Municipal solid waste incineration (MSWI) technology has developed rapidly worldwide. Carbon monoxide (CO) is one of the to be controlled key operating index of such processes. CO emission concentration prediction is a challenge problem duo to its large fluctuation range. A new CO emission concentration prediction method based on concept drift detection using kernel principal component analysis (KPCA) is proposed. The proposed approach includes off-line model construction module, on-line concept drift detection prediction and updating module. First, we construct the LSTM-based CO prediction model using historical data and KPCA-based concept drift detection model for calculating the evaluation index. Then, recursive KPCA is used to adaptive monitor the concept drift of the time-varying process. Finally, based on continuous updating of the historical LSTM mode with the concept drift samples, we achieve higher prediction accuracy. The rationality and validity are verified with the actual data of MSWI processes.
Credit score prediction is a crucial task in financial industry, as it helps lenders and financial institutions evaluate the creditworthiness of borrowers and manage credit risk. In this work, we present a comparative study of deep learning (DL)-based credit score prediction models. To achieve this objective, we compare the performance of DL models against traditional methods in credit scoring. We train and test the models using a real-world dataset of credit histories, containing various features such as credit card balances, payment history, and employment status. Our experimental results show that the hybrid LSTM-GRU model outperform both the LSTM and GRU models in credit score prediction, as well as traditional methods. The hybrid LSTM-GRU model demonstrates higher accuracy and better predictive power, indicating its potential for improving credit scoring models in the financial industry.
Driver behavior plays a crucial role in mitigating traffic accidents. Unlike previous studies focusing on drivers' actual maneuvers, our study examines actions preceding a maneuver or those not manifest in conventional maneuvers. We gathered data from a car-sharing service frequented by university students, scrutinizing the frequency of pedal changes via foot camera images. These data were compared with the pedal depression measurements from Controller Area Network (CAN) bus data to discern instances of potentially reduced safety due to cognitive load. We identified locations with high pedal change frequency, even in the absence of any recorded pedal operation in the CAN bus data. This result suggests the existence of unrecorded driver behaviors that could precipitate traffic accidents. Our findings, therefore, bear substantial implications for enhancing traffic safety measures.
Cardiovascular diseases are the leading cause of death in the world, with arrhythmias being a significant symptom and risk factor. Advancements in technologies such as low-cost and low-power wearable devices, and machine learning techniques for analyzing big volumes of data offer opportunities to address this issue. However, low-cost devices may have limitations, including reduced data quality due to lower sampling rates, bit depth, and the number of leads recorded. These limitations might produce a significant decrease in machine learning models' performance in detecting arrhythmias. This study investigates the impact of data quality reduction on arrhythmia classification using deep neural network models. High-quality ECG data with 12 leads, 500Hz sampling rate, and 32-bits resolution were transformed into low-quality versions with varying leads (from one to six), 100Hz sampling rate, and 8-bits resolution. Training a state-of-the-art deep learning arrhythmia detection model on both high-quality and low-quality datasets revealed a decrease in performance from 95.3% to 93.9% in the worst case, which is concerning given the critical nature of the domain. To mitigate this performance loss, we propose an ensembling method that compensates for 42% of the loss, achieving an accuracy of 94.5% even with the low-quality dataset. The analysis also identifies the leads with the most promising classification performances. These results can aid in making better design decisions when creating cost-effective wearable ECG devices.
In an increasingly digitalized commerce landscape, the proliferation of credit card fraud and the evolution of sophisticated fraudulent techniques have led to substantial financial losses. Automating credit card fraud detection is a viable way to accelerate detection, reducing response times and minimizing potential financial losses. However, addressing this challenge is complicated by the highly imbalanced nature of the datasets, where genuine transactions vastly outnumber fraudulent ones. Furthermore, the high number of dimensions within the feature set gives rise to the “curse of dimensionality". In this paper, we investigate subspace learning-based approaches centered on One-Class Classification (OCC) algorithms, which excel in handling imbalanced data distributions and possess the capability to anticipate and counter the transactions carried out by yet-to-be-invented fraud techniques. The study highlights the potential of subspace learning-based OCC algorithms by investigating the limitations of current fraud detection strategies and the specific challenges of credit card fraud detection. These algorithms integrate subspace learning into the data description; hence, the models transform the data into a lower-dimensional subspace optimized for OCC. Through rigorous experimentation and analysis, the study validated that the proposed approach helps tackle the curse of dimensionality and the imbalanced nature of credit card data for automatic fraud detection to mitigate financial losses caused by fraudulent activities.
