Machine Learning Classification Models Research Articles

Breaking the silence: leveraging social interaction data to identify high-risk suicide users online using network analysis and machine learning

Suicidal thought and behavior (STB) is highly stigmatized and taboo. Prone to censorship, yet pervasive online, STB risk detection may be improved through development of uniquely insightful digital markers. Focusing on Sanctioned Suicide, an online pro-choice suicide forum, this work derived 17 egocentric network features to capture dynamics of social interaction and engagement within this uniquely uncensored community. Using network data generated from over 3.2 million unique interactions of N = 192 individuals, n = 48 of which were determined to be highest risk users (HRUs), a machine learning classification model was trained, validated, and tested to predict HRU status. Model prediction dynamics were analyzed using introspection techniques to uncover patterns in feature influence and highlight social phenomena. The model achieved a test AUC = 0.73 ([0.61, 0.85], 95% CI), suggesting that network-based socio-behavioral patterns of online interaction can signal for heightened suicide risk. Transitivity, density, and in-degree centrality were among the most important features driving this performance. Moreover, predicted HRUs tended to be targets of social exchanges with lesser frequency and possessed egocentric networks with “small world” network properties. Through the implementation of an underutilized method on an unlikely data source, findings support future incorporation of network-based social interaction features in descriptive, predictive, and preventative STB research.

A comparative analysis of classical and machine learning methods for forecasting TB/HIV co-infection

TB/HIV coinfection poses a complex public health challenge. Accurate forecasting of future trends is essential for efficient resource allocation and intervention strategy development. This study compares classical statistical and machine learning models to predict TB/HIV coinfection cases stratified by gender and the general populations. We analyzed time series data using exponential smoothing and ARIMA to establish the baseline trend and seasonality. Subsequently, machine learning models (SVR, XGBoost, LSTM, CNN, GRU, CNN-GRU, and CNN-LSTM) were employed to capture the complex dynamics and inherent non-linearities of TB/HIV coinfection data. Performance metrics (MSE, MAE, sMAPE) and the Diebold-Mariano test were used to evaluate the model performance. Results revealed that Deep Learning models, particularly Bidirectional LSTM and CNN-LSTM, significantly outperformed classical methods. This demonstrates the effectiveness of Deep Learning for modeling TB/HIV coinfection time series and generating more accurate forecasts.

Product differentiation of Sarawak premium rice using MOS gas sensors: comparison between transient and frequency response

In Malaysia, rice, ranked as the third most crucial crop, faces challenges due to domestic consumption outpacing production, resulting in increased instances of rice adulteration. This underscores the imperative of maintaining integrity and quality standards across the entire supply chain. This study uses an electronic nose, comprising four metal oxide semiconductor (MOS) gas sensors, and employing temperature modulation, Principal Component Analysis (PCA) and supervised machine learning (classification models) to distinguish rice varieties such as Bario, Bajong, Borneo Fragrant, Biris, and Jasmine. The study evaluated 30 classifiers based on their classification and validation accuracy. Sensor data was first extracted from the transient response of sensors output voltage, yielding a 12-dimension dataset with response times of 30 s, 50 s, and 95 s. Classification models trained from this dataset achieved classification (training) accuracy of up to 100% and validation accuracy of up to 96%, where the best performing models are subspace discriminant and kernel naïve bayes classifiers. An attempt was also made to analyze the sensor data frequency response for rice classification. Comparison between the prediction results in the transient and frequency domains showed that transient response is better suited for the classification of rice.

