Feature subset selection for classification is inherently a bi-objective optimization problem, where the task is to obtain a feature subset yielding maximum possible area under the receiver operating characteristic curve by minimizing cardinality. In today's world, an humungous amount of data is generated in all human activities. To mine such voluminous data, which is often high-dimensional, there is a need to develop parallel and scalable frameworks. In the first-of-its-kind study, we proposed and developed three chaotic bi-objective evolutionary algorithms based wrappers with a migration strategy under Spark, namely, (i) parallel chaotic non-dominated sorting algorithm (P-C-NSGA-II-IS), (ii) parallel chaotic non-dominated sorting particle swarm optimization (P-C-NSPSO-IS), and (iii) parallel chaotic multi-objective evolutionary algorithm based on decomposition (P-C-MOEA/D-IS). We employed logistic map and tent map for each of the parallel chaotic algorithm. The performance of the chaotic variants is compared with their corresponding parallel, non-chaotic algorithms. Throughout the study, AUC is computed by invoking the logistic regression classifier on various datasets. The experimental results demonstrate that P-C-NSGA-II-LM-IS, P-NSPSO-IS and P-NSGA-II-IS secured top-3 in terms of mean HV and Formula 1 racing based ranking. We also presented the statistical test of significance, empirical attainment plots, speedup analysis, and mean AUC obtained by the most repeated feature subset, and diversity analysis using hypervolume.

Gaining the trust and confidence of customers is the essence of the growth and success of financial institutions and organizations. Of late, numerous instances of fraudulent activities impacted the financial industry. Furthermore, owing to the generation of large voluminous datasets, it is highly essential that the underlying framework is scalable and meets real-time needs. To address this issue, in the study, we proposed ATM fraud detection in static and streaming contexts respectively. In the static context, we investigated parallel and scalable machine learning algorithms for ATM fraud detection that are built on Spark and trained with various machine learning (ML) models, including Naive Bayes (NB), Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), Gradient Boosting Tree (GBT), and Multi-layer perceptron (MLP). We also employed several balancing techniques such as Synthetic Minority Oversampling Technique (SMOTE) and its variants, Generative Adversarial Networks (GAN), to address the rarity in the dataset. Additionally, we proposed a streaming-based ATM fraud detection in the streaming context. Our sliding window-based method collects ATM transactions that are performed within a specified time interval and then utilizes them to train several ML models, including NB, RF, DT, and K-Nearest Neighbour (KNN). We selected these models based on their less model complexity and quicker response time. In both contexts, RF turned out to be the best model. RF obtained the best mean AUC of 0.975 in the static context and a mean AUC of 0.910 in the streaming context. RF is also empirically proven to be statistically significant than the next-best performing models.

Federated Learning (FL) has gained significant traction due to its decentralized training approach while preserving privacy. The Extreme Learning Machine (ELM) emerged as a competitive technique for diverse classification and regression tasks, owing to its advantages such as fewer trainable parameters and expedited training times. Nonetheless, it suffers from overfitting and suboptimal generalization. Motivated by these considerations, we introduce FedELM-SGD, an iterative variant of ELM which integrates mini-batch stochastic gradient descent (SGD) within privacy-preserving federated learning settings. We also propose an extreme learning factory under federated settings (FedELF) to counteract the impact of random parameter initialization in ELM. In this study, we outline a two-stage methodology: firstly, the generation of privacy-preserving datasets through perturbation of real datasets using Differential Privacy techniques; secondly, the invocation of either FedELM-SGD or FedELF to construct a globally shared classification model. We demonstrate the efficacy of our proposed methods across various finance and medical-related datasets.

