Feature Selection Techniques for Big Data Analytics

Abstract

Big data applications have tremendously increased due to technological developments. However, processing such a large amount of data is challenging for machine learning algorithms and computing resources. This study aims to analyze a large amount of data with classical machine learning. The influence of different random sampling techniques on the model performance is investigated by combining the feature selection techniques and machine learning classifiers. The experiments used two feature selection techniques: random subset and random projection. Two machine learning classifiers were also used: Naïve Bayes and Bayesian Network. This study aims to maximize the model performance by reducing the data dimensionality. In the experiments, 400 runs were performed by reducing the data dimensionality of a video dataset that was more than 40 GB. The results show that the overall performance fluctuates between 70% accuracy to 74% for using sampled and non-sample (all the data), a slight difference in performance compared to the non-sampled dataset. With the overall view of the results, the best performance among all combinations of experiments is recorded for combination 3, where the random subset technique and the Bayesian network classifier were used. Except for the round where 10% of the dataset was used, combination 1 has the best performance among all combinations.

1. Introduction

Big data applications have tremendously increased due to technological developments that led to increased data size and the conversion of regular data into large datasets. Big data in the form of extensive data thus requires high-speed servers for speedy processing [1]. Large servers are needed to save this big data that make the data available on a request basis [2]. Big data assists decision-making and validation in organizational processes [3]. Conversely, there always is a critical trade-off between application size and efficiency, i.e., the larger the application data, the lower the efficiency of the application [4]. Big data applications always require large amounts of data to model the system, whereas extensive data requires significant storage and methods to handle the data efficiently. To handle such scenarios, big data handling methods are required to divide the data into subgroups and handle them in the same manner as the source data [5].

Besides data handling, big data is also prone to risks and constraints such as data validity, theoretical relevance, appropriate attribute association, controls, audibility, and precision. These parameters are meant to ensure the quality of information and the quality of big data [6]. In addition to these constraints, many other factors are associated with big data, such as data security, sorting the data, the management of servers, and privileges related to data [7]. The number of digital tools for data handling exceeded almost 92% by 2002 and is still increasing, leading to the big data business of about 46.4 billion [8].

1.1. Motivation

Besides numerous big data applications, it has become challenging for applications in semantic networks, data mining, social networks, and information fusion [9]. Likewise, many research interests were developed in pattern mining, data tracking, data storage, data visualization, analysis of user behavior, and data processing [10]. This led to big data solutions in the form of technologies such as computational intelligence and machine learning that made it possible to attain many solutions using these technologies for scanning and processing data. These solutions include data condensation, incremental learning, distributed computing, divide and conquer, data sampling density-based approaches, and others [8,9,10].

Sampling data has been of more importance in data handling problems with the issues of computational burden, complexity, and inefficiency associated with the tasks under consideration [11]. The richness of the data quality is compromised in most scenarios due to biased estimations made per sample [12]. To handle this issue, reverse sampling procedures are exploited using information from external sources where big data is an ensemble with probabilistic sampling [13]. Sampling size is the most crucial factor for the system’s accuracy [14]. The processes of non-probabilistic sampling, Zig Zag, inverse sampling, and cluster sampling have been introduced as a solution for big data sampling [13,15,16].

This study aims to analyze a large amount of data and sample it randomly into subsets. The experiments show the influence of different techniques on the model performance is analyzed by combining the feature selection techniques and machine learning classifiers. The experiments use two feature selection techniques: random subset and random projection. It is quite possible that the random selection of a small portion of the data can produce as good results as the original data, making processing the whole data a waste of computing resources. It is worth saying that if a slight difference in performance can be achieved when the entire dataset is used, then it is possible to be neglected in favor of using just a small portion of the data with a close performance rate of the whole data.

1.2. Related Work

The information from data can be learned via machine learning algorithms to generate decision-making and prediction models [16]. The machine learning techniques can learn the behavior and trends in data for future predictions by training the model via communication, comparisons, problem-solving, discoveries, and strategies [4]. The more significant the amount of data, the higher the accuracy of machine learning, but machine learning good performance also depends on the data’s simplicity. This leads to the problem of machine learning and big data with unstructured data, unclassified data, and rapidly changing data [17].

