Underlying Machine Learning Models Research Articles

A novel approach for assessing fairness in deployed machine learning algorithms

Fairness in machine learning (ML) emerges as a critical concern as AI systems increasingly influence diverse aspects of society, from healthcare decisions to legal judgments. Many studies show evidence of unfair ML outcomes. However, the current body of literature lacks a statistically validated approach that can evaluate the fairness of a deployed ML algorithm against a dataset. A novel evaluation approach is introduced in this research based on k-fold cross-validation and statistical t-tests to assess the fairness of ML algorithms. This approach was exercised across five benchmark datasets using six classical ML algorithms. Considering four fair ML definitions guided by the current literature, our analysis showed that the same dataset generates a fair outcome for one ML algorithm but an unfair result for another. Such an observation reveals complex, context-dependent fairness issues in ML, complicated further by the varied operational mechanisms of the underlying ML models. Our proposed approach enables researchers to check whether deploying any ML algorithms against a protected attribute within datasets is fair. We also discuss the broader implications of the proposed approach, highlighting a notable variability in its fairness outcomes. Our discussion underscores the need for adaptable fairness definitions and the exploration of methods to enhance the fairness of ensemble approaches, aiming to advance fair ML practices and ensure equitable AI deployment across societal sectors.

Using machine learning to develop smart reflex testing protocols.

Reflex testing protocols allow clinical laboratories to perform second line diagnostic tests on existing specimens based on the results of initially ordered tests. Reflex testing can support optimal clinical laboratory test ordering and diagnosis. In current clinical practice, reflex testing typically relies on simple "if-then" rules; however, this limits the opportunities for reflex testing since most test ordering decisions involve more complexity than traditional rule-based approaches would allow. Here, using the analyte ferritin as an example, we propose an alternative machine learning-based approach to "smart" reflex testing. Using deidentified patient data, we developed a machine learning model to predict whether a patient getting CBC testing will also have ferritin testing ordered. We evaluate applications of this model to reflex testing by assessing its performance in comparison to possible rule-based approaches. Our underlying machine learning models performed moderately well in predicting ferritin test ordering (AUC=0.731 in reference to actual ordering) and demonstrated promising potential to underlie key clinical applications. In contrast, none of the many traditionally framed, rule-based, hypothetical reflex protocols we evaluated offered sufficient agreement with actual ordering to be clinically feasible. Using chart review, we further demonstrated that the strategic deployment of our model could avoid important ferritin test ordering errors. Machine learning may provide a foundation for new types of reflex testing with enhanced benefits for clinical diagnosis.

A WEB APPLICATION FOR REAL-TIME PHISHING WEBSITE DETECTION

The rapid proliferation of the internet has led to an alarming increase in cyber threats, particularly phishing attacks. Phishing websites impersonate legitimate ones, luring users into divulging sensitive information such as login credentials and financial details. Detecting and preventing such attacks are crucial for safeguarding user privacy and security. This paper presents the development of a web application for real-time phishing website detection. The application employs machine learning techniques, specifically the XGBoost algorithm, to analyze the characteristics of URLs and classify them as either legitimate or phishing. It offers users a simple and effective way to check the authenticity of websites before interacting with them.The web application provides an intuitive user interface, allowing users to enter a URL and receive an instant assessment of its legitimacy. The underlying machine learning model is trained on a diverse dataset of legitimate and phishing URLs, ensuring robust detection capabilities. Additionally, the application provides insights into the features used for classification, enhancing user awareness of potential threats. The proposed web application serves as a valuable tool in the fight against phishing attacks, empowering users to make informed decisions when navigating the web. Its real-time detection capabilities and user-friendly interface make it accessible to a wide audience, contributing to a safer online environment. KEYWORDS Web Application, Phishing Detection, Machine Learning, XGBoost, User Privacy.

