Quantifying (Hyper) Parameter Leakage in Machine Learning

We propose a machine learning model that aims to balance the trade-off between explainability and predictive accuracy by combining black-box and white-box models. Our method uses an additional machine learning model to classify input data according to whether only the black-box model makes a correct prediction. Input data with a low probability of being correctly predicted by only the black-box model are allocated to the white-box model. Experimental results for 16 classification and 16 regression datasets show that our method can maintain overall accuracy and enhance explainability compared to an existing method. Our method is promising for improving the explainability of machine learning models in practical applications.

This paper is based on a presentation given in December 2019 at the Lund’s University conference ‘Artificial Intelligence (AI), Data Protection and Intellectual Property in a European context’. The purpose of this article is to analyse the suitability of the copyright system over Machine Learning (ML) models, the so-called ‘core components’ of ML systems. Computer programs protection has always caused certain difficulties for intellectual property law. Internationally, agreement was reached in the 1970s to protect computer programs as literary works of copyright. ML models have been called ‘learning algorithms’, ‘AI computer programs’, and ‘software 2.0’. Yet there is no unanimity about what they are technically. This is relevant from a copyright perspective, because the regime of protection granted by copyright will be different depending on whether the ML model qualifies as a computer program, as a mathematical method, or as another type of work. Additionally, all proprietary and open source software licensing relies on copyright protection. In most open licenses, the license is not triggered if it is applied to something that is not protected by copyright (or related rights). Thus, it seems pertinent to question whether EU copyright law provides adequate protection for the core components of machine learning systems, the ML models.

The best practice of watershed management is through the understanding of the hydrological processes. As a matter of fact, hydrological processes are highly associated with stochastic, non-linear, and non-stationary phenomena. Hydrological processes simulation and modeling are challenging issues in the domains of hydrology, climate and environment. Hence, the development of machine learning (ML) models for solving those complex hydrological problems took essential place over the past couple decades. It can be observed, hydrological data availability has increased remarkably, and thus computational resources has led to a resurgence in ML models’ development. It has been witnessed huge efforts on the hydrological processes modeling using the facility of ML models and several review researches have been conducted. Literature studies approved the capacity of ML models in the field of hydrology over the classical “traditional models” based on their forecastability, flexibility, precision, generalization, and modeling execution convergence speed. However, although several potential merits were observed in ML model’s development, several limitations are allied such as the interpretability of those black-box models, the practicality of the ML models in watershed management, and difficulty to explain the physical hydrological processes. In this survey, an exhibition for all the published review articles on the development of ML models for hydrological processes and recognize all the research gaps and potential research direction. The ultimate aim of the current survey is to establish a new milestone for the interested hydrology, environment and climate researchers on the applications of ML models.

Accurate prediction of energy consumption in district heating systems plays an important role in supporting effective and clean energy production and distribution in dense urban areas. Predictive models are needed for flexible and cost-effective operation of energy production and usage, e.g., using peak shaving or load shifting to compensate for heat losses in the pipeline. This helps to avoid exceedance of power plant capacity. The purpose of this study is to automate the process of building machine learning (ML) models to solve a short-term power demand prediction problem. The dataset contains a district heating network’s measured hourly power consumption and ambient temperature for 415 days. In this paper, we propose a hybrid evolutionary-based algorithm, named GA-SHADE, for the simultaneous optimization of ML models and feature selection. The GA-SHADE algorithm is a hybrid algorithm consisting of a Genetic Algorithm (GA) and success-history-based parameter adaptation for differential evolution (SHADE). The results of the numerical experiments show that the proposed GA-SHADE algorithm allows the identification of simplified ML models with good prediction performance in terms of the optimized feature subset and model hyperparameters. The main contributions of the study are (1) using the proposed GA-SHADE, ML models with varying numbers of features and performance are obtained. (2) The proposed GA-SHADE algorithm self-adapts during operation and has only one control parameter. There is no fine-tuning required before execution. (3) Due to the evolutionary nature of the algorithm, it is not sensitive to the number of features and hyperparameters to be optimized in ML models. In conclusion, this study confirms that each optimized ML model uses a unique set and number of features. Out of the six ML models considered, SVR and NN are better candidates and have demonstrated the best performance across several metrics. All numerical experiments were compared against the measurements and proven by the standard statistical tests.