Effective risk management and compliance adherence are paramount for the success of financial institutions and organizations. However, they often face significant challenges due to fraudulent activities, with ATM fraud, among others, emerging as a prevalent issue in today's banking landscape. We proposed a novel profiling-based one-class classification (OCC) method to solve this problem. Then training phase of our approach employs the K-Means clustering algorithm to cluster non-fraudulent transactions exhibiting similar characteristics and patterns. A rule is generated from each cluster, thereby in a rule set comprising K rules, each consisting of conditions based on the lower and upper bounds on all features. This rule set is employed to identify fraudulent transactions presented in the test phase because ours is an OCC method. One distinctive feature of our approach is its interpretability and explainability, which is crucial for understanding the model's predictions. Overall, our proposed approach demonstrates the best performance vis-à-vis that of various state-of-the-art OCC methods in terms of classification rate. Additionally, we provide sensitivity analysis by varying the number of conditions violated across the K rules.

Given the inherent rarity observed in imbalanced datasets, adopting one-class classification (OCC) stands out as a pragmatic approach to counteract bias toward the predominant class. This research endeavors to thoroughly assess ten state-of-the-art OCC methodologies across a spectrum of five diverse challenges within the Banking, Insurance, and Cybersecurity sectors. Moreover, we introduce an innovative unsupervised learning technique wherein (i) K-Means clustering is utilized during the training phase on negative sample data. To determine the optimal number of clusters, we utilized the Silhouette index and employed the maximum intra-class centroid distance as a cluster-specific threshold. These thresholds play a pivotal role in distinguishing between positive and negative samples. (ii) In the testing phase, a majority voting mechanism is employed to evaluate the discriminative capability of these thresholds, facilitating precise classification of test data. Empirical findings unequivocally demonstrate that our proposed approach outperforms several state-of-the-art techniques, achieving superior classification accuracy across four out of the five datasets. This underscores the effectiveness and potential applicability of our novel methodology in tackling the intricate challenges prevalent in the Banking, Insurance, and Cybersecurity sectors, particularly in the domains of fraud detection and related areas.

Online streaming feature subset selection (OSFSS) presents a noteworthy challenge when data samples arrive rapidly and in a time-dependent manner. The complexity of this problem is further exacerbated when features arrive as a stream. Despite several attempts to solve OSFSS over feature streams, existing methods lack scalability, cannot handle interaction effects among features, and fail to efficiently handle high-velocity feature streams. To address these challenges, we propose a novel wrapper-for OSFSS named OSFSS-W (wrapper-for OSFSS), specifically designed to mine feature streams within the Apache Spark environment. Our proposed method incorporates (i) two vigilance tests: for removing (a) irrelevant features and (b) redundant features (ii) incremental learning and (iii) a tolerance-based feedback mechanism that retains and utilizes previous knowledge while adhering to the predefined tolerance thresholds. Additionally, for the purpose of optimization, we introduce a Bare Bones Particle Swarm Optimization (BBPSO-L) algorithm driven by the logistic distribution. Further, the BBPSO-L is parallelized within Apache Spark, following an island-based approach. We evaluated the performance of the proposed algorithm on the datasets taken from the cybersecurity, bioinformatics, and finance domains. The results demonstrate that incorporating two vigilance tests coupled with a tolerance-based feedback mechanism significantly improved the median Area under the receiver operating characteristic curve (AUC) and median cardinality across all datasets.

Feature subset selection (FSS) employing a wrapper approach is fundamentally a combinatorial optimization problem maximizing the area under the receiver operating characteristic curve (AUC) of a classifier built on this subset under single objective environment. To balance both the AUC and the cardinality of the selected feature subset, we propose a novel multiplicative fitness function that combines AUC and a decreasing function of cardinality. Although the differential evolution algorithm is robust, it is prone to premature convergence, which can result in entrapment in local optima. To address this challenge, we propose chaotic binary differential evolution coupled with feature-level elitism (CE-BDE), where the chaotic maps are introduced at the initialization and the crossover operator. We also introduce feature-level elitism to improve the exploitation capability. Feature-level elitism involves preserving those features, which are chosen based on their frequency of occurrence in the population in the evolution process. Dealing with big data entails computational complexity, which motivates us to propose an effective parallel/ distributed strategy island model. The results demonstrate that the parallel CE-BDE outperformed the rest of the algorithms in terms of mean AUC and cardinality. The speedup and computational gain yielded by the proposed parallel approach further accentuate its superiority. Overall, the top-performing algorithm with the multiplicative fitness function turned out to be statistically significant compared to that with the additive fitness function across 5 out of 6 datasets.