FEW questions, many answers: using machine learning to assess how students connect food–energy–water (FEW) concepts

There is growing support and interest in postsecondary interdisciplinary environmental education, which integrates concepts and disciplines in addition to providing varied perspectives. There is a need to assess student learning in these programs as well as rigorous evaluation of educational practices, especially of complex synthesis concepts. This work tests a text classification machine learning model as a tool to assess student systems thinking capabilities using two questions anchored by the Food-Energy-Water (FEW) Nexus phenomena by answering two questions (1) Can machine learning models be used to identify instructor-determined important concepts in student responses? (2) What do college students know about the interconnections between food, energy, and water, and how have students assimilated systems thinking into their constructed responses about FEW? Reported here is a broad range of model performances across 26 text classification models associated with two different assessment items, with model accuracy ranging from 0.755 to 0.992. Expert-like responses were infrequent in our dataset compared to responses providing simpler, incomplete explanations of the systems presented in the question. For those students moving from describing individual effects to multiple effects, their reasoning about the mechanism behind the system indicates advanced systems thinking ability. Specifically, students exhibit higher expertise in explaining changing water usage than discussing trade-offs for such changing usage. This research represents one of the first attempts to assess the links between foundational, discipline-specific concepts and systems thinking ability. These text classification approaches to scoring student FEW Nexus Constructed Responses (CR) indicate how these approaches can be used, in addition to several future research priorities for interdisciplinary, practice-based education research. Development of further complex question items using machine learning would allow evaluation of the relationship between foundational concept understanding and integration of those concepts as well as a more nuanced understanding of student comprehension of complex interdisciplinary concepts.

Predicting the translation efficiency of messenger RNA in mammalian cells.

The degree to which translational control is specified by mRNA sequence is poorly understood in mammalian cells. Here, we constructed and leveraged a compendium of 3,819 ribosomal profiling datasets, distilling them into a transcriptome-wide atlas of translation efficiency (TE) measurements encompassing >140 human and mouse cell types. We subsequently developed RiboNN, a multitask deep convolutional neural network, and classic machine learning models to predict TEs in hundreds of cell types from sequence-encoded mRNA features, achieving state-of-the-art performance (r=0.79 in human and r=0.78 in mouse for mean TE across cell types). While the majority of earlier models solely considered 5' UTR sequence, RiboNN integrates contributions from the full-length mRNA sequence, learning that the 5' UTR, CDS, and 3' UTR respectively possess ~67%, 31%, and 2% per-nucleotide information density in the specification of mammalian TEs. Interpretation of RiboNN revealed that the spatial positioning of low-level di- and tri-nucleotide features (i.e., including codons) largely explain model performance, capturing mechanistic principles such as how ribosomal processivity and tRNA abundance control translational output. RiboNN is predictive of the translational behavior of base-modified therapeutic RNA, and can explain evolutionary selection pressures in human 5' UTRs. Finally, it detects a common language governing mRNA regulatory control and highlights the interconnectedness of mRNA translation, stability, and localization in mammalian organisms.

MLino bench: A comprehensive benchmarking tool for evaluating ML models on edge devices

In today’s rapidly evolving technological landscape, Machine Learning (ML) has become an integral part of our daily lives, ranging from recommendation systems to advanced medical diagnostics and autonomous vehicles. As ML continues to advance, its applications extend beyond conventional boundaries. With the continuous refinement of the models and frameworks, the possibilities for leveraging these technologies into devices that traditionally lacked any form of computational autonomy are ever-expanding. This shift towards embedding ML capabilities directly into edge devices brings new challenges due to the stringent limitations these devices have in terms of memory, power consumption, and cost. The ML models implemented on such devices must find an equilibrium between memory footprint and performance, attaining a classification time that fulfills real-time demands and maintains a similar level of accuracy as the desktop version. Without automated assistance in managing these considerations, the complexity of evaluating multiple models can lead to suboptimal decisions.In this paper, we introduce MLino Bench, an open-source benchmarking tool tailored for assessing lightweight ML models on edge devices with limited resources and capabilities. The tool accommodates various models, frameworks, and platforms, presenting a meticulous design that enables a comprehensive evaluation directly on the target device. It encompasses crucial metrics such as on-target accuracy, classification time, and model size, providing a versatile framework that assists practitioners in decision-making when deploying models to such devices.The tool employs a fully streamlined benchmark flow involving training the ML model in a high-level interpreted language, porting, compiling, flashing, and finally benchmarking on the actual target. Our experimental evaluation of the tool highlights its flexibility in assessing multiple ML models across different model hyperparameters, frameworks, datasets, and embedded platforms. Furthermore, a distinctive advantage compared to state-of-the-art ML benchmarking tools is the inclusion of classical ML models, including Random Forests, Decision Trees, Support Vector Machines, Naive Bayes, and more. This sets our tool apart from others that predominantly emphasize only neural network models. Due to this inclusive approach, our tool facilitates the evaluation of ML models across a broad spectrum of devices, ranging from resource-constrained edge devices to those with medium and advanced computational capabilities.