One of the deep learning paradigms, such as Convolutional Neural Networks (CNNs), can deal with the problem of data classification [18,19] for images and textual data [20]. CNN has an excellent performance in image data classification and detection, but it requires large amounts of input data and, subsequently, high processing power. The CNN comprises multiple layers such as convolutional, pooling, and fully connected layers that require enormous resources to perform efficiently. The feature extraction mechanisms have gained significant attention [21,22] due to their ability to reduce massive data optimally. The dimensionality of data affects the performance of machine learning techniques and data handling mechanisms. More significant data require technological resources such as powerful processing tools unavailable in most scenarios.

The data attributes play a vital role in machine learning and data handling to develop better models, but they may complicate the scenarios due to inappropriate data coverage and classification. Many attributes and instances create complex dimensionality issues in large data sets. Thus, this article explores the NDPI video dataset [23] divided into three categories: Acceptable data, flagged data, and unacceptable data for analysis. The data from image filtering is utilized for data sampling. The data used is well organized for three reasons, i.e., it is well categorized into three categories, the data can be converted into numerical values, and a large amount of data, up to 40GB, is available. This makes a case for large data processing and is thus appropriate for comprehensive machine learning-based analysis.

Many research articles, such as [24,25,26,27] in the literature, aim to solve similar problems in identical domains. The work in [22,28] explores large datasets with many classes to increase productivity and efficient machine learning models. The color transformation methods are studied in [26]. Several evidence-based and adaptive sampling methods are explored for filtering in [29,30,31]. The analysis for website filtering is provided in [31]. The analysis of keyframes is illustrated in [32]. The research works in [33] and [34] offer functions to visual attributes to make multimedia accessible. Content retrieval applications are explored in [35,36,37,38,39,40,41,42]. Feature analysis and reduction based on several related areas are also explored in [19,20,41,42,43,44,45,46,47,48]. Neighborhood rough sets are proposed in [49] as tools for reducing the attributes in big data. This method provides the best choice for attribute selection. A hierarchical framework based on supervised models is proposed in [50] using a support vector machine as a machine learning algorithm to reduce the attributes in big data. Gabor filters are exploited for noise reduction, and Elephant Herd Optimization is used for feature selection. Two effective feature selection methods, such as Principal Component Analysis and Linear Discriminant Analysis, are exploited in [51] to reduce the attribute sets for machine learning algorithms: Random Forest, Naïve Bayes, Support Vector Machine, and Decision Tree. The dominance-based neighborhood rough sets (DNRS) method is exploited in [52] for parallel attribute reduction that considers partial order for numerical and categorical attributes. Neighborhood decision with some consistency is explored for attribute reduction based on multi-criterion [53]. The classification variations in varying attribute scenarios are handled with neighborhood decision consistency. Reduced error is attained with a new attribute reducing method, a heuristic method to derive the redact.

Recently, Rostami et al. [54] offered a genetic algorithm based on community detection for the aim of feature selection, which acts in three phases. The feature similarities are determined in the first stage. During the second step, community detection algorithms classify the features into clusters. In the third stage, a genetic algorithm is used to select traits for a new community-based repair procedure. Nine benchmark classification problems were analyzed in terms of the performance of the presented approach. Additionally, the authors have compared the efficiency of the suggested technique with the results from four known algorithms for feature selection. Comparing the performance of the proposed technique with three new feature selection methods based on PSO, ACO, and ABC algorithms on three classifiers indicated that the accuracy of the proposed method is on average 0.52% higher than the PSO, 1.20% higher than ACO, and 1.57 higher than the ABC algorithm. Rajendran et al. [55] concentrate on the development of a big data classification model using chaotic pigeon-inspired optimization (CPIO)-based feature selection in conjunction with an optimum deep belief network (DBN) model. The suggested model is performed in the Hadoop MapReduce environment to handle big data. The CPIO method is first employed to pick a subset of valuable features. The Harris Hawks Optimization (HHO)-based DBN model is also created as a classifier to provide suitable class labels. The invention of the HHO method to adjust the hyperparameters of the DBN model contributes to the improvement of classification performance. Several simulations were conducted to determine the superiority of the provided approach, and the results were analyzed from many dimensions.