In today's world, sensors play a crucial role, as they feed information to make accurate decisions and take actions; therefore, making sure that sensors behave correctly is critical. In this work, we focus on inspecting the data provided by sensors, aiming at discovering any issue due to malfunction, misuse, or any other source of error before the issue is propagated through the system. To achieve that, we propose a novel approach based on wavelets embedded in a microcontroller to analyze data from sensors. The objective is to determine whether the sensor is issuing correct data (normal behavior) or not (abnormal behavior), to prevent the error from reaching other parts of the system.
We study Off-Policy Evaluation (OPE) in contextual bandit settings with large action spaces. The benchmark estimators suffer from severe bias and variance tradeoffs. Parametric approaches suffer from bias due to difficulty specifying the correct model, whereas ones with importance weight suffer from variance. To overcome these limitations, Marginalized Inverse Propensity Scoring (MIPS) was proposed to mitigate the estimator's variance via embeddings of an action. Nevertheless, MIPS is unbiased under the no direct effect, which assumes that the action embedding completely mediates the effect of an action on a reward. To overcome the dependency on these unrealistic assumptions, we propose a Marginalized Doubly Robust (MDR) estimator. Theoretical analysis shows that the proposed estimator is unbiased under weaker assumptions than MIPS while reducing the variance against MIPS. The empirical experiment verifies the supremacy of MDR against existing estimators with large action spaces.
Algorithmic trading is a topic with major developments in the last years. Investors rely mostly on indicators derived from fundamental (FA) or technical analysis (TA), while sentiment analysis (SA) has also received attention in the last decade. This has led to great financial advantages with algorithms being the main tool to create pre-programmed trading strategies. Although the three analysis types have been mainly considered individually, their combination has not been studied as much. Given the ability of each individual analysis type in identifying profitable trading strategies, we are motivated to investigate if we can increase the profitability of such strategies by combining their indicators. Thus, in this paper we propose a novel Genetic Programming (GP) algorithm that combines the three analysis types and we showcase the advantages of their combination in terms of three financial metrics, namely Sharpe ratio, rate of return and risk. We conduct experiments on 30 companies and based on the results, the combination of the three analysis types statistically and significantly outperforms their individual results, as well as their pairwise combinations. More specifically, the proposed GP algorithm has the highest mean and median values for Sharpe ratio and rate of return, and the lowest (best) mean value for risk. Moreover, we benchmark our GP algorithm against multilayer perceptron and support vector machine, and show that it statistically outperforms both algorithms in terms of Sharpe ratio and risk.
Concept drift refers to a change in the data distribution affecting the data stream of future samples. Consequently, learning models operating on the data stream might become obsolete, and need costly and difficult adjustments such as retraining or adaptation. Existing methods usually implement a local concept drift adaptation scheme, where either incremental learning of the models is used, or the models are completely retrained when a drift detection mechanism triggers an alarm. This paper proposes an alternative approach in which an unsupervised and model-agnostic concept drift adaptation method at the global level is introduced, based on autoencoders. Specifically, the proposed method aims to ``unlearn'' the concept drift without having to retrain or adapt any of the learning models operating on the data. An extensive experimental evaluation is conducted in two application domains. We consider a realistic water distribution network with more than 30 models in-place, from which we create 200 simulated data sets / scenarios. We further consider an image-related task to demonstrate the effectiveness of our method.
Incomplete or sparse non-dominated fronts are unavoidable in multi-objective optimization due to complexity of problems, morphology of Pareto optimal fronts, and stochasticity involved in evolutionary optimization algorithms. It is pragmatic to develop methods that can alleviate some of these issues after the optimization run is complete, without the need for re-optimization or additional solution evaluations. Previously developed methods demonstrated that it is possible to predict Pareto-optimal solutions from pseudo-weight vectors using Gaussian Process Regression (GPR) models. We extend the GPR-based method to predict new Pareto-optimal solutions using reference vectors as unique identifiers and demonstrate that like the pseudo-weight vectors, reference vectors can also used instead in learning the association between identifiers and corresponding variable vectors. Results on many test problems indicate that the choice of a suitable identifier makes a large impact on the decision-making process, particularly for visualizing the newly created non-dominated (ND) solutions. In this study, we discuss the advantages and disadvantages of using pseudo-weights and reference vectors as unique identifiers for ND solutions, paving the way to devise further identifiers for predicting new Pareto-optimal solutions.
