Big Data Analytics (BDA) is progressively becoming a popular practice implemented by many organizations, because of their potential to discover valuable in-sights for improved decision-making. The International Data Corporation predicts that the Global Datasphere will grow from 33 Zettabytes in 2018 to 175 Zettabytes in 2025. Obviously, we are currently in the era of Big Data (BD), and the rate of data growth is very alarming. Unfortunately, BD does not offer a lot of value in its unprocessed form. Therefore, to unlock the great potentials of BD, we need efficient BDA methods. Machine Learning (ML) algorithms are one of the most efficient tools suitable for data analytics, however, some ML algorithms cannot effectively handle BD; their computational complexity increases with in-crease in data size. Therefore, some researchers introduced various techniques for improving the speed of ML algorithms, including feature selection techniques, in-stance selection techniques, sampling, and distributed computing. However, most of them failed to achieve a balanced trade-off between storage reduction and predictive accuracy [1]. Therefore, this paper introduces a boundary detection and instance selection technique for improving the speed of ML-based BDA, called Ant Colony Optimization Instance Selection Algorithm for Machine Learning (ACOISA_ML)). The key highlights of ACOISA_ML are outlined below: Boundary identification: The first stage of ACOISA_ML is the boundary identification stage. Unlike other ACO-based instance selection techniques that directly use ACO algorithm for instance or feature selection, ACOISA_ML use ACO algorithm for boundary identification. It adopts the concept of ACO edge selection to search for different boundaries (not to select instances). To the best of author's knowledge, this study is one of the first studies that adopt the concept of ACO edge detection for instance selection problems. This concept is mostly used for image edge detection (and not data boundary identification) [2][3][4].
The second stage of the proposed technique is the boundary instance selection stage. After identifying different boundaries, ACOISA selects the best boundary and use k-NN to select the relevant instances for training (that is, instances close to the best-identified boundary).
This study introduces a novel method for computing heuristic value for ACO. This method is suitable for boundary instance se-lection problems. ACOISA_ML is designed to use the proposed computation method to cal-
culate the heuristic value for each instance in the dataset. As afore-mentioned, ACO is used to identify the best boundary instance, that is, the in-stance with the highest pheromone value. Hence, the heuristic value for each in-stance is designed to reflect the boundary information for each instance.
The technique was evaluated on five ML algorithms, namely: (Artificial Neural Network (ANN), Random Forest (RF), Naïve Bayes (NB), k Nearest Neighbor (k-NN), and Logistic Regression (LR)). In this study, we refer to the models produced by the full dataset as standard models, and the models produced by the reduced subset as hybrid models. Finally, we compare the hybrid models to the standard models based on the following criteria: (i) the ability to preserve prediction accuracy (ii) training speed (iii) storage reduction percentage, and (iv) algorithm time (or instance selection time). All the datasets used in this study were obtained from the UCI data repository [5]. Annexures 1 and 2 shows the average training speed and prediction accuracy produced by the standard models (denoted as Standard) and hybrid models (de-noted as Hybrid). As shown in the Annexures, the hybrid models achieved better training speed than the standard models without significantly affecting their pre-diction accuracy. Moreover, the right-hand side of Annexure 1 shows the average algorithm time (denoted as Alg-T) and the average storage reduction percentage (denoted as Av-Sto) achieved by ACOISA_ML. The storage reduction percentage represents the fraction of instances selected after data reduction. As shown in the Annexure, ACOISA_ML reduced the storage size of the evaluated big datasets by over 55% (in most cases) without substantially affecting their quality. Moreover, ACOISA_ML achieved good instance selection time. It used an average of 36.8 seconds to reduce the largest dataset evaluated in this study (i.e. Twitter dataset). This shows the effectiveness of ACOISA_ML for BDA.
In addition, ACOISA_ML was compared to four recent instance selection algorithms, namely: LDIS, LSSM, LSBO, and ISDSP. The algorithms were evaluated on SVM, hence we first evaluated ACOISA_ML on SVM before comparing it to the algorithms. Annexure 3 shows the prediction accuracy (denoted as accuracy) and storage reduction percentage (denoted as storage) for the four algorithms. The best prediction accuracy for each dataset is underlined. As shown, ACOISA_ML outperformed LSSM in prediction accuracy in 6 out of 11 datasets and outperformed LSBO in 7 out of 11 datasets. Moreover, the results show that ACOISA_ML outperformed LDIS and ISDSP in 9 out of 11 datasets. Furthermore, T-test statistical analysis was performed to evaluate the speed improving capacity of ACOISA_ML. Specifically, we compared the training speed produced by the hybrid models of ANN and RF to the training speed produced by the standard algorithms. The P-values produced by all the test analysis are less than 0.05, hence we can conclude with 95% confidence level that ACOISA_ML is significantly faster, in terms of training speed, than the analyzed standard algorithms. Overall, the results show that the proposed technique is suit-able for fast and simplified BDA and ML speed optimization.
Keywords: Big data analytics, Machine learning, Instance selection, Data reduction, Speed optimization.