China’s bioenergy potential will be stable and decoupling from economy and population at national level: A study based on machine regression prediction model

Policymakers and industry participants strongly demand accurate and convenient methods for predicting bioenergy potential. To address the limitations of existing biophysical models, which require complex inventories and technical support, this study utilizes readily available economic and population data from 1978 to 2021 as independent variables, with bioenergy as the dependent variable. After evaluating ten classical machine learning models and six newly published optimization algorithms, the hybrid of Back Propagation and Marine Predators Algorithm was selected as an effective bioenergy regression prediction model. The prediction results of 2022–2050 indicate that China’s bioenergy potential will reach 13 ± 3.8 Bt equivalent of standard coal and remain stable, decoupled from economic and population trends at the national level over the next three decades. Additionally, developing measures to promote the conversion of biomass into hydrogen, methanol, and sustainable aviation fuels will be a significant opportunity for the sustainable development of the bioenergy industry.

Evaluation of Machine Learning Classification Models for False-Positive Reduction in Prostate Cancer Detection Using MRI Data.

In this work, several machine learning (ML) algorithms, both classical ML and modern deep learning, were investigated for their ability to improve the performance of a pipeline for the segmentation and classification of prostate lesions using MRI data. The algorithms were used to perform a binary classification of benign and malignant tissue visible in MRI sequences. The model choices include support vector machines (SVMs), random decision forests (RDFs), and multi-layer perceptrons (MLPs), along with radiomic features that are reduced by applying PCA or mRMR feature selection. Modern CNN-based architectures, such as ConvNeXt, ConvNet, and ResNet, were also evaluated in various setups, including transfer learning. To optimize the performance, different approaches were compared and applied to whole images, as well as gland, peripheral zone (PZ), and lesion segmentations. The contribution of this study is an investigation of several ML approaches regarding their performance in prostate cancer (PCa) diagnosis algorithms. This work delivers insights into the applicability of different approaches for this context based on an exhaustive examination. The outcome is a recommendation or preference for which machine learning model or family of models is best suited to optimize an existing pipeline when the model is applied as an upstream filter.

Predicting the Hallucinogenic Potential of Molecules Using Artificial Intelligence.

The development of new drugs addressing serious mental health and other disorders should avoid the psychedelic experience. Analogs of psychedelic drugs can have clinical utility and are termed "psychoplastogens". These represent promising candidates for treating opioid use disorder to reduce drug dependence, with rarely reported serious adverse effects. This drug abuse cessation is linked to the induction of neuritogenesis and increased neuroplasticity, a hallmark of psychedelic molecules, such as lysergic acid diethylamine. Some, but not all psychoplastogens may act through the G-protein coupled receptor (GPCR) 5HT2A whereas others may display very different polypharmacology making prediction of hallucinogenic potential challenging. In the process of developing tools to help design new psychoplastogens, we have used artificial intelligence in the form of machine learning classification models for predicting psychedelic effects using a published in vitro data set from PsychLight (support vector classification (SVC), area under the curve (AUC) 0.74) and in vivo human data derived from books from Shulgin and Shulgin (SVC, AUC, 0.72) with nested five-fold cross validation. We have also explored conformal predictors with ECFP6 and electrostatic descriptors in an effort to optimize them. These models have been used to predict known 5HT2A agonists to assess their potential to act as psychedelics and induce hallucinations for PsychLight (SVC, AUC 0.97) and Shulgin and Shulgin (random forest, AUC 0.71). We have tested these models with head twitch data from the mouse. This predictive capability is desirable to reliably design new psychoplastogens that lack in vivo hallucinogenic potential and help assess existing and future molecules for this potential. These efforts also provide useful insights into understanding the psychedelic structure activity relationship.