In a separate effort, Rostami et al. [56] conducted a comparative analysis of several feature selection approaches and categorized these methods generally. In addition, the current state of the art in swarm intelligence is examined, as are the most recent feature selection approaches based on these algorithms. Furthermore, the merits and limitations of the various examined feature selection approaches based on swarm intelligence are appraised. Song et al. [57] present a novel three-phase hybrid Feature Selection technique (HFS-C-P) based on correlation-guided clustering and particle swarm optimization (PSO) to address the two difficulties mentioned above simultaneously. To do this, the suggested algorithm integrates three types of Feature Selection approaches depending on their benefits. In the first and second stages, a filter Feature Selection approach and a feature clustering-based method with low computing cost are developed to limit the search space required in the third phase. The third step then involves locating an ideal subset of features using an evolutionary algorithm with global searchability. In addition, a symmetric uncertainty-based feature deletion approach, a rapid correlation-guided feature clustering strategy, and an enhanced integer PSO are proposed to improve the performance of the three phases, respectively. The suggested technique is finally evaluated on 18 publicly accessible real-world datasets in contrast to nine Feature Selection algorithms.

Jain et al. [58] proposed a model that undergoes initial preprocessing to eliminate unwanted words. The set of feature vectors is then extracted using Term Frequency-Inverse Document Frequency (TF-IDF) as a feature extraction technique. In addition, a Binary Brain Storm Optimization (BBSO) algorithm is applied to the Feature Selection procedure, resulting in enhanced classification performance. In addition, Fuzzy Cognitive Maps (FCMs) are used as a classifier to categorize the incidence of positive or negative emotions. A comprehensive analysis of experimental results ensures that the presented BBSO-FCM model performs better on the benchmark dataset. Abu Khurma et al. [59] present a complete summary of 156 papers concerning NIA’s improvements for combating Feature Selection. They supplement the conversations with analytical perspectives, illustrated data, practical examples, and open-source software solutions and debate Feature Selection and NIA-related open topics. The study concludes with a summary of the fundamentals of NIAs-Feature Selection, investigating around 34 distinct operators. Chaotic maps are the most common operator. Hybridization is the most common kind of alteration. There are three forms of hybridization: NIA integration, NIA integration with a classifier, and NIA integration without a classifier. The most prevalent hybridization is the combination of a classifier and the NIA. Medical and microarray applications account for most NIA-Feature Selection modifications and use. Big data has benefited many fields. Recently, besides many new fusions of big data applications, the security paradigm has seen exponential usage of big data [60,61,62]. Security and safety applications need precise, accurate, and expedited decision-making based on big data analytics. Our work also contributes toward the rapid model creation from smaller sets and deploying these models for several application scenarios.

2. Methodology

Our experimental setup analyzes a large amount of data with classical machine learning. Since our main objective is the evaluation and insight into the effect of sampled sets versus the significant sets, we choose the classical feature extraction and classical ML models for experimental assessment and the proof of concepts.

Figure 1 shows the evaluation process. The input images are subjected to feature extraction. For feature extraction, we use the auto-correlogram features as a feature set. There are multiple reasons why we have chosen these feature vectors. Firstly, they are fast and thoroughly researched for extracting meaningful information from the image. Additionally, the auto-correlogram feature set represents the critical information in the image and is computationally feasible to implement, can be efficiently pre-computed, and stored. The selection of auto-correlogram is also based on the pre-comparative study, where we compare auto-correlogram features, Gabor features, Color Layout features, and Pyramid Histogram of Oriented Gradients (PHOG) features. We obtained the highest performance for the auto-correlogram features. As shown in Figure 1, after the feature set is obtained, we can perform three types of evaluation.