REALITY-ENHANCED SYNTHETIC IMAGE TRAINING DATASET FOR COMPUTER VISION CONSTRUCTION MONITORING

Recent growth in the availability of aerial and ground-level images from construction sites has created a surge in computer vision-based techniques for construction monitoring applications. To address the need for large volumes of visual data with ground truth for training the underlying machine learning models, the application of synthetic data, particularly from 3D/4D BIM, has gained traction in the research community. However, the existing domain gap between real and synthetic images, such as rendering noise from repeated material texture patterns, randomized positional camera parameters that collect the ground truth of BIM elements, and the poorly visibility of BIM elements to the camera viewpoints within these environments, has negatively impacted the potential use of these synthetic datasets as scale. To address these limitations, this paper presents a new synthetic image generation pipeline that optimizes and integrates camera extrinsic parameters with a novel synthetic image appearance enhancement technique to generate high volume of quality synthetic data with respective ground-truth. The use of synthetic datasets for training progress monitoring models is validated through several real-data segmentation cases that incorporate: 1) the automatic collection of synthetic images and ground-truth annotations from high-LoD BIM model disciplines, 2) optimization of positional camera parameters using element visibility metrics, and 3) the enhancement of realism of synthetic images using a patch-based generative approach. The benefits and the current limitations to automated progress monitoring, as well as robotic path automation and optimization for progress motoring real data collection are discussed.

Membership Inference Attack and Defense for Wireless Signal Classifiers With Deep Learning

An over-the-air membership inference attack (MIA) is presented to leak private information from a wireless signal classifier. Machine learning (ML) provides powerful means to classify wireless signals, e.g., for PHY-layer authentication. As an adversarial machine learning attack, the MIA infers whether a signal of interest has been used in the training data of a target classifier. This private information incorporates waveform, channel, and device characteristics, and if leaked, can be exploited by an adversary to identify vulnerabilities of the underlying ML model (e.g., to infiltrate the PHY-layer authentication). One challenge for the over-the-air MIA is that the received signals and consequently the RF fingerprints at the adversary and the intended receiver differ due to the discrepancy in channel conditions. Therefore, the adversary first builds a surrogate classifier by observing the spectrum and then launches the black-box MIA on this classifier. The MIA results show that the adversary can reliably infer signals (and potentially the radio and channel information) used to build the target classifier. Therefore, a proactive defense is developed against the MIA by building a shadow MIA model and fooling the adversary. This defense can successfully reduce the MIA accuracy and prevent information leakage from the wireless signal classifier.

Bridging the gap between mechanistic biological models and machine learning surrogates.

Mechanistic models have been used for centuries to describe complex interconnected processes, including biological ones. As the scope of these models has widened, so have their computational demands. This complexity can limit their suitability when running many simulations or when real-time results are required. Surrogate machine learning (ML) models can be used to approximate the behaviour of complex mechanistic models, and once built, their computational demands are several orders of magnitude lower. This paper provides an overview of the relevant literature, both from an applicability and a theoretical perspective. For the latter, the paper focuses on the design and training of the underlying ML models. Application-wise, we show how ML surrogates have been used to approximate different mechanistic models. We present a perspective on how these approaches can be applied to models representing biological processes with potential industrial applications (e.g., metabolism and whole-cell modelling) and show why surrogate ML models may hold the key to making the simulation of complex biological systems possible using a typical desktop computer.

CAML: Contextual augmented meta-learning for cold-start recommendation

The performance of recommendation engines is challenged by the cold-start issues. Classical recommendation techniques have limited capability to address this problem because the underlying machine learning models are data-thirsty. Conversely, the recent success of meta-learning attracted a lot of attention due to few-shot learning capabilities which can be employed to accommodate new tasks even when there are few data samples for training. Most of the meta-learning recommendation models adopt model-agnostic meta-learning to initialize parameters that may lead to stuck into local optima instead of global optima for some users. To leverage the learning process, we propose a Contextually Augmented Meta-Learning recommender system (CAML). The proposed method augments the contextual features into a meta-learning model which considerably improves the tasks adaptation capability. We constructed a Data Augmentation Unit (DAU) that used a hybrid similarity to augment data samples of similar neighbors. The augmented samples are then forwarded to a Meta-Learner (MetaL) to learn user preferences and generate relevant recommendations. Additionally, we highlighted that the proposed method is generalized and can be adapted for various meta-learning-based recommendation models. We validate our proposal on three benchmark datasets for both ranking and rating prediction perspectives. The experimental outcomes prove the significance of our proposal over the state-of-the-art recommendation methods.