This paper proposes a bi-objective evolutionary approach to perform technical indicator-based stock trading. The objective is to find the optimal combinations of technical indicators in order to generate buy and sell strategies such that the objective functions, namely, Sharpe ratio and Maximum Drawdown, are maximized and minimized, respectively. In this study, Adaptive geometry-based MOEA (AGE-MOEA) and AGE-MOEA2 are proposed to accomplish the optimization owing to their popularity and power. This study incorporates a rolling-window-based approach (two years of training followed by a year for testing), and thus, the results of the approach seem to be considerably better in stable periods without major economic fluctuations. For the baseline comparison purpose, we employ Non-dominated sorting genetic algorithm-II (NSGA-II), Multi-objective evolutionary algorithm based on decomposition (MOEA/D) too for the problem. Further, we incorporate the transaction cost and domain expertise in the whole modeling approach. It is observed that AGE-MOEA turned out to be the best in 6 out of 11 time horizons by devising a better optimal strategy. However, MOEA/D selected less number of indicators in most of the buy strategy cases and stood first in terms of interpretability. The same observation is noticed with AGE-MOEA in sell strategy cases.

The clever hybridization of quantum computing concepts and evolutionary algorithms (EAs) resulted in a new field called quantum-inspired evolutionary algorithms (QIEAs). Unlike traditional EAs, QIEAs employ quantum bits to adopt a probabilistic representation of the state of a feature in a given solution. This unprecedented feature enables them to achieve better diversity and perform global search, effectively yielding a trade-off between exploration and exploitation. We conducted a comprehensive survey across various publishers and gathered 56 papers published between 2002 and 2024. We thoroughly analyzed these publications, focusing on the novelty elements and types of heuristics employed by the extant quantum-inspired evolutionary algorithms (QIEAs) proposed to solve the feature subset selection (FSS) problem. Importantly, we provided a detailed analysis of the different types of objective functions and popular quantum gates, i.e., rotation gates, employed throughout the literature. Further, we provided merits and demerits of QIEAs vis-à-vis EAs. Additionally, we suggested several open research problems which could be extremely useful to the budding researchers.

This is a comprehensive survey of the applications of causal inference in the Banking, Financial Services and Insurance (BFSI) domain based on 45 papers published from 1992 to 2023. It categorizes papers into (i) Banking and risk management (ii) Finance (covering investment, asset and portfolio management; behavioral finance and time series), (iii) Financial markets and (iv) Insurance. Exploring methods such as Bayesian Causal Network, Granger Causality, and counterfactuals, the article emphasizes significance of causal inference in explaining predictions of AI/ML models. This survey also recommends promising future research directions in the intersection of causal inference and these domains making it helpful for the professionals working therein.

Machine Learning (ML) has successfully made inroads into almost every field. This very fact makes the ML models a target for fraudsters who perpetrate various adversarial attacks, thereby hindering the performance of ML models. Evasion and data-poisoning-based attacks are more notorious, especially in finance, healthcare, and other critical sectors. This motivated us to propose a novel, computationally less expensive method for generating adversarial samples by employing a Variational Autoencoder (VAE). It is well known that the Wavelet Neural Network (WNN) is considered computationally efficient in solving image and audio processing, speech recognition, and time-series forecasting. This paper proposes a VAE-Deep-Wavelet neural network (VAE-Deep-WNN), where the encoder and decoder employ a WNN instead of a multi-layer perceptron (MLP). Recently, a chaotic VAE (C-VAE) was reported to be more effective in one-class classification [1], which motivated us to propose a chaotic variant of Deep-WNN and MLP-based VAE, named C-VAE-Deep-WNN and C-VAE-MLP, respectively. Our chaotic variants employed a logistic map to generate chaotic numbers, replacing the random noise in the latent space of the traditional VAE. In this paper, we performed both vanilla and chaos VAE-based adversary sample generation and applied them to various problems related to finance and cybersecurity domains, such as loan default, credit card fraud, and churn modeling, etc; we performed both evasion and data-poisoning attacks on Logistic Regression, Decision Tree, Gradient boosting, Light gradient boosting models. In the majority of the datasets, VAE-Deep-WNN/C-VAE-Deep-WNN outperformed the other VAE-based variants in both evasion and data poisoning attacks.