Snarcase - Regain Control over Your Predictions with Low-Latency Machine Unlearning

The "right-to-be-forgotten" requires the removal of personal data from trained machine learning (ML) models with machine unlearning. Conducting such unlearning with low latency is crucial for responsible data management. Low-latency unlearning is challenging, but possible for certain classes of ML models when treating them as "materialised views" over training data, with carefully chosen operations and data structures for computing updates. We present Snapcase, a recommender system that can unlearn user interactions with sub-second latency on a large grocery shopping dataset with 33 million purchases and 200 thousand users. Its implementation is based on incremental view maintenance with Differential Dataflow and a custom algorithm and data structure for maintaining a top- k aggregation over the result of a sparse matrix-matrix multiplication. We demonstrate how interactive low-latency unlearning empowers users in critical scenarios to get rid of sensitive items in their recommendations and to drastically reduce their data's negative influence on other users' predictions.

Early Warning Models Using Machine Learning to Predict Sepsis-Associated Chronic Critical Illness: A Study Based on the Medical Information Mart for Intensive Care Database.

Background Patients with chronic critical illness (CCI) experience poor prognoses and incur high medical costs. However, there is currently limited clinical awareness of sepsis-associated CCI, resulting in insufficient vigilance.Therefore, it is necessary to build a machine learning model that can predict whether sepsis patients will developCCI. Methods Clinical data on19,077 sepsis patients from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database were analyzed. Predictive factors were identified using the Student's t-test, Mann-Whitney Utest, or χ 2test. Six machine learning classification models, namely, the logistic regression, support vector machine, decision tree, random forest, extreme gradient enhancement, and artificial neural network, were established. The optimal model was selected on the basis of its performance. Calibration curves wereused to evaluate the accuracy of model classification, while the external validation dataset was used to evaluate the performance of the model. Results Thirty-seven characteristics,such as elevated alanine aminotransferase, rapid heart rate, and high Logistic Organ Dysfunction System scores, were identified as risk factors for developingCCI. The area under the receiver operating characteristic curve (AUROC) values forall models were above 0.73 onthe internal test set. Among them, the extreme gradient enhancement model exhibited superior performance (F1score=0.91, AUROC=0.91, Brierscore=0.052). It alsoexhibited stable prediction performance onthe external validation set (AUROC =0.72). Conclusion A machine learning model was established to predict whether sepsis patients will develop CCI. It can provide useful predictive information for clinical decision-making.

Predictive Model Comparison for Predicting Condom Use: Comparison of Conventional Logistic Regression and Other Machine Learning

Condom use at first sex remains an important issue as it shapes future sexual behavior. This study aimed to deploy and predict condom use using five different machine learning classification models. Dataset used for this study was from Indonesian Demographic and Health Survey (IDHS) 2017 with a population of interest was male adolescents. We evaluated five different models, namely logistic regression, naïve bayes, K-Nearest Neighbors, support vector machines, and decision tree. Performances of each model were assessed using metrics such as accuracy, specificity, sensitivity, ROC Curve, and AUC Score. Study found that different models exhibit different accuracy, specificity, sensitivity, ROC Curve, and AUC Score. The decision tree and naïve bayes models remained the models with the highest specificity and sensitivity, however the KNN model expressed the highest AUC score. Result from the conventional logistic regression also explained that condom use was associated with education level, age at first sex, and attitude towards condom use. The government is advised to create equal education opportunities for every adolescent and shape better knowledge and condom attitudes. Future studies are advised to enhance the performance of machine learning models using hyperparameter tuning and other methods.