<p>Ensemble learning is one of the most powerful extensions for improving upon individual machine learning models. Rather than a single model being used, several models are trained and the predictions combined to make a more informed decision. Such combinations will ideally overcome the shortcomings of any individual member of the ensemble. Most ma- chine learning competition winners feature an ensemble of some sort, and there is also sound theoretical proof to the performance of certain ensem- bling schemes. The benefits of ensembling are clear in both theory and practice. Despite the great performance, ensemble learning is not a trivial task. One of the main difficulties is designing appropriate ensembles. For exam- ple, how large should an ensemble be? What members should be included in an ensemble? How should these members be weighted? Our first contribution addresses these concerns using a strongly-typed population- based search (genetic programming) to construct well-performing ensem- bles, where the entire ensemble (members, hyperparameters, structure) is automatically learnt. The proposed method was found, in general, to be significantly better than all base members and commonly used compari- son methods trialled. With automatically designed ensembles, there is a range of applica- tions, such as competition entries, forecasting and state-of-the-art predic- tions. However, often these applications also require additional prepro- cessing of the input data. Above the ensemble considers only the original training data, however, in many machine learning scenarios a pipeline is required (for example performing feature selection before classification). For the second contribution, a novel automated machine learning method is proposed based on ensemble learning. This method uses a random population-based search of appropriate tree structures, and as such is em- barrassingly parallel, an important consideration for automated machine learning. The proposed method is able to achieve equivalent or improved results over the current state-of-the-art methods and does so in a fraction of the time (six times as fast). Finally, while complex ensembles offer great performance, one large limitation is the interpretability of such ensembles. For example, why does a forest of 500 trees predict a particular class for a given instance? In an effort to explain the behaviour of complex models (such as ensem- bles), several methods have been proposed. However, these approaches tend to suffer at least one of the following limitations: overly complex in the representation, local in their application, limited to particular fea- ture types (i.e. categorical only), or limited to particular algorithms. For our third contribution, a novel model agnostic method for interpreting complex black-box machine learning models is proposed. The method is based on strongly-typed genetic programming and overcomes the afore- mentioned limitations. Multi-objective optimisation is used to generate a Pareto frontier of simple and explainable models which approximate the behaviour of much more complex methods. We found the resulting rep- resentations are far simpler than existing approaches (an important con- sideration for interpretability) while providing equivalent reconstruction performance. Overall, this thesis addresses two of the major limitations of existing ensemble learning, i.e. the complex construction process and the black- box models that are often difficult to interpret. A novel application of ensemble learning in the field of automated machine learning is also pro- posed. All three methods have shown at least equivalent or improved performance than existing methods.</p>

Explainable Boosting Machine for Choosing Radiation Dose-Volume Constraints on Cardio-Pulmonary Substructures Associated with Overall Survival in the NRG Oncology RTOG 0617 Clinical Trial

In recent years, a significant amount of research has focused on analyzing the effectiveness of machine learning (ML) models for malware detection. These approaches have ranged from methods such as decision trees and clustering to more complex approaches like support vector machine (SVM) and deep neural networks. In particular, neural networks have proven to be very effective in detecting complex and advanced malware. This, however, comes with a caveat. Neural networks are notoriously complex. Therefore, the decisions that they make are often just accepted without questioning why the model made that specific decision. The black box characteristic of neural networks has challenged researchers to explore methods to explain black-box models such as SVM and neural networks and their decision-making process. Transparency and explainability give the experts and malware analysts assurance and trustworthiness about the ML models’ decisions. In addition, it helps in generating comprehensive reports that can be used to enhance cyber threat intelligence sharing. As such, this much needed analysis drives our work in this paper to explore the explainability and interpretability of ML models in the field of online malware detection. In this paper, we used the Shapley Additive exPlanations (SHAP) explainability technique to achieve efficient performance in interpreting the outcome of different ML models such as SVM Linear, SVM-RBF (Radial Basis Function), Random Forest (RF), Feed-Forward Neural Net (FFNN), and Convolutional Neural Network (CNN) models trained on an online malware dataset. To explain the output of these models, explainability techniques such as KernalSHAP, TreeSHAP, and DeepSHAP are applied to the obtained results.