Gaining the empathy and trust of customers is paramount in the financial domain. However, the recurring occurrence of fraudulent activities undermines both of these factors. ATM fraud is a prevalent issue faced in today's banking landscape. The critical challenges in fraud datasets are highly imbalanced datasets, evolving fraud patterns, and lack of explainability. In this study, we handled these techniques on an ATM transaction dataset collected from India. In binary classification, we investigated the effectiveness of various oversampling techniques, such as the Synthetic Minority Oversampling Technique (SMOTE) and its variants, Generative Adversarial Networks (GAN), to achieve oversampling. Gradient Boosting Tree (GBT), outperformed the rest of the techniques by achieving an AUC of 0.963, and Decision Tree (DT) stands second with an AUC of 0.958. In terms of complexity and interpretability, DT is the winner. Among the oversampling approaches, SMOTE and its variants performed better. We incorporated explainable artificial intelligence (XAI) and Causal Inference (CI) in the fraud detection framework and studied them via various analyses. Further, we provided managerial impact.

Federated Learning (FL) has garnered widespread attention in finance, banking, and healthcare due to its decentralized, distributed training and the ability to protect privacy while obtaining a global shared model. However, FL faces challenges such as communication overhead and limited resource capability. This motivated us to propose a first-of-its-kind, two-stage FL approach as follows: (i) During phase I, under non-federated settings, synthetic dataset is generated by employing two different probability distributions as noise to the vanilla conditional tabular generative adversarial neural network (CTGAN) resulting in modified CTGAN. We also employed standard metrics to assess the quality of synthetic datasets. (ii) In phase II, the Federated Probabilistic Neural Network (FedPNN) is developed for building globally shared classification model. Despite PNN being a one-pass learning classifier, its complexity depends on the training data size. Therefore, we employed a modified evolving clustering method (ECM), another one-pass algorithm, to cluster the training data, in between the input and pattern layers of the FedPNN. The effectiveness of our approach is validated on credit card fraud detection and Polish bankruptcy prediction datasets.

The need to protect data privacy is critical in industries like healthcare, finance, and banking. Federated Learning (FL) has attracted widespread attention due to its decentralized, distributed training and the potential to protect privacy. However, FL presents challenges like privacy concerns, network, and data heterogeneity. This motivated us to propose a two-phase, one-pass FL algorithm, where (I) In the first phase, the differentially private dataset is generated by incorporating the Laplacian mechanism, and (ii) during the second phase, we employed one-pass privacy-preserving federated extreme learning machine (OP-FedELM) for generating globally shared training dataset and build the global model. ELM generally tends to overfit training data at a client. Therefore, we employed a modified version of the Evolving clustering algorithm (ECM), an online one-pass clustering algorithm, to cluster training datasets at the clients. Further, it will reduce the computational training time of ELM and communication overhead. At the server, we employed a meta-clustering algorithm to cluster the updates from the clients. We also proposed another one-shot FL algorithm called privacy-preserving federated ELM based on minimum least square solution (FedELM-LS). The experimental analysis concludes that OP-FedELM is computationally less expensive yet achieves higher AUC in three of four datasets.