On the transferability of residence time distributions in two 10-km long river sections with similar hydromorphic units

Quantifying hydrologic exchange fluxes (HEFs) at the stream-groundwater interface and their residence time distributions (RTDs) in the subsurface are important for managing the water quality and ecosystem health in dynamic river corridors. However, field measurements and directly simulating high-spatial resolution HEFs and RTDs can be time-consuming, particularly at watershed-scale. Recent research has proposed the potential application of using hydromorphic units (HUs), which cluster the bathymetry and surface water hydrodynamics attributes, to aid in RTD estimation. However, more pioneering studies have demonstrated that underground structure and large-scale river channel geomorphology are the primary factors controlling the HEFs and RTDs. To address this contradiction, this work evaluates the HEFs and resulting HU-RTD relationships for two 10-km long river sections along the Columbia River, leveraging a one-way coupled three-dimensional transient surface–subsurface water transport modeling framework. Applying such a framework at the two river sections with similar HUs allows for quantitative comparisons of HEFs and RTDs using both statistical tests and machine learning classification models. The comparison reveals that the similarity and transferability of the HU-RTD relationship is very low for the two investigated river sections. This suggests that, creating a general algorithm to estimate RTDs in large-scale river reaches based solely on surface water hydrodynamics and short-distance bathymetry topography data may be nearly impossible.

Interpretable AI and Machine Learning Classification for Identifying High-Efficiency Donor-Acceptor Pairs in Organic Solar Cells.

To enhance the efficiency of organic solar cells, accurately predicting the efficiency of new pairs of donor and acceptor materials is crucial. Presently, most machine learning studies rely on regression models, which often struggle to establish clear rules for distinguishing between high- and low-performing donor-acceptor pairs. This study proposes a novel approach by integrating interpretable AI, specifically using Shapely values, with four supervised machine learning classification models, namely, support vector machines, decision trees, random forest, and gradient boosting. These models aim to identify high-efficiency donor-acceptor pairs based solely on chemical structures and to extract important features that establish general design principles for distinguishing between high- and low-efficiency pairs. For validation purposes, an unsupervised machine learning algorithm utilizing loading vectors obtained from the principal component analysis is employed to identify crucial features associated with high-efficiency donor-acceptor pairs. Interestingly, the features identified by the supervised machine learning approach were found to be a subset of those identified by the unsupervised method. Noteworthy features include the van der Waals surface area, partial equalization of orbital electronegativity, Moreau-Broto autocorrelation, and molecular substructures. Leveraging these features, a backward-working model can be developed, facilitating exploration across a wide array of materials used in organic solar cells. This innovative approach will help navigate the vast chemical compound space of donor and acceptor materials essential in creating high-efficiency organic solar cells.

Machine Learning Predicts Peripherally Inserted Central Catheters-Related Deep Vein Thrombosis Using Patient Features and Catheterization Technology Features.