• Model creation using the full set of data/attributes.
• Model creation using the reduced set obtained through the random subset.
• Model creation using the reduced set obtained through the random projection.

Figure 1. Proposed evaluation methodology.

Figure 1. Proposed evaluation methodology.

The random subset and random projection can be further used to extract the desired amount of data starting from 10% to 90%. A 100% selection will mean the actual amount of data. Therefore, we do not need to extract the 100% set because it is already available as a full set in Figure 1.

The output of previous steps is thus a reduced or full feature set that the classifiers can use for evaluation. The classifier selected (in the ML algorithm block) is used to learn a model from the features extracted in the previous steps. We employ the two classifiers, Naïve Bayesian and the Bayesian Network, for experimental evaluation and the proof of concepts. We selected probabilistic models, including Naïve Bayesian and Bayesian networks, due to their feasibility for feature independence. The Naïve Bayesian also has the advantage that it can be incrementally trained and tested for large datasets. It is thus an updatable classifier that is optimal for real-time and large sets of data that cannot be loaded entirely into the memory. We believe that the selection of the two machine learning algorithms, Naïve Bayesian and the Bayesian Network, optimally address our proposed evaluation’s theme.

Once a model is learned, 10-fold cross-validation ensures that the model performance is more robust and can apply to real-world data. The performance reported in all the experiments is an average of 10-fold. The experimental evaluation section explains the evaluation of full sets and reduced sets by the random subset and random projection.

2.1. Naïve Bayesian

Naïve Bayes is a type of supervised learning for classification based on the Bayesian Theorem. Naïve Bayes takes its concept on the assumption that there is no relation between the existence and non-existence of one feature over another. It uses the maximum likelihood technique to estimate a parameter. For a set of attributes X: A₁, A₂, A₃, and A₄, the classification function F (X) assumes that the attributes have no parent except C, the primary parent in Naïve Bayes [15] shown in Figure 2. In Naïve Bayes, all the attributes may or may not depend on each other and are equally essential for the model. The technique can generate an estimation based on little data input for training. Depending upon the characteristics of the probability model, the classifier with the most significant value leads the hypothesis. The Naive Bayes classifiers are efficient in many complex real-world situations, despite having a basic design and oversimplified assumptions.

2.2. Bayesian Network

The Bayesian Network has the same assumptions as Naïve Bayes, such as there is no relation between the existence and non-existence of one feature over another [15]. The pictorial illustration of the Bayes Network is provided in Figure 3. The Bayes net structure comprises a root node N, leaves (N₁, N₂, …, N_n), and the edges (E₁, E₂, …, E_n). The nodes can be represented as random variables, which are observable quantities. The edges represent the conditional dependencies, and the leaves represent the hypothesis of the given problem. The network explores all the possible edge combinations to attain an optimal model in the form of an acyclic graph for a given problem statement. The learning process of the Bayesian Network has two stages; the first stage is learning a network structure, and the second is learning the probability tables. Many efficient algorithms perform inference and learning in Bayesian networks.

3. Experimental Evaluation

3.1. Rational

We employ the two classifiers, Naïve Bayesian and the Bayesian Network, for experimental evaluation and the proof of concepts. We selected probabilistic models, including Naïve Bayesian and Bayesian networks, due to their feasibility for feature independence. The Naïve Bayesian also has the advantage that it can be incrementally trained and tested for large datasets. It is thus an updatable classifier that is optimal for real-time and large sets of data that cannot be loaded entirely into the memory. We believe that the selection of the two machine learning algorithms, Naïve Bayesian and the Bayesian Network, optimally addresses the experimental evaluation’s theme.

3.2. Dataset

For experimental evaluation, the NDPI large video dataset [23] is used and divided into three categories: acceptable, flagged, and unacceptable data for analysis. The data from image filtering is utilized for data sampling. The data used are well organized due to three reasons, i.e., they are well categorized into three categories, the data can be converted into numerical representations, and a large amount of data, up to 40 GB, is available, which makes a case for large data processing and thus appropriate for comprehensive machine learning-based analysis.