Feature selection (FS) is an important pre-processing step in building machine learning models. minimum Redundancy and Maximum Relevance (mRMR) approach has emerged as one of the successful algorithms in obtaining irredundant feature subset involving only bi-variate computations. In the current digital age, owing to the prevalence of very large scale datasets, an imminent need has arisen for scalable solutions using distributed/parallel algorithms. MapReduce solutions are proven to be one of the best approaches to design fault-tolerant and scalable solutions. This work analyses the existing Horizontal MapReduce approaches for mRMR feature selection and identifies the limitations thereof. It is observed that existing approaches involve redundant and repetitive computations and lacks a metadata framework to diminish them. This motivated us to propose Horizontal partitioning based MapReduce solutions namely HMR_mRMR, is an Iterative MapReduce algorithms and is designed under Apache Spark. Appropriate usage of metadata framework and solution formulation optimizes the computations in the proposed approaches. The comparative experimental study is conducted with existing approaches to establish the importance of HMR_mRMR.

Wavelet neural networks (WNN) have been applied in many fields to solve regression as well as classification problems. After the advent of big data, as data gets generated at a brisk pace, it is imperative to analyze it as soon as it is generated owing to the fact that the nature of the data may change dramatically in short time intervals. This is necessitated by the fact that big data is all pervasive and throws computational challenges for data scientists. Therefore, in this paper, we built an efficient Scalable, Parallelized Wavelet Neural Network (SPWNN) which employs the parallel stochastic gradient algorithm (SGD) algorithm. SPWNN is designed and developed under both static and streaming environments in the horizontal parallelization framework. SPWNN is implemented by using Morlet and Gaussian functions as activation functions. This study is conducted on big datasets like gas sensor data which has more than 4 million samples and medical research data which has more than 10,000 features, which are high dimensional in nature. The experimental analysis indicates that in the static environment, SPWNN with Morlet activation function outperformed SPWNN with Gaussian on the classification datasets. However, in the case of regression, there is no clear trend was observed. In contrast, in the streaming environment i.e., Gaussian outperformed Morlet on the classification and Morlet outperformed Gaussian on the regression datasets. Overall, the proposed SPWNN architecture achieved a speedup of 1.22-1.78.

Feature selection stands out to be an important preprocessing step that is used to handle the uncertainty and vagueness in the data. In recent times, the minimum Redundancy and Maximum Relevance (mRMR) approach has been proven to be effective in obtaining the irredundant feature subset. Owing to the generation of voluminous datasets, it is essential to design scalable solutions using distributed/parallel paradigms. MapReduce solutions are proven to be one of the best approaches to designing fault-tolerant and scalable solutions. This work analyses the existing vertical partitioning MapReduce approaches for mRMR feature selection and identifies the limitations thereof. In the current study, we proposed VMR_mRMR, an efficient vertical partitioning-based approach using a memorization approach thereby overcoming the extant approaches limitations. The experiment analysis says that VMR_mRMR significantly outperformed extant approaches and achieved a better computational gain (C.G). We also conducted a comparative analysis with the horizontal partitioning approach HMR_mRMR to assess the strengths and limitations of the proposed approach.

Performing online feature subset selection (OFS) when data samples arrive at a high velocity in a time-dependent manner is a critical problem to solve. The situation becomes more difficult when the features also arrive in a stream. Several efforts were made by the researchers to perform OFS over feature streams. However, they are not scalable and cannot analyze feature streams coming at a high velocity. Further, optimal feature subsets must be identified by scalable approaches. It is noteworthy that evolutionary algorithms (EAs) which are inherently parallel and scalable are least employed for OFS in a streaming feature case. This motivated us to address the challenges and propose a generic EA-based wrapper for OFS to mine feature streams under the Apache Spark environment.