A nascent design theory for explainable intelligent systems

Due to computational advances in the past decades, so-called intelligent systems can learn from increasingly complex data, analyze situations, and support users in their decision-making to address them. However, in practice, the complexity of these intelligent systems renders the user hardly able to comprehend the inherent decision logic of the underlying machine learning model. As a result, the adoption of this technology, especially for high-stake scenarios, is hampered. In this context, explainable artificial intelligence offers numerous starting points for making the inherent logic explainable to people. While research manifests the necessity for incorporating explainable artificial intelligence into intelligent systems, there is still a lack of knowledge about how to socio-technically design these systems to address acceptance barriers among different user groups. In response, we have derived and evaluated a nascent design theory for explainable intelligent systems based on a structured literature review, two qualitative expert studies, a real-world use case application, and quantitative research. Our design theory includes design requirements, design principles, and design features covering the topics of global explainability, local explainability, personalized interface design, as well as psychological/emotional factors.

Mitigating voltage fingerprint spoofing attacks on the controller area network bus

The Controller Area Network (CAN) bus suffers security vulnerabilities that allow message spoofing and masquerading Electronic Control Units (ECUs). A popular provision for mitigating these vulnerabilities is through the use of machine learning (ML) to derive ECU fingerprints based on the physical properties of bus signals. Particularly, voltage-based intrusion detection systems associate the message transmitter with its voltage fingerprint to detect conflicting logical ECU identifiers in the presence of cyberattacks. However, the signal characteristics depend on the operating conditions and hence the fingerprints need to be adapted overtime by online training of the underlying ML model. An adversary may exploit such a shortcoming to superimpose training data based on its own transmissions and thus bypass the protection mechanism. Such an attack not only allows device impersonation but also leads to rejecting transmissions of a legitimate ECU. This paper proposes an effective approach to thwart these attack scenarios. Our approach introduces unpredictably-scheduled transmissions involving one or multiple ECUs to confuse the adversary and ensure the generation of a legitimate fingerprinting dataset for online training. We validate the robustness of our approach using data collected from a real vehicle and show that it outperforms a prominent competing scheme by over 30% in terms of identifying malicious ECUs when the attacker could overwrite 50% of the retraining transmissions.

Amortized Generation of Sequential Algorithmic Recourses for Black-Box Models

Explainable machine learning (ML) has gained traction in recent years due to the increasing adoption of ML-based systems in many sectors. Algorithmic Recourses (ARs) provide "what if" feedback of the form "if an input datapoint were x' instead of x, then an ML-based system's output would be y' instead of y." Recourses are attractive due to their actionable feedback, amenability to existing legal frameworks, and fidelity to the underlying ML model. Yet, current recourse approaches are single shot that is, they assume x can change to x' in a single time period. We propose a novel stochastic-control-based approach that generates sequential recourses, that is, recourses that allow x to move stochastically and sequentially across intermediate states to a final state x'. Our approach is model agnostic and black box. Furthermore, the calculation of recourses is amortized such that once trained, it applies to multiple datapoints without the need for re-optimization. In addition to these primary characteristics, our approach admits optional desiderata such as adherence to the data manifold, respect for causal relations, and sparsity identified by past research as desirable properties of recourses. We evaluate our approach using three real-world datasets and show successful generation of sequential recourses that respect other recourse desiderata.

Explainable artificial intelligence (XAI) post-hoc explainability methods: risks and limitations in non-discrimination law

Organizations are increasingly employing complex black-box machine learning models in high-stakes decision-making. A popular approach to addressing the problem of opacity of black-box machine learning models is the use of post-hoc explainability methods. These methods approximate the logic of underlying machine learning models with the aim of explaining their internal workings, so that human examiners can understand them. In turn, it has been alluded that the insights from post-hoc explainability methods can be used to help regulate black-box machine learning. This article examines the validity of these claims. By examining whether the insights derived from post-hoc explainability methods in post-model deployment can prima facie meet legal definitions in European (read European Union) non-discrimination law, we argue that machine learning post-hoc explanation methods cannot guarantee the insights they generate.Ultimately, we argue that the use of post-hoc explanatory methods is useful in many cases, but that these methods have limitations that prohibit reliance as the sole mechanism to guarantee fairness of model outcomes in high-stakes decision-making. By way of an ancillary function, the inadequacy of European Non-Discrimination Law for algorithmic decision-making is demonstrated too.