3.3. Experimental Setup

The experiments of this study aim to analyze the influence of different techniques on the model performance by combining feature selection techniques and machine learning classifiers. The features are extracted using the autocorrelogram approach. The autocorrelogram approach considers color and texture and caters to the spatial arrangements in the image. To evaluate smaller and large sets, two feature selection techniques are used in the experiments; random subset and random projection. Two machine learning classifiers were used as well; Naïve Bayes and Bayesian Network. Our experiment approach maximizes the model performance by reducing the data dimensionality.

Table 1. Four combinations of the experiments.

		Feature Selection
		Random Subset	Random Projection
ML Model	Naïve Bayes	Combination 1	Combination 2
	Bayesian Network	Combination 3	Combination 4

presents the four combinations of feature selection techniques and machine learning classifiers.

In each combination of the experiments, ten rounds were performed. The first round used the entire dataset. In each subsequent round, the dataset is reduced by 10% using the feature selection technique. The last round used only 10% of the original dataset. The run was repeated ten times in each round, and the average performance was recorded. With a total number of 400 runs, the goal was to avoid any biased or chance results. Repeating the runs and taking the average performance provides more confidence in the performance evaluation of the model. In each run, 90% of the data were used for training the classifier and 10% for testing the generated model. In the following figures, 40 average performance values were recorded for the four different combinations of the experiments and are explained in the next section.

4. Evaluation, Results, and Discussion

In the first combination of the experiments (

Table 2. F-measure for Naïve Bayesian with a random subset.

Percentage	F-Measure Naïve Bayesian (Random Subset)
10	0.721
20	0.693
30	0.707
40	0.708
50	0.727
60	0.717
70	0.722
80	0.729
90	0.721
100	0.721

), the random subset technique and the Naïve Bayes classifier were used. Starting with the entire dataset repeated for ten-time runs and taking the average performance, further similar nine rounds have the same scenario by reducing the dataset by 10% in each round, reaching the last round with only 10% of the original dataset. The measurement of accuracy used for evaluating the generated model’s performance is F-measure, calculated in each round.

The F-measure of the first round, called the base round, where the original dataset was used in the experiment, is 0.721. Commonly, it is thought that the F-measure of 100% of the dataset would result in the best performance among all the other rounds of combination 1 experiments. However, this is not the case, at least for combination 1 experiments. With an overall look at the performances of the ten rounds, it is clear that the best F-measure value goes for the third round, where 80% of the dataset was used. The F-measure for this round is 0.729. This result highlights the advantage of the random subset reduction approach.

Interestingly, the performance in this round, where only 80% of the dataset was used, is better than the performance of the base round, where the full dataset was used. Although the difference in the performance between the two rounds is slight, it reduces the processing resources. It makes an interesting remark about the potential of using sampling approaches for big data processing. The F-measure for the rest of the rounds ranges between 0.727 in the sixth round and 0.693 in the ninth round, where the model’s performance is at the bottom. The most interesting outcome of the combination 1 experiments is that three rounds have the same performance, rounds one, two, and ten. It is worth mentioning that the base round that uses the full data set has the same performance as the tenth round, where only 10% of the original dataset is used. When comparing the amount of two sets of data used for these two rounds, a massive reduction in processing resources has been made with the same model performance. Again, this is not always the case, but at least this combination of the experiments proves that the 90% reduction of the dataset did not affect the model’s performance. In other words, sampling the data can add advantages to big data processing by reducing the processing resources and increasing the accuracy of the generated model in some cases. Figure 4 shows the F-measure for Naïve Bayesian with a random subset.

In the second combination (

Table 3. F-measure for Naïve Bayesian with a random projection.

Percentage	F-Measure Naïve Bayesian (Random Projection)
10	0.675
20	0.714
30	0.724
40	0.731
50	0.749
60	0.748
70	0.747
80	0.754
90	0.748
100	0.721

), the random projection technique and the Naïve Bayes classifier were used.