In a first-of-its-kind study, we propose an online feature selection (OFS) framework for streaming data under big data paradigm, by proposing (i) parallel hybrid particle swarm optimization (PSO)-based wrapper, (ii) a robust method to handle high velocity and voluminous datasets, and (iii) two vigilance tests for detecting novelty. Our framework involves a continuous and adaptive learning process, by reducing the number of retrains of the wrapper. Moreover, it is scalable by virtue of the parallelization of the PSO and its variants/hybrids resulting in quick responses in real-time. We proposed BBPSO-L+TNS, a hybrid of bare-bones particle swarm optimization guided by logistic distribution (BPPSO-L) and threshold based neighbourhood search (TNS) heuristic, to achieve better exploitation capability and avoid entrapped in local optima. The findings demonstrate the robustness of the proposed streaming framework, yielding cost-effective solutions. Further, BBPSO-L+TNS outperformed the baseline algorithms.

Literature abounds with various statistical and machine learning techniques for stock market forecasting. However, Reinforcement Learning (RL) is conspicuous by its absence in this field and is little explored despite its potential to address the dynamic and uncertain nature of the stock market. In a first-of-its-kind study, this research precisely bridges this gap, by forecasting stock prices using RL, in the static as well as streaming contexts using deep RL techniques. In the static context, we employed three deep RL algorithms for forecasting the stock prices: Deep Deterministic Policy Gradient (DDPG), Proximal Policy Optimisation (PPO) and Recurrent Deterministic Policy Gradient (RDPG) and compared their performance with Multi-Layer Perceptron (MLP), Support Vector Regression (SVR) and General Regression Neural Network (GRNN). In addition, we proposed a generic streaming analytics-based forecasting approach leveraging the real-time processing capabilities of Spark streaming for all six methods. This approach employs a sliding window technique for real-time forecasting or nowcasting using the above-mentioned algorithms. We demonstrated the effectiveness of the proposed approach on the daily closing prices of four different financial time series dataset as well as the Mackey–Glass time series, a benchmark chaotic time series dataset. We evaluated the performance of these methods using three metrics: Symmetric Mean Absolute Percentage (SMAPE), Directional Symmetry statistic (DS) and Theil’s U Coefficient. The results are promising for DDPG in the static context and GRNN turned out to be the best in streaming context. We performed the Diebold–Mariano (DM) test to assess the statistical significance of the best-performing models.

In this paper, we solve the feature subset selection (FSS) problem with three objective functions namely, cardinality, area under receiver operating characteristic curve (AUC) and Matthews correlation coefficient (MCC) using novel multi-objective evolutionary algorithms (MOEAs). MOEAs often encounter poor convergence due to the increase in non-dominated solutions and getting entrapped in the local optima. This situation worsens when dealing with large, voluminous big and high-dimensional datasets. To address these challenges, we propose parallel, fractional dominance-based MOEAs for FSS under Spark. Further, to improve the exploitation of MOEAs, we introduce a novel batch opposition-based learning (BOP) along with a cardinality constraint on the opposite solution. Accordingly, we propose two variants, namely, BOP1 and BOP2. In BOP1, a single neighbour is randomly chosen in the opposite solution space, whereas in BOP2, a group of randomly chosen neighbours in the opposite solution space. In either case, the opposite solutions are evaluated to improve the exploitation capability of the underlying MOEAs. We observe that in terms of mean optimal objective function values and across all datasets, the proposed BOP2 variant of parallel fractional dominance-based algorithms emerges as the top performer in obtaining efficient solutions. Further, we introduce a novel metric, namely the ratio of hypervolume (HV) and inverted generated distance (IGD), HV/IGD, that combines both diversity and convergence. With respect to the mean HV/IGD computed over 20 runs and Formula 1 racing, the BOP1 variants of fractional dominance-based MOEAs outperformed other algorithms.