Unconventional oil and gas reservoirs are usually classifiedbyextremely low porosity and permeability values. The most economicalway to produce hydrocarbons from such reservoirs is by creating artificiallyinduced channels. To effectively design hydraulic fracturing jobs,accurate values of rock breakdown pressure are needed. Conductinghydraulic fracturing experiments in the laboratory is a very expensiveand time-consuming process. Therefore, in this study, different machinelearning (ML) models were efficiently utilized to predict the breakdownpressure of tight rocks. In the first part of the study, to measurethe breakdown pressures, a comprehensive hydraulic fracturing experimentalstudy was conducted on various rock specimens. A total of 130 experimentswere conducted on different rock types such as shales, sandstone,tight carbonates, and synthetic samples. Rock mechanical propertiessuch as Young’s modulus (E), Poisson’sratio (ν), unconfined compressive strength, and indirect tensilestrength (σt) were measured before conducting hydraulicfracturing tests. ML models were used to correlate the breakdown pressureof the rock as a function of fracturing experimental conditions androck properties. In the ML model, we considered experimental conditions,including the injection rate, overburden pressures, and fracturingfluid viscosity, and rock properties including Young’s modulus(E), Poisson’s ratio (ν), UCS, and indirecttensile strength (σt), porosity, permeability, andbulk density. ML models include artificial neural networks (ANNs),random forests, decision trees, and the K-nearest neighbor. Duringtraining of ML models, the model hyperparameters were optimized bythe grid-search optimization approach. With the optimal setting ofthe ML models, the breakdown pressure of the unconventional formationwas predicted with an accuracy of 95%. The accuracy of all ML techniqueswas quite similar; however, an explicit empirical correlation fromthe ANN technique is proposed. The empirical correlation is the functionof all input features and can be used as a standalone package in anysoftware. The proposed methodology to predict the breakdown pressureof unconventional rocks can minimize the laboratory experimental costof measuring fracture parameters and can be used as a quick assessmenttool to evaluate the development prospect of unconventional tightrocks.

This paper proposes to explain the black-box feature of data-driven machine learning (ML) models used for controlling power electronic converters for the first time. As the name suggests, their “black box” feature prevents a clear understanding of the physical insights behind these ML models. It remains a fundamental aspect, if one plans to take action based on a prediction, or deploy a new ML model. Moreover, leaked and corrupted data during the training process can easily augment unexplainable actions from them. To address these issues, we first interpret the actions of the black box models by calculating a conditional entropy for each input with respect to an output. Using this metric, the averaged relationships between each input-output can be mapped and representative conclusions are firstly drawn on identifying erroneous data. Finally, these abnormal data are then removed from the training database to improve the interpretability & classification abilities of the ML model. We illustrate our findings on the performance of a regression based learning tool used for controlling a grid-connected voltage source inverter (VSI).

Maintaining and rehabilitating pavement in a timely manner is essential for preserving or improving its condition, with roughness being a critical factor. Accurate prediction of road roughness is a vital component of sustainable transportation because it helps transportation planners to develop cost-effective and sustainable pavement maintenance and rehabilitation strategies. Traditional statistical methods can be less effective for this purpose due to their inherent assumptions, rendering them inaccurate. Therefore, this study employed explainable and supervised machine learning algorithms to predict the International Roughness Index (IRI) of asphalt concrete pavement in Sri Lankan arterial roads from 2013 to 2018. Two predictor variables, pavement age and cumulative traffic volume, were used in this study. Five machine learning models, namely Random Forest (RF), Decision Tree (DT), XGBoost (XGB), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN), were utilized and compared with the statistical model. The study findings revealed that the machine learning algorithms’ predictions were superior to those of the regression model, with a coefficient of determination (R2) of more than 0.75, except for SVM. Moreover, RF provided the best prediction among the five machine learning algorithms due to its extrapolation and global optimization capabilities. Further, SHapley Additive exPlanations (SHAP) analysis showed that both explanatory variables had positive impacts on IRI progression, with pavement age having the most significant effect. Providing accurate explanations for the decision-making processes in black box models using SHAP analysis increases the trust of road users and domain experts in the predictions generated by machine learning models. Furthermore, this study demonstrates that the use of explainable AI-based methods was more effective than traditional regression analysis in IRI prediction. Overall, using this approach, road authorities can plan for timely maintenance to avoid costly and extensive rehabilitation. Therefore, sustainable transportation can be promoted by extending pavement life and reducing frequent reconstruction.

One of the important production processes in the textile industry is the spinning process. Starting from cotton fibers, yarns are (usually) created on a rotor-spinning machine. The spinnability of a fiber is dependant on its quality and on the machine settings of the spinning machine. It would be a great benefit to be able to predict the spinnability and resulting strength of the yarn starting from a certain quality and from machine settings. To this end, two totally diiffferent modeling approaches can be considered: the so-called 'white' modeling and the so-called 'black box' modeling. In white modeling the process is described by mathematical equations, which are based upon (theoretical) physical knowledge of the process. Extensive physical information about the process is in this case available through physical, chemical or mechanical equations giving the user a thorough insight into the operation of the process. However, due to the large input (and output) dimensions of the fiber-to-yarn process and their complex interactions, no exact mathematical model of a spinning machine is known to exist nor is it likely that such a model will ever be constructed. A black box model, in contrast to white modeling, simply connects input parameters to the output without giving or containing any substantial physical information about the process itsellf Black box models have been successfully constructed by Pynckels et al. to predict the spinnability (Pynckels, 1995) and the characteristics (Pynckels, 1997) of the yarn using neural networks with a Backpropagation learning rule. Apart from the lack of physical information, these models also have no fault indication or measure of uncertainty about the results.