In the first round, where the full dataset was used, the F-measure for the generated model was 0.721. Assuming that the entire dataset would result in the best performance among all rounds since the whole dataset was used, the second round, where 90% of the dataset was used, recorded a better performance at 0.748. Looking at the third round that uses 80% of the dataset, the performance becomes even the best among all the rounds of this combination, where the F-measure has the value of 0.754. This raised a question about the need for using the full dataset for big data analytics, where sampling can provide even better performance of the model and fewer processing resources used. Again, similar to combination 1, the best performance in this combination was recorded for the third round, where 80% of the dataset was used. This is an indication but not a conclusion about the best setting for the Naïve Bayes classifier when using the random subset and projection techniques to reduce the dataset to 80% of the original collection. The performance of the rest of the rounds ranges between 0.749 in the sixth round and 0.675 in the tenth round, where the model’s performance is at the bottom. It is worth mentioning that the worst performance was recorded when only 10% of the dataset was used. This could indicate that reducing the dataset to this amount results in missing some valuable data. While sampling techniques could provide valuable improvements to big data processing, the threshold of reducing the amount of data is another factor that needs to be investigated in detail. While the performance improved in the fifth and sixth rounds, it drops down afterward in each round with an inverse relation to the amount of the dataset reduced. Again, this can prove that it is not necessary to improve the performance by reducing the dataset but to understand the nature of the data and the sampling technique used to reach the optimal reducing amount that provides the best performance for the generated model of the classifier. Thus, it is a multi-faced methodology that the decision-maker must consider when processing big data. Figure 5 shows the F-measure for Naïve Bayesian with a random projection.

In the third combination (

Table 4. F-measure for Bayesian Network with a random subset.

Percentage	F-measure Bayesian Network (Random Subset)
10	0.647
20	0.741
30	0.756
40	0.757
50	0.759
60	0.763
70	0.766
80	0.762
90	0.767
100	0.77

), the random subset technique and the Bayesian network classifier were used. Unlike the previous two combinations, the best performance recorded for this combination was in the first round, where the full dataset was used.

The F-measure for the first round is 0.77, which decreases afterward to 0.767 in the second round, where 90% of the dataset was used, and 0.762 in the third round, where 80% of the dataset was used. The F-measure returned to increase again at 0.766 in the fourth round. However, it kept decreasing slightly afterward until the ninth round, where the performance at that point was 0.741. A sharp drop was recorded for the tenth round at 0.647, where only 10% of the dataset was used. The scenario in this combination has a different behavior than the previous two experiment combinations. The random subset technique has not provided advantages to the classifier’s performance. However, one can conclude that with the slight decrease in the performance of the generated model, it is worth the great deal of the reduction in the dataset amount, the thing that the decision-maker can decide given that the sampling technique would provide an improvement concerning the processing resources used. Figure 6 shows the F-measure for Bayesian Network with a random subset.

In the fourth combination (

Table 5. F-measure for Bayesian Network with a random projection.

Percentage	F-measure Bayesian Network (Random Projection)
10	0.672
20	0.72
30	0.727
40	0.742
50	0.737
60	0.743
70	0.738
80	0.748
90	0.747
100	0.77

), the random projection technique and the Bayesian network classifier were used.

Similar to the third combination, the best performance was recorded in the first round with an F-measure value of 0.77. The performance afterward fluctuated, with slight ups and downs between 0.748 in the third round and 0.72 in the ninth round. The worst performance was recorded in the tenth round at 0.672, where only 10% of the dataset was used. It can be inferred from this combination that the random projection technique did not improve the performance of the generated model. However, looking at the slight decrease in the performance with the dramatic saving on processing resources by reducing the amount of dataset, one can point out that the sampling technique provided an advantage of reducing the resources used. Figure 7 shows the F-measure for Bayesian Network with a random projection.