Multiple sentences that reveal the same meaning are considered to be paraphrases. Paraphrases restate a given text, passage or statement using different words in which the original context and the meaning are kept intact. It can be used to expand, clarify or summarize the content of essays, research papers and journals. Semantic identity of sentences are detected during the process of paraphrase detection. Paraphrase detection can be related to different applications, like plagiarism detection, text summarizing, text mining, question answering, and query ranking, in the domain of Natural Language Processing. Effective paraphrase detection could be implemented if the semantics of the language and their interactions are adequately captured. The process of paraphrase detection is considered to be a difficult and challenging task due to the wide range of complex morphological structures and vocabulary that prevails in most of the Indian languages. The approaches that exist for paraphrase detection include machine learning techniques like Multinomial Logistic Regression model and Recursive Auto Encoders, which lacks in hand-crafted feature engineering. The problem could be solved when deep learning approaches are used for paraphrase detection. In the proposed system, the classification of paraphrase, semi-paraphrase and non-paraphrase sentences are implemented using an ensemble of three deep learning algorithms which includes BERT (Bidirectional Encoder Representations from Transformers), USE (Universal Sentence Encoder) and Seq2Seq (Sequence to Sequence). The DPIL corpus has been used for the evaluation of the proposed system and the highest accuracy obtained considering languages Hindi and Punjabi are 85.22% and 85.80% respectively.

This paper proposes to model chaos in the automated teller machine (ATM) cash withdrawal time series of a large Indian commercial bank and forecast the withdrawals using deep learning (DL) and hybrid DL methods. It also considers the influence of “day-of-the-week” on the results. We first modelled the chaos present in the withdrawal time series by reconstructing the state space of each series using the optimal lag and embedding dimension. This process converts the original univariate time series into a multi variate time series. The “day-of-the-week” dummy variable is converted into seven variables using one-hot encoding and augmented to the multivariate or univariate time series depending on whether chaos was present or absent. For forecasting the future cash withdrawals, we employed (i) statistical technique namely autoregressive integrated moving average (ARIMA), (ii) machine learning techniques such as random forest (RF), support vector regression (SVR), multi-layer perceptron (MLP), group method of data handling (GMDH), and general regression neural network (GRNN), and (iii) DL techniques such as long short term memory (LSTM) neural network, Gated Recurrent Unit (GRU) and 1-dimensional convolutional neural network (1D-CNN). We also explored hybrid DL techniques such as 1D-CNN + LSTM and 1D-CNN + GRU. We observed improvements in the forecasts for all techniques when “day-of-the-week” variable was included. It is observed that chaos was present in 28 ATMs, whereas in the remaining 22 ATMs chaos was absent. In both the cases, LSTM yielded the best Symmetric Mean Absolute Percentage Error (SMAPE) on the test data. However, LSTM showed statistically different performance than the 1D-CNN + LSTM in chaos category but equal performance with 1D-CNN in non– category yet statistically significant than 1D-CNN.

Extant sequential wrapper-based feature subset selection (FSS) algorithms are not scalable and yield poor performance when applied to big datasets. Hence, to circumvent these challenges, we propose parallel and distributed hybrid evolutionary algorithms (EAs) based wrappers under Apache Spark. We propose two hybrid EAs based on the Binary Differential Evolution (BDE), and Binary Threshold Accepting (BTA), namely, (i) Parallel Binary Differential Evolution and Threshold Accepting (PB-DETA), where BDE and BTA work in tandem in every iteration, and (ii) its ablation variant, Parallel Binary Threshold Accepting and Differential Evolution (PB-TADE). Here, BTA is invoked to enhance the search capability and avoid premature convergence of BDE. For comparison purposes, we also parallelized two state-of-the-art algorithms: adaptive DE (ADE) and permutation based DE (DE-FSPM), and named them PB-ADE and P-DE-FSPM respectively. Throughout, logistic regression (LR) is employed to compute the fitness function, namely, area under the receiver operator characteristic curve (AUC). The effectiveness of the proposed algorithms is tested over the five big datasets of varying dimensions. It is noteworthy that the PB-TADE turned out to be statistically significant than the rest. All the algorithms have shown the repeatability property. The proposed parallel model attained a speedup of 2.2–2.9. We also reported feature subset with high AUC and least cardinality.