This study aims to use patient feature and catheterization technology feature variables to train the corresponding machine learning (ML) models to predict peripherally inserted central catheters-deep vein thrombosis (PICCs-DVT) and analyze the importance of the two types of features to PICCs-DVT from the aspect of "input-output" correlation. To comprehensively and systematically summarize the variables used to describe patient features and catheterization technical features, this study combined 18 literature involving the two types of features in predicting PICCs-DVT. A total of 21 variables used to describe the two types of features were summarized, and feature values were extracted from the data of 1,065 PICCs patients from January 1, 2021 to August 31, 2022, to construct a data sample set. Then, 70% of the sample set is used for model training and hyperparameter optimization, and 30% of the sample set is used for PICCs-DVT prediction and feature importance analysis of three common ML classification models (i.e. support vector classifier [SVC], random forest [RF], and artificial neural network [ANN]). In terms of prediction performance, this study selected four metrics to evaluate the prediction performance of the model: precision (P), recall (R), accuracy (ACC), and area under the curve (AUC). In terms of feature importance analysis, this study chooses a single feature analysis method based on the "input-output" sensitivity principle-Permutation Importance. For the mean model performance, the three ML models on the test set are P = 0.92, R = 0.95, ACC = 0.88, and AUC = 0.81. Specifically, the RF model is P = 0.95, R = 0.96, ACC = 0.92, AUC = 0.86; the ANN model is P = 0.92, R = 0.95, ACC = 0.88, AUC = 0.81; the SVC model is P = 0.88, R = 0.94, ACC = 0.85, AUC = 0.77. For feature importance analysis, Catheter-to-vein rate (RF: 91.55%, ANN: 82.25%, SVC: 87.71%), Zubrod-ECOG-WHO score (RF: 66.35%, ANN: 82.25%, SVC: 44.35%), and insertion attempt (RF: 44.35%, ANN: 37.65%, SVC: 65.80%) all occupy the top three in the ML models prediction task of PICCs-DVT, showing relatively consistent ranking results. The ML models show good performance in predicting PICCs-DVT and reveal a relatively consistent ranking of feature importance from the data. The important features revealed might help clinical medical staff to better understand and analyze the formation mechanism of PICCs-DVT from a data-driven perspective.

QKSAN: A Quantum Kernel Self-Attention Network.

The Self-Attention Mechanism (SAM) excels at distilling important information from the interior of data to improve the computational efficiency of models. Nevertheless, many Quantum Machine Learning (QML) models lack the ability to distinguish the intrinsic connections of information like SAM, which limits their effectiveness on massive high-dimensional quantum data. To tackle the above issue, a Quantum Kernel Self-Attention Mechanism (QKSAM) is introduced to combine the data representation merit of Quantum Kernel Methods (QKM) with the efficient information extraction capability of SAM. Further, a Quantum Kernel Self-Attention Network (QKSAN) framework is proposed based on QKSAM, which ingeniously incorporates the Deferred Measurement Principle (DMP) and conditional measurement techniques to release half of quantum resources by mid-circuit measurement, thereby bolstering both feasibility and adaptability. Simultaneously, the Quantum Kernel Self-Attention Score (QKSAS) with an exponentially large characterization space is spawned to accommodate more information and determine the measurement conditions. Eventually, four QKSAN sub-models are deployed on PennyLane and IBM Qiskit platforms to perform binary classification on MNIST and Fashion MNIST, where the QKSAS tests and correlation assessments between noise immunity and learning ability are executed on the best-performing sub-model. The paramount experimental finding is that the QKSAN subclasses possess the potential learning advantage of acquiring impressive accuracies exceeding 98.05% with far fewer parameters than classical machine learning models. Predictably, QKSAN lays the foundation for future quantum computers to perform machine learning on massive amounts of data while driving advances in areas such as quantum computer vision.

VQCNN: variational quantum convolutional neural networks based on quantum filters and fully connected layers

Abstract Classical machine learning is more susceptible to adversarial examples due to its linear and non-robust nature, which results in a severe degradation of the recognition accuracy of classical machine learning models. Quantum techniques are shown to have a higher robustness advantage and are more resistant to attacks from adversarial examples than classical machine learning. Inspired by the robustness advantage of quantum computing and the feature extraction advantage of convolutional neural networks, this paper proposes a novel variational quantum convolutional neural network model (VQCNN), whose quantum fully connected layer consists of a combination of a quantum filter and a variational quantum neural network to increase the model’s adversarial robustness. The network intrusion detection model based on VQCNN is verified on KDD CUP99 and UNSW-NB datasets. The results show that under the attack of Fast Gradient Sign Method, the decline values of accuracy, precision, and recall rate of the intrusion detection model based on VQCNN are less than those of the other four models, and it has higher adversarial robustness.

