Design of metaheuristic rough set-based feature selection and rule-based medical data classification model on MapReduce framework

3.1 Preprocessing

At the initial stage, the data preprocessing takes place to transform the data into a compatible format. Primarily, the pre-processed step was implemented to change non-traditional into traditional datasets that improve the accuracy of presented approach. Here, the min–max normalization manner was applied. Among the most prevalent methods for normalizing information is min–max normalization. For each attribute, the significance level is converted to a 0, the highest value is converted to a 1, and all other values are converted to a fraction within 0 and 1. In this approach, the feature is being rescaled to the range of 0 to 1 or –1 to 1. This rescaling is done by the utilization of linear interpretation equation defined as follows:

where ( y max − y min ) = 0 ; if ( x max − x min ) = 0 to a feature, it defines the continuous rate of feature from the data. Once the value of feature has been identified with constant value from the data, it needs the normalized value as it does not deliver some data to NN. Once the min–max normalization was implemented, all the features lie from the novel range of values that in fit remain similar. The normalization utilizing min–max is an advantage of maintaining every connection from the data correctly. The min–max normalization approach (y = (x – min)) is utilized; however, there are additional choices. The original picture information will be changed in the range of 0 to 1 by using min–max normalization (inclusive).

3.2 Design of BOA-based minimal rough set selection

Once the medical data are preprocessed, the feature selection process is carried out by the use of BOA-based minimal rough set selection. The BOA has been metaheuristic technique, which is simulated as foraging and mating nature of BFs. An essential feature of BOA from other metaheuristics is that all BFs hold their individual scent. The fragrance is represented using the following equation:

where f i denotes the apparent order of fragrance, c implies the sensory modality, I is the stimulus intensity, and a signifies the power exponent that depends on degree of fragrance absorption. In practice, the values of sensory morphology coefficient c from the interval of [ 0 , ∞ ] are assumed. Then, the value is decided utilizing the discrimination of optimized problem from the iterated system of BOA [19]. The sensory modality c from the optimal searching stage of the BOA is defined using the following equation:

where T max denotes higher iteration count and the beginning value of c is fixed as 0.01. Also, two major processes are involved in the BOA, namely global search and local search stages. The global searching movement of the BFs can be defined using the following equation:

where x i t signifies the solution vector x i of the i th BF in t round and r denotes an arbitrary number from the range of [0, 1]. In this case, g best denotes the present optimum solution attained from every solution attained in the present stage. Specifically, f i signifies the fragrance of the i th BF. The local searching process can be defined using the following equation:

where x j t and x i k are j th and k th BFs arbitrarily selected from the solution area. When x j t and x i k belong to the identical round, they denote that the BF will become a local random walk. Else, this type of arbitrary movement diversified the solutions. The global as well as local searching process for food and mating partners by the BF takes place naturally. Therefore, a switching probability p has been considered for transforming the common global and intensive local searching process. At all rounds, the BOA arbitrarily makes a number from the range of [0, 1] that has undergone comparison with the switch probability p in deciding to perform a global or local searching process [21].

The concept of BOA can be used for the minimal rough set selection issue. Assume a huge feature space with entire feature subsets. Every feature subset can be considered as a point or location in the space. When a total of N features exist, a set of 2 N types of subsets occur varying from one another in the length and features exist in every subset. The optimum location is the subset with minimal length and maximum classification accuracy. There were multiple BFs in the feature space, and each BF was assigned a position. If the BF searches the food in the space, the goal is to move toward the optimum location. Later, the position gets changed, they communicate with one other and search for the local as well as global optimum position. The process involved in the BOA-based rough set selection process is detailed in the following.

Initially, the BF position is considered as the binary bit string of length N , where N denotes the feature count. Every bit denotes a feature, and particularly, the value “1” implies that the respective feature is chosen whereas “0” denotes unselected. Every position implies a feature subset. Consider X , Y as the binary bit strings which denote the position of two BFs.