Toward feasible machine learning model updates in network-based intrusion detection

Over the last years, several works have proposed highly accurate machine learning (ML) techniques for network-based intrusion detection systems (NIDS), that are hardly used in production environments. In practice, current intrusion detection schemes cannot easily handle network traffic’s changing behavior over time, requiring frequent and complex model updates to be periodically performed. As a result, unfeasible amounts of labeled training data must be provided for model updates as time passes, making such proposals unfeasible for the real world. This paper proposes a new intrusion detection model based on stream learning with delayed model updates to make the model update task feasible with a twofold implementation. First, our model maintains the intrusion detection accuracy through a classification assessment approach, even with outdated underlying ML models. The classification with a reject option rationale also allows suppressing potential misclassifications caused by new network traffic behavior. Second, the rejected instances are stored for long periods and used for incremental model updates. As an insight, old rejected instances can be easily labeled through publicly available attack repositories without human assistance. Experiments conducted in a novel dataset containing a year of real network traffic with over 2.6 TB of data have shown that current techniques for intrusion detection cannot cope with the network traffic’s evolving behavior, significantly degrading their accuracy over time if no model updates are performed. In contrast, the proposed model can maintain its classification accuracy for long periods without model updates, even improving the false-positive rates by up to 12% while rejecting only 8% of the instances. If periodic model updates are conducted, our proposal can improve the detection accuracy by up to 6% while rejecting only 2% of network events. In addition, the proposed model can perform model updates without human assistance, waiting up to 3 months for the proper event label to be provided without impact on accuracy, while demanding only 3.2% of the computational time and 2% of new instances to be labeled as time passes, making model updates in NIDS a feasible task.

Deep Learning-Based Conformal Prediction of Toxicity.

Predictive modeling for toxicity can help reduce risks in a range of applications and potentially serve as the basis for regulatory decisions. However, the utility of these predictions can be limited if the associated uncertainty is not adequately quantified. With recent studies showing great promise for deep learning-based models also for toxicity predictions, we investigate the combination of deep learning-based predictors with the conformal prediction framework to generate highly predictive models with well-defined uncertainties. We use a range of deep feedforward neural networks and graph neural networks in a conformal prediction setting and evaluate their performance on data from the Tox21 challenge. We also compare the results from the conformal predictors to those of the underlying machine learning models. The results indicate that highly predictive models can be obtained that result in very efficient conformal predictors even at high confidence levels. Taken together, our results highlight the utility of conformal predictors as a convenient way to deliver toxicity predictions with confidence, adding both statistical guarantees on the model performance as well as better predictions of the minority class compared to the underlying models.

SDDM: an interpretable statistical concept drift detection method for data streams

Machine learning models assume that data is drawn from a stationary distribution. However, in practice, challenges are imposed on models that need to make sense of fast-evolving data streams, where the content of data is changing and evolving over time. This change between the distributions of training data seen so-far and the distribution of newly coming data is called concept drift. It is of utmost importance to detect concept drifts to maintain the accuracy and reliability of online classifiers. Reactive drift detectors monitor the performance of the underlying machine learning model. That is, to detect a drift, feedback on the classifier output has to be given to the drift detector, known as prequential evaluation. In many real-life scenarios, immediate feedback on classifier output is not possible. Thus, drift detection is delayed and gets out of context. Moreover, the drift detector output is in the form of a binary answer if there is a drift or not. However, it is equally important to explain the source of drift. In this paper, we present the Statistical Drift Detection Method (SDDM) which can detect drifts by monitoring the change of data distribution without the need for feedback on classifier output. Moreover, the detection is quantified and the source of drift is identified. We empirically evaluate our method against the state-of-the-art on both synthetic and real life data sets. SDDM outperforms other related approaches by producing a smaller number of false positives and false negatives.