In this paper a new Intellectual Property (IP) VLSI architecture for the implementation of real time and low complexity fast motion estimation for. multimedia applications is proposed. The motion estimation is a kev issue either in H.263/MPEG video coding ol in image filtering. Specially, the algorithms bused on the predictive spatio-temporal technique achieve high coding quality at reasonable computational power by exploiting the spatio-temporal correlation of the video motion field. The novel architecture, obtained by a design-reuse methodology: is parametric and configurable and hence it allows for the implementation of different predictive algorithms. Ir also features hardware complexity scalability and it is suitable for the design of ASICs optimized for a wide range of multimedia applications. The IF, synthesized for a 0.25 mum CMOS technology: achieves a computational power lip to 740x10(6) absolute differences per second, for a maximum 0.96 mm(2) core size, and permits the processing of typical video images at clock frequencies of few MHz.

Stone columns have been extensively advocated as a traditional approach to increase the undrained bearing capacity and reduce the settlement of footings sitting on cohesive ground. However, due to the complex interaction between the soil and the stone columns, there currently needs to be a commonly acknowledged approach that can be used to precisely predict the undrained bearing capacity of the system. For this reason, the bearing capacity of a sandy bed reinforced with geogrid and sitting above a collection of geogrid-encased stone columns floating in soft clay was studied in this research. Using a white-box machine learning (ML) technique called Multivariate Polynomial Regression (MPR), this work aims to develop a model for predicting the bearing capacity of the referred foundation system. For this purpose, two hundred and forty-five experimental results were collected from the literature. In addition, the model was compared to two other ML models, namely, a black-box model known as Random Forest (RF) and a white-box ML model called Linear Regression (LR). In terms of R2 (coefficient of determination) and RMSE (Root Mean Absolute Error) values, the newly proposed model outperforms the two other referred models and demonstrates robust estimation capabilities. In addition, a parametric analysis was carried out to determine the contribution of each input variable and its relative significance on the output.

Machine learning has achieved remarkable success across diverse domains. Nevertheless, concerns about interpretability in black-box models, especially within Deep Neural Networks (DNNs), have become pronounced in safety-critical fields like healthcare and finance. Classical machine learning (ML) classifiers, known for their higher interpretability, are preferred in these domains. Similar to DNNs, classical ML classifiers can exhibit bugs that could lead to severe consequences in practice. Test input prioritization has emerged as a promising approach to ensure the quality of an ML system, which prioritizes potentially misclassified tests so that such tests can be identified earlier with limited manual labeling costs. However, when applying to classical ML classifiers, existing DNN test prioritization methods are constrained from three perspectives: 1) Coverage-based methods are inefficient and time-consuming; 2) Mutation-based methods cannot be adapted to classical ML models due to mismatched model mutation rules; 3) Confidence-based methods are restricted to a single dimension when applying to binary ML classifiers, solely depending on the model’s prediction probability for one class. To overcome the challenges, we propose MLPrior, a test prioritization approach specifically tailored for classical ML models. MLPrior leverages the characteristics of classical ML classifiers (i.e., interpretable models and carefully engineered attribute features) to prioritize test inputs. The foundational principles are: 1) tests more sensitive to mutations are more likely to be misclassified, and 2) tests closer to the model’s decision boundary are more likely to be misclassified. Building on the first concept, we design mutation rules to generate two types of mutation features (i.e., <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">model mutation features</b> and <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">input mutation features</b> ) for each test. Drawing from the second notion, MLPrior generates <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">attribute features</b> of each test based on its attribute values, which can indirectly reveal the proximity between the test and the decision boundary. For each test, MLPrior combines all three types of features of it into a final vector. Subsequently, MLPrior employs a pre-trained ranking model to predict the misclassification probability of each test based on its final vector and ranks tests accordingly. We conducted an extensive study to evaluate MLPrior based on 185 subjects, encompassing natural datasets, mixed noisy datasets, and fairness datasets. The results demonstrate that MLPrior outperforms all the compared test prioritization approaches, with an average improvement of 14.74%∼66.93% on natural datasets, 18.55%∼67.73% on mixed noisy datasets, and 15.34%∼62.72% on fairness datasets.

Theory-Guided Machine Learning Finds Geometric Structure-Property Relationships for Chemisorption on Subsurface Alloys