Figure 8 and Figure 9 show the comparative F-measures for the Naïve Bayesian and the Bayesian Network for random subset and the random projection approaches. Figure 8 shows the random subset F-measure distribution for the two classifiers. In Figure 8, we can see that the overall model accuracy of the Bayesian network is higher than the Naïve Bayesian. Initially, with 10% data, the Naïve Bayesian has higher accuracy, but as the data increases, the Naïve Bayesian is outperformed by the Bayesian Network. With the data increase per scenario, the Bayesian Network receives approximately a 5% increase in performance. Figure 9 shows the random projection F-measure for the two classifiers. Figure 9 indicates that the Bayesian network in many rounds still outperforms the Naïve Bayesian, but the influence is not as dominant as that of Figure 8.

Figure 10 summarizes the overall statistics in one graph. With the overall view of Figure 10, the best performance among all combinations is recorded for combination 3, where the random subset technique and the Bayesian network classifier were used. Except for the round where 10% of the dataset was used, combination 1 has the best performance among all combinations.

5. Comparative Analysis

For comparison, we include experiments covering different approaches that have similar feature/data extraction capabilities. For the comparison, we fix (select) the proposed Bayesian with Random Projection, represented as BN-RP. By selection here, we mean that since there are many permutations in this article, we select one good performance setting for our evaluation. We fix the 50% data as the baseline for all the feature/data selection approaches. For the comparison, we use reservoir sampling [8], Pure random sampling, and Subset-Eva. [41], Correlation Eval. [63], Gain-R [19], Info-Gain [20], OneR [43], PCA [44], Relief [45], and Symmetrical-Uncertain-Evaluation [64]. These are among the widely used features/data selection approaches in the state of the art. We have used similar settings for all the feature/data extraction approaches. These are among the most widely used features/data selection approaches in the state of the art. We have used similar settings for all the feature/data extraction approaches.

Table 6. Comparison of the proposed approach with feature/data selection approaches.

Proposed Approach (BN-RP)	0.759
Reservoir sampling	0.751222222
Pure random sampling	0.749922222
Subset Evaluation	0.7498
Correlation Evaluation	0.7392
Gain Ratio Evaluation	0.735
Info Gain Evaluation	0.751
OneR Evaluation	0.7387
Principal Components	0.753
Relief Evaluation	0.751
Symmetrical Uncertain. Evaluation	0.73

and Figure 11 show the details comparison of the different approaches based on the F-measure. The proposed approach has the highest F-measure of 0.759. The reservoir sampling achieves an F-measure of 0.751. Pure random sampling achieves an F-measure of 0.749. Pure random sample here refers to an approach that takes the first 50% of the data if arranged in order. The subset evaluation has almost similar performance to that of pure random sampling. The correlation evaluation and gain ratio have almost similar F-measures. The info gain, PCA, and relief evaluations have a good F-measure of over 0.75. The OneR and info gain evaluations reduced the F-measure to 0.73. Thus, the proposed approach outperforms all the other approaches.

6. Conclusions

Big data analytics is still a challenging process in both levels of data processing and computing resources. This study has analyzed a large amount of data with classical machine learning. As such, the main objective of this article was to show that in large datasets, it is quite possible that a random selection of features can be as good as the selection of features by optimization algorithms such as the Pareto-front, ant colony optimization, particle swarm optimization, and many others. Thus, this study has analyzed a large amount and sampled it randomly into subsets. The influence of different techniques on the model performance is analyzed by combining the feature selection techniques and machine learning classifiers. The experiments used two feature selection techniques: random subset and random projection. In the evaluation, it was noted that the overall performance fluctuates between 70% accuracy to 74% accuracy for using sampled and non-sampled (all the data). Thus, in large datasets, it is quite possible that the random selection of the small portion of the data can produce as good results as the original data. The difference in performance is only slightly over 3%, which is negligible. We thus argue that if optimized algorithms are used, the performance will still be the same because the actual performance of 100% data differs very little from the reduced data samples. Thus, we argue that whether or not the reduced samples are selected by optimization algorithms such as the Pareto-front, ant-colony, particle swarm optimization, and many others, their performance will be the same or close because, with sampled versions, we should not be able to outperform the model of the 100% data.