Predicting With Confidence: Using Conformal Prediction in Drug Discovery

One of the challenges with predictive modeling is how to quantify the reliability of the models' predictions on new objects. In this work we give an introduction to conformal prediction, a framework that sits on top of traditional machine learning algorithms and which outputs valid confidence estimates to predictions from QSAR models in the form of prediction intervals that are specific to each predicted object. For regression, a prediction interval consists of an upper and a lower bound. For classification, a prediction interval is a set that contains none, one, or many of the potential classes. The size of the prediction interval is affected by a user-specified confidence/significance level, and by the nonconformity of the predicted object; i.e., the strangeness as defined by a nonconformity function. Conformal prediction provides a rigorous and mathematically proven framework for in silico modeling with guarantees on error rates as well as a consistent handling of the models’ applicability domain intrinsically linked to the underlying machine learning model. Apart from introducing the concepts and types of conformal prediction, we also provide an example application for modeling ABC transporters using conformal prediction, as well as a discussion on general implications for drug discovery.

Enterprise AI Canvas Integrating Artificial Intelligence into Business

ABSTRACT Artificial Intelligence (AI) and Machine Learning have enormous potential to transform businesses and disrupt entire industry sectors. However, companies wishing to integrate algorithmic decisions into their organization face multiple challenges: They have to identify use-cases in which artificial intelligence can create value, as well as decisions that can be supported or executed automatically. Furthermore, the organization will need to be transformed to be able to integrate AI-based systems into their human work-force. In addition, the more technical aspects of the underlying machine learning model have to be discussed in terms of how they impact the various units of a business: Where do the relevant data come from, which constraints have to be considered, how is the quality of the data and the prediction evaluated? The Enterprise AI canvas is designed to bring data scientist and business expert together to discuss and define all relevant aspects which need to be clarified in order to integrate AI-based systems into a digital enterprise. It consists of two parts, where part one focuses on the business view and organizational aspects, whereas part two focuses on the underlying machine learning model and the data it uses.

Feedback-aided data acquisition improves myoelectric control of a prosthetic hand

Objective. Pattern-recognition-based myocontrol can be unreliable, which may limit its use in the clinical practice and everyday activities. One cause for this is the poor generalization of the underlying machine learning models to untrained conditions. Acquiring the training data and building the model more interactively can reduce this problem. For example, the user could be encouraged to target the model’s instabilities during the data acquisition supported by automatic feedback guidance. Interactivity is an emerging trend in myocontrol of upper-limb electric prostheses: the user should be actively involved throughout the training and usage of the device. Approach. In this study, 18 non-disabled participants tested two novel feedback-aided acquisition protocols against a standard one that did not provide any guidance. All the protocols acquired data dynamically in multiple arm positions to counteract the limb position effect. During feedback-aided acquisition, an acoustic signal urged the participant to hover with the arm in specific regions of her peri-personal space, de facto acquiring more data where needed. The three protocols were compared on everyday manipulation tasks performed with a prosthetic hand. Main results. Our results showed that feedback-aided data acquisition outperformed the acquisition routine without guidance, both objectively and subjectively. Significance. This indicates that the interaction with the user during the data acquisition is fundamental to improve myocontrol.

Similarity Based Feature Transformation for Network Anomaly Detection

The fundamental objective behind any network intrusion detection system is to automate the detection process whenever intrusions occur in the network. The problem of the network anomaly detection is to determine, if the network incoming traffic is legitimate (or) anomalous. Automated detection systems designed to identify incoming anomalous traffic patterns usually apply widely used machine learning techniques. However, irrespective of any system model which is developed to identify anomalous traffic, all these models requires comparing anomalous and normal traffic patterns. Such comparisons implicitly depend on the ability of the underlying machine learning model to gauge the similarity between a known legitimate observation and the target. The efficiency of any network anomalous detection system depends on the use of distance (or) similarity measures and how they are actually applied. A novel distance function which can be applied to determine the similarity between two conditional feature pattern vectors is an important contribution of present research. Feature dimensionality is another important issue for any machine learning algorithm. In the present work, feature reduction is achieved using the proposed feature transformation technique. However, our approach for feature transformation uses the proposed gaussian distance function to achieve dimensionality reduction to represent the original input dataset in the new transformation space. We have also proposed new computation expressions for determining equivalent deviation and threshold in gaussian space. Experiments are performed on KDD and NSL-KDD datasets by considering widely applied classifier algorithms in various state-of-art research contributions. For performance validation of machine learning models, k-fold cross validation is applied by setting k to 10 through considering evaluation parameters such as accuracy, precision and recall. Experiment results have proved that our approach for anomaly detection that applies the proposed feature transformation technique proved comparatively better to detection methods CANN, GARUDA, and UTTAMA addressed in the recent research literature.

