing generalized models is critical since the models have to perform well under threats that they havent trained on. However, despite the broad body of work on using DL for ensuring the security of CPSs, to our best knowledge very little work exists where the focus is on the generalization capabilities of these DL applications. In this paper, we intend to provide a concise survey of the regularization methods for DL algorithms used in security-related applications in CPSs and thus could be used to improve the generalization capability of DL based cyber-physical system based security applications. Further, we provide a brief insight into the current challenges and future directions as well.

Epileptic seizure source identification involves neurologists combing through a substantial amount of data manually, which sometimes takes weeks per patient. This paper presents a methodology for minimizing the amount of data a neurologist has to analyze to identify the seizure focus. The method keeps the neurologist as the final decision maker and aids in the decision making process. It has to be noted that the primary focus of the work was not improving the accuracy of interictal spike detection but reduction of the volume of data. The presented methodology is based on Artificial Neural Networks (ANN) and is implemented on EEG data collected on 5 patients using a dense array EEG reader. As a baseline, a simple template matching was implemented on the same dataset. Experimental results showed that the ANN based methodology was able to reduce the dataset by 98%, a significant improvement on the template matching method.

Control systems monitor and command other devices, systems, and software within an infrastructure. Typically, control systems employ human-in-the-loop control for critical decision making and response. These end-users require easy access to accurate, actionable and relevant data to ensure quick and effective decision making. This work presents a framework for creating dynamic visual interfaces for improved situational awareness. The proposed framework determines the relevance of available information pieces and then applies the derived relevance scores to a visualization so that the most relevant and important information are emphasized to the end-users. In the presented work, a priori expert knowledge is encoded in the system through the use of Fuzzy Logic (FL) and the resulting FL inference system assigns scores to information pieces based on system state information and user defined relevance. These scores can then be used to organize and display the relevant data given the current situation and end-user roles. The proposed FL based scoring system was implemented on a real world control system dataset and we demonstrate how the information visualization is dynamically adapted to improve situational awareness. Further, we discuss potential methods the relevance scores can be incorporated into real world visualizations to increase the situational awareness in control systems.

Machine control using electroencephalography (EEG) based brain computer interfaces (BCI) has been extensively researched in the past decade. However, research is often based on event bound methods such as motor imagery. Despite being useful in medical applications, even bound methods limit users' operational capability while performing BCI control. To alleviate the said limitation, we explore a robot control framework based on abstract thought. Abstract thought in this context is defined as conscious mental tasks that are not bound with any particular event or bodily movement. This paper presents an initial step in the framework, which is a methodology for optimal feature selection for abstract thought EEG data classification. The presented method contains 2 steps: 1) generational Genetic Algorithm (GA) based feature selection, and, 2) EEG data classification using selected features. The presented method was implemented on an EEG dataset acquired from a consumer grade EEG device. Abstract thought EEG data were collected for three actions pertaining to robot control; 1) “rest”, 2) “move forward”, and, 3) “turn left”. The presented method was compared to EEG classification without any feature selection. Experimental results showed that the presented method outperformed the method without feature selection for all the tested classifiers with a 10% or higher improvement in classification accuracy.

The efficient use of energy is one method of potentially reducing the amount of fossil fuels that are used world-wide to generate electricity. Nuclear energy presents a promising sustainable energy source for the future that generates few to no greenhouse gases. While it creates energy that can be used for electricity it also has many byproducts that are treated as waste or not utilized to their full potential. Waste heat is a major product of nuclear energy that should be harnessed and applied to other processes. These include water desalination/water purification hydrogen production, use in industrial chemical operations, or electricity production. Nuclear energy is most often base load following power and very inflexible. One way to address this is to use Hybrid Energy Systems (HES). By doing this the system can be load following and hence more efficient and sustainable. Therefore, the goal of this project is to create a system that mimics the waste heat from a reactor, demonstrate how to utilize that heat, and show that when energy demand is low the reactor does not need to reduce power; the energy can be directed elsewhere to create goods. This paper presents the details of the Energy Conversion Loop (ECL) that was developed to act as a test bed for experimentation of the aforementioned processes and test the possibility of nuclear HES. Further, the paper presents a proposed intelligent control system that can adapt to the system requirements for the ECL. In addition, the system has great potential for education on critical infrastructure protection and testing of future control logic systems and mechanical systems.

Ensuring sustainability demands more efficient energy management with minimized energy wastage. Therefore, the power grid of the future should provide an unprecedented level of flexibility in energy management. To that end, intelligent decision making requires accurate predictions of future energy demand/load, both at aggregate and individual site level. Thus, energy load forecasting have received increased attention in the recent past. However, it has proven to be a difficult problem. This paper presents a novel energy load forecasting methodology based on Deep Neural Networks, specifically, Long Short Term Memory (LSTM) algorithms. The presented work investigates two LSTM based architectures: 1) standard LSTM and 2) LSTM-based Sequence to Sequence (S2S) architecture. Both methods were implemented on a benchmark data set of electricity consumption data from one residential customer. Both architectures were trained and tested on one hour and one-minute time-step resolution datasets. Experimental results showed that the standard LSTM failed at one-minute resolution data while performing well in one-hour resolution data. It was shown that S2S architecture performed well on both datasets. Further, it was shown that the presented methods produced comparable results with the other deep learning methods for energy forecasting in literature.

Heating, Ventilation and Air Conditioning (HVAC) system is largest energy consumer in buildings. Worldwide, buildings consume 20% of the total energy production. Therefore, increasing efficiency of the HVAC system will result in significant financial savings. As one solution, Thermal Energy Storage (TES) tanks are being utilized with buildings to store excess energy to be reused later. An optimal control strategy is crucial for optimal usage. Therefore, this paper presents a novel control framework based on Artificial Neural Networks (ANN) for optimally controlling a TES for achieving increased savings. The presented ANN controller utilizes 3 main inputs: 1) current TES energy availability, 2) predicted building power requirement, and 3) predicted utility load/price. In addition to the design details of the control framework, this paper presents implementation details of the ANN controller. Further, experiments on several test cases were carried out and the paper presents the experimental setup and obtained results for each test case. Performance of the presented ANN control framework was compared against a classical proportional derivative (PD) controller. It was observed that the presented framework resulted in better cost savings than the classical controller consistently for all the experimental test cases.

Brain Computer Interfaces (BCI) have gained significant interest over the last decade as viable means of human machine interaction. Although many methods exist to measure brain activity in theory, Electroencephalography (EEG) is the most used method due to the cost efficiency and ease of use. However, thought pattern based control using EEG signals is difficult due two main reasons; 1) EEG signals are highly noisy and contain many outliers, 2) EEG signals are high dimensional. Therefore the contribution of this paper is a novel methodology for recognizing thought patterns based on Self Organizing Maps (SOM). The presented thought recognition methodology is a three step process which utilizes SOM for unsupervised clustering of pre-processed EEG data and feed-forward Artificial Neural Networks (ANN) for classification. The presented method was tested on 5 different users for identifying two thought patterns; “move forward” and “rest”. EEG Data acquisition was carried out using the Emotiv EPOC headset which is a low cost, commercial-off-the-shelf, noninvasive EEG signal measurement device. The presented method was compared with classification of EEG data using ANN alone. The experimental results for the 5 users chosen showed an improvement of 8% over ANN based classification.