A self-attention TCN-based model for suicidal ideation detection from social media posts

Early suicidal ideation detection has long been regarded as an important task that can benefit both society and individuals. In this regard, it has been shown that, very frequently, the first symptoms of this problem can be identified by analyzing the contents shared on social media. Machine learning classification models have proven promising in capturing behavioral and textual features from posts shared on social media. This study proposes a novel machine-learning model to detect the risk of suicide from social media posts, employing both natural language processing and state-of-the-art deep learning techniques. We propose an ensemble LSTM-TCN model that benefits from a self-attention mechanism to detect suicidal ideation among users of two well-known social networks, Twitter (X) and Reddit. Furthermore, we present a comprehensive analysis of the data, examining the suicidal posts both statistically and semantically, which can provide rich knowledge about suicidal ideation. Our proposed model (AL-BTCN) outperforms the compared state-of-the-art models, resulting in over 94% accuracy, recall, and F1-score. Researchers, mental health specialists, and social media service providers can all benefit from the findings of this study.

VEXP: A virtual enhanced cross screen panel for off-target pharmacology alerts

We describe the development of the GSK vEXP (virtual enhanced cross screen panel) for off-target pharmacology alerts. The derivation of a panel of machine learning classification models or QSAR models (Quantitative Structure-Activity Relationship) for off-target safety assessment allows early alerting to risk factors in candidate drugs. The models are matched to an internal in-vitro biochemical screening panel described previously with some updates reported here. The extreme imbalance of some internal GSK datasets and most of the related external ChEMBL datasets is shown when considering potency thresholds relevant to in-vitro screening. The small size and bias to the active class make many ChEMBL datasets un-modellable using such thresholds. Although larger, many GSK datasets remain too imbalanced to give a performant model. The value of merging internal and external data to help rebalance datasets and improve the domain of applicability is demonstrated with improvements in model performance frequently seen on merged data. Efforts to collate public datasets with a far better balance of the missing in-actives would likely do more to improve opensource models than simply increasing dataset size. We investigate the use of moving the probability threshold and applying imbalanced learners to help overcome the imbalance problem. Both methods can produce models with improved performance when applied to imbalanced datasets. Datasets with class imbalance 95:5 % or with <100 compounds were un-modellable. Where datasets had a class imbalance of 90:10 % the imbalanced learn methods were often more performant than standard tree-based classifiers. No one classification algorithm consistently out-performed all others and our approach emphasises a standardised, automated build and evaluate approach across all classifiers to identify the best model. The application of vEXP includes ranking of hit compounds for fast prioritisation, flagging of hit series that contain systematic scaffold or functional group related risks and the confirmation that late-stage optimisation is not introducing new off-target activities in established chemical series.

Massively Parallel Evasion Attacks and the Pitfalls of Adversarial Retraining

Even with widespread adoption of automated anomaly detection in safety-critical areas, both classical and advanced machine learning models are susceptible to first-order evasion attacks that fool models at run-time (e.g. an automated firewall or an anti-virus application). Kernelized support vector machines (KSVMs) are an especially useful model because they combine a complex geometry with low run-time requirements (e.g. when compared to neural networks), acting as a run-time lower bound when compared to contemporary models (e.g. deep neural networks), to provide a cost-efficient way to measure model and attack run-time costs. To properly measure and combat adversaries, we propose a massively parallel projected gradient descent (PGD) evasion attack framework. Through theoretical examinations and experiments carried out using linearly-separable Gaussian normal data, we present (i) a massively parallel naive attack, we show that adversarial retraining is unlikely to be an effective means to combat an attacker even on linearly separable datasets, (ii) a cost effective way of evaluating models defences and attacks, and an extensible code base for doing so, (iii) an inverse relationship between adversarial robustness and benign accuracy, (iv) the lack of a general relationship between attack time and efficacy, and (v) that adversarial retraining increases compute time exponentially while failing to reliably prevent highly-confident false classifications.