The Hamming distance is a statistic that may be used to compare two binary value sequences. When two discrete sequences of similar duration are compared, the Hamming distance is the amount of bit locations where the data pairs disagree. d is indicated as hamming distance and it can be illustrated as d(a,b). For every binary bit string X and Y , the Hamming distance is equivalent to the number of 1’s in X XOR Y . Assume X , Y are the two binary bit strings denoting the position of two BFs, and the Hamming distance can be equated using the following equation:

where ⊕ denotes a modulo‐2 addition, x i , y i ∈ { 0 , 1 } . The parameter x i signifies a binary bit in X . Consider X 1 , X 2 , … , X n are the binary bit strings representing the position of n BFs. The midpoint position of n BFs can be formulated as follows:

where x i j characterizes the i th bit of BF position X j , and X c is the middle position of n BFs.

Every BF begins with an arbitrary position in each round [22]. Every BF aims to modify each step in the searching space based on the behavior of searching, swarming, and following. A fitness function can be derived based on the three behaviors and the one with higher fitness value can be chosen for updating the succeeding position. It can be represented as follows:

where γ R ( D ) denotes the classification quality (also called dependency), | R | indicates the total number of “ 1 ” numbers in a binary BF position, indicating the chosen feature count, also termed as the feature subset length. | C | denotes the total feature count. α and β are two variables implying the significance of classifier quality and length of the subsets, α ∈ [ 0 , 1 ] , and β = 1 − α . It indicates that the classification quality and feature length might hold distinct significance to the feature selection process. The fitness function (sometimes referred to as the analytical solution) determines how near a particular approach is to the optimization method of a particular topic. It assesses how appropriate a proposal is. The importance of every position is determined using the fitness function.

Once the BF reached a maximum fitness value, it is perished with getting a rough set select. It indicates that the BF constructs the local optimum solution. The succeeding round beings when every BF gets perished. The termination condition is set as the maximum number of iterations or attaining an identical set of feature selects under three succeeding rounds. [23]

3.3 Design of RIPPER in MapReduce environment

During the classification process, the RIPPER technique gets executed on the MapReduce platform to classify the healthcare data. MapReduce performs two important capabilities as it classifies and distributes activity to different devices in the system or mapping, a service known as the mapper, and it organizes and combines the data for every server into a coherent response to a question, known as the reduction. The RIPPER has been extremely utilized as a rule induction technique. It scales linearly with the training sample count utilized and is suitable to structure techniques with data uncertainty. Also, it utilizes a validation set (VS) for preventing the technique from overfitting. The RIPPER orders the classes based on the frequency. When ( y 1 , y 2 , … , y c ) denotes the class label and y 1 refers to the minimum frequency and y c the maximum frequency, afterward, RIPPER primarily forms rules to y 1 taking residual class record as negative record. Then, RIPPER removes principles to y 2 . This procedure has been repeated till y c is left, which is considered as default class.

In order to generate rules, RIPPER utilizes an approach which primarily considers all rules are empty and afterward it can be constructed with more conjuncts to it consecutively. It utilizes FOIL IG for adding conjunct to rule. Assume that rule R : A → class cover p 0 positive record and n 0 negative record. Then, a more novel conjunct B , the rule R ′ : A ∧ B → class cover p 1 positive record and n 1 negative record. Next, the FOIL IG is computed as:

The conjunct is increasing till the rule begins covering negative instances. The rule was pruned dependent upon their efficiency on the VS utilizing the subsequent metric ( p − n ) / ( p + n ) , where p refers to the amount of positive record concealed with rule from the VS and n implies the amount of negative record supporting the rule under the VS.

After creating the rule, every record under the rule is removed. The minimum description length (MDL) concept is a strong inductive assessment approach that serves as the foundation for data analysis, analytical thinking, and computer vision. It asserts that the optimal interpretation, provided a restricted collection of observable facts, is the one which allows for the most quantization. This technique then continues with constructing a novel rule. The rule has been made if the rule set does not violate the MDL rule and the error on the VS was lesser than 50%.

RIPPER can be applied in Java using Hadoop Java library. The dataset has been separated horizontally for supporting the Hadoop MapReduce structure and making sure parallel implementation of code. The mapper–reducer is used for three purposes: one in one to rule generating, rule pruning, and computing accuracy [20]. Therefore, all the mappers implement their code on some of the datasets, and the reducer aggregates on the outcome of mapper for producing one general output. Figure 2 illustrates the framework of MapReduce.

Figure 2

Structure of MapReduce.

In order to rule generating, the mapper–reducer function computes the