An Easy-to-Use and Fast Assessment of Patient-Specific DBS-Induced Changes in Hand Motor Control in Parkinson's Disease.

For Parkinson's disease (PD), efficient and fast monitoring of fine motor function is fundamental for capturing transient phenomena induced by deep brain stimulation (DBS), thus, enabling a fast and accurate tuning of stimulation parameters. Tuning of DBS parameters is important for obtaining a patient-specific optimal clinical effect and to regularly compensate for disease progress. We propose a fine motor function assessment framework for capturing transient DBS-induced changes. The main goals are to obtain a fast, repeatable, objective, robust, and DBS-sensitive motor-score, in addition to a high-dimensional characterization of motor components by means of an interpretable data-driven model. To achieve this, we combine a hand motor-task, termed the copy-draw test, with a linear model for analyzing features extracted from the proposed task. The approach was tested with four patients totaling eight sessions analyzed. Our approach delivers a motor-score that is sensitive to DBS-induced changes in motor function. It can be applied repeatedly within seconds. The interpretability of the underlying machine learning model provides a direct overview of the feature relevance. This analysis allows to detect and characterize single movement components that are sensitive to DBS. The proposed assessment framework is an useful tool to push forward the data-driven identification of PD-relevant neural markers. Consequent to this end, the source code of the paradigm is made publicly available.

NERBio: using selected word conjunctions, term normalization, and global patterns to improve biomedical named entity recognition

BackgroundBiomedical named entity recognition (Bio-NER) is a challenging problem because, in general, biomedical named entities of the same category (e.g., proteins and genes) do not follow one standard nomenclature. They have many irregularities and sometimes appear in ambiguous contexts. In recent years, machine-learning (ML) approaches have become increasingly common and now represent the cutting edge of Bio-NER technology. This paper addresses three problems faced by ML-based Bio-NER systems. First, most ML approaches usually employ singleton features that comprise one linguistic property (e.g., the current word is capitalized) and at least one class tag (e.g., B-protein, the beginning of a protein name). However, such features may be insufficient in cases where multiple properties must be considered. Adding conjunction features that contain multiple properties can be beneficial, but it would be infeasible to include all conjunction features in an NER model since memory resources are limited and some features are ineffective. To resolve the problem, we use a sequential forward search algorithm to select an effective set of features. Second, variations in the numerical parts of biomedical terms (e.g., "2" in the biomedical term IL2) cause data sparseness and generate many redundant features. In this case, we apply numerical normalization, which solves the problem by replacing all numerals in a term with one representative numeral to help classify named entities. Third, the assignment of NE tags does not depend solely on the target word's closest neighbors, but may depend on words outside the context window (e.g., a context window of five consists of the current word plus two preceding and two subsequent words). We use global patterns generated by the Smith-Waterman local alignment algorithm to identify such structures and modify the results of our ML-based tagger. This is called pattern-based post-processing.ResultsTo develop our ML-based Bio-NER system, we employ conditional random fields, which have performed effectively in several well-known tasks, as our underlying ML model. Adding selected conjunction features, applying numerical normalization, and employing pattern-based post-processing improve the F-scores by 1.67%, 1.04%, and 0.57%, respectively. The combined increase of 3.28% yields a total score of 72.98%, which is better than the baseline system that only uses singleton features.ConclusionWe demonstrate the benefits of using the sequential forward search algorithm to select effective conjunction feature groups. In addition, we show that numerical normalization can effectively reduce the number of redundant and unseen features. Furthermore, the Smith-Waterman local alignment algorithm can help ML-based Bio-NER deal with difficult cases that need longer context windows.