The last decade has seen significant changes in the power grid complexity due to the increased integration of multiple heterogeneous distributed energy resources. Accurate and fast power flow analysis tools have then become essential to guarantee grid stability, reliable operation, strategic planning, and market strategies. State-of-the-art approaches to power flow analysis are based on iterative numerical techniques which exhibit high accuracy but slow-, or even no-, convergence. For this reason, researchers have investigated the use of data-driven techniques that, while exhibiting lower accuracy with respect to iterative numerical ones, have the advantage of being extremely fast. To address the lack of accuracy, physics-informed data-driven techniques, i.e., techniques that leverage both the data and domain knowledge to generate simultaneously fast and accurate models, have been proposed. Nevertheless, these works exhibit two main limitations: i) they do not fully leverage the physical knowledge, and ii) they do not fairly compare the different approaches. In this paper, we propose a novel physics informed data-driven model able to address both limitations by fully leveraging the physical knowledge into the data-driven, i.e., constraining the model and augmenting the available data, and proposing a framework able to fairly compare the different approaches proving the actual effectiveness of the proposal. Results on the IEEE 57 realistic power network will support the proposal.
Training Artificial Neural Networks poses a challenging and critical problem in machine learning. Despite the effectiveness of gradient-based learning methods, such as Stochastic Gradient Descent (SGD), in training neural networks, they do have several limitations. For instance, they require differentiable activation functions, and cannot optimize a model based on several independent non-differentiable loss functions simultaneously; for example, the F1-score, which is used during testing, can be used during training when a gradient-free optimization algorithm is utilized. Furthermore, the training in any DNN can be possible with a small size of the training dataset. To address these concerns, we propose an efficient version of the gradient-free Coordinate Search (CS) algorithm, an instance of General Pattern Search methods, for training neural networks. The proposed algorithm can be used with non-differentiable activation functions and tailored to multi-objective/multi-loss problems. Finding the optimal values for weights of ANNs is a large-scale optimization problem. Therefore instead of finding the optimal value for each variable, which is the common technique in classical CS, we accelerate optimization and convergence by bundling the weights. In fact, this strategy is a form of dimension reduction for optimization problems. Based on the experimental results, the proposed method, in some cases, outperforms the gradient-based approach, particularly, in situations with insufficient labeled training data. The performance plots demonstrate a high convergence rate, highlighting the capability of our suggested method to find a reasonable solution with fewer function calls. As of now, the only practical and efficient way of training ANNs with hundreds of thousands of weights is gradient-based algorithms such as SGD or Adam. In this paper we introduce an alternative method for training ANN.
Hopfield networks are an attractive choice for solving many types of computational problems because they provide a biologically plausible mechanism. The Self-Optimization (SO) model adds to the Hopfield network by using a biologically founded Hebbian learning rule, in combination with repeated network resets to arbitrary initial states, for optimizing its own behavior towards some desirable goal state encoded in the network. In order to better understand that process, we demonstrate first that the SO model can solve concrete combinatorial problems in SAT form, using two examples of the Liars problem and the map coloring problem. In addition, we show how under some conditions critical information might get lost forever with the learned network producing seemingly optimal solutions that are in fact inappropriate for the problem it was tasked to solve. What appears to be an undesirable side-effect of the SO model, can provide insight into its process for solving intractable problems.
Deep learning and federated learning show significant success in cybersecurity for Intrusion Detection Systems (IDS). This paper presents the Federated Self-Supervised Learning (FSSL) framework proposed for IDSs. FSSL combines Self-Supervised Learning (SSL) with federated learning to obtain a global model. SSL works at the client level, where only unlabeled data is available, and thus it enables the learning from these data. This knowledge enhances the training of the target model. Therefore, FSSL follows a federated learning approach, where private data from multiple clients help to create a global model. Each client learns an unsupervised model, which is then transmitted to a server and combined into a single model. The communication between clients and the server aims to improve model performance and convergence. Conducted experiments compare FSSL with a baseline approach using limited data and a deep learning model. FSSL utilizes an autoencoder to learn a representational model on unlabeled data and transfers knowledge by initializing deep learning model weights with the encoder layers. Results show that FSSL significantly improves the F1-Score of detection systems across three well-known datasets (NSL-KDD, TonIoT, and BotIoT). Moreover, the proposed model demonstrated a noteworthy capability to detect previously unidentified attacks when compared to the baseline.
This study investigates how the introduction of morphological novelty affects the fitness and diversity of a population of modular robots. Novelty is usually measured in behavioral space, while the approach discussed in this paper assesses novelty solely using morphologies. The proposed algorithm is inspired by the histogram of oriented gradients, in combination with elements of principal component analysis, and the Wasserstein distance. The experiments conducted utilize novelty in parent selection, with different configurations. Analyzing the results, the introduction of morphological novelty promotes beneficial effects on fitness and diversity when applied correctly.
We propose a new swarm intelligence numerical optimization algorithm that represents individuals as dynamic graphs in the Euclidean search space. We call it Graph Building Optimization Algorithm or GBO. The unique point of GBO is that an individual is represented by a dynamic graph whose nodes have coordinates (search points) in the Euclidean search space. Due to this unique point, we can draw a GBO's search process as a generation-transition of a feature of a graph. It is expected that we can obtain better understandings on a given problem by comparing the generation-transition for the given problem to the baseline for the simplest unimodal problem. We assume the maximum node degree in the best individual as the feature and the generation-transition of the feature for F1 in the CEC'13 test problems as the baseline. We demonstrate that we can guess the characteristics of other 27 problems in the CEC'13 test problems by comparing their generation-transitions to the baseline. In addition, we evaluate GBO using the same problems and show that GBO is capable of finding good solutions for various problems.
In this work we present a novel approach for generating cardiovascular data using a modified WaveNet architecture. This can enable further research in areas where data is scarce and hard to obtain. By generating additional time series data in a set of animal tests performance of existing models could be improved and more difficult approaches, that require substantial amounts of data, attempted. We validate our approach on a classification task and compare it to similar methods of data augmentation.
AI technologies have made significant advancements across various sectors, especially healthcare. Although AI algorithms in healthcare showcase remarkable predictive capabilities, apprehensions have emerged owing to errors, biases, and a lack of transparency. These concerns have led to a decline in trust among clinicians and patients, while also posing the risk of further accentuating pre-existing biases against marginalized groups and exacerbating inequities. This paper presents a scenario-based preferences risk register 1 1 Denotes a methodically arranged document or database detailing potential risks linked to particular scenarios or situations. framework for identifying and accounting AI algorithm biases in diagnosing diseases. The framework is demonstrated with a realistic case study on cardiac sarcoidosis. The framework identifies success criteria, initiatives, emergent conditions and the most and least disruptive scenarios. The success criteria align with the National Institute of Standards and Technology AI Risk Management Framework (NIST AI RMF) trustworthy AI characteristics, and the scenarios are based on various statistical/computational bias that causes algorithmic bias. The framework provides valuable guidance for leveraging AI in healthcare, enhancing objective designs, and mitigating risks by adopting a figure of merit to score the initiatives and measuring the disruptive order. By prioritizing transparency, trustworthy AI, and identifying the most and least disruptive scenarios/biases, the framework promotes responsible and effective use of AI technologies in healthcare.
The hand has excellent functional, aesthetic and social importance. However, Parkinson's disease, stroke, and other myopathies can cause motor impairments. Patients require a rehabilitation program to follow their progress, and one of the tools used to do that is the electromyographic (EMG) signals. This article proposes using algorithms to characterize and classify EMG signals during force exercises for the wrist and forearm. Eight healthy subjects participated in this study. They performed seven exercises, making five trials for each one. Signal features were analyzed in different time windows using a genetic algorithm and machine learning techniques to select the window that maximizes the classification. Combining four electrodes, seven exercises, and 14 algorithms achieved a classification accuracy of 92.41 % using the Multilayer Perceptron classifier. The study demonstrates a highly reliable method for classifying forearm and wrist exercises based on EMG signals, useful for exoskeletons or rehabilitation platforms. Future work will focus on implementing EMG signals to enhance motor rehabilitation therapy and provide findings that will help the scientific community investigate the combination of EEG signals for rehabilitation purposes.
The fusion of inertial and visual data is an effective approach to human motion analysis, with applications in areas such as sports or rehabilitation exercise monitoring. Employing wireless, low-cost, external inertial sensors and a built-in camera on mobile devices provides a convenient acquisition system, available for wide range of potential users. In order to take advantage of both data modalities, robust time synchronization is required. We consider consumer-grade devices, for which direct access to internal clocks is not available and only high-level API is provided. At the same time, we aim to avoid event-based synchronization that would require additional user actions. We investigate sources of acquisition errors on mobile devices, and then we propose and evaluate a novel synchronization method for inertial and visual data. Experimental results indicate that the proposed method provides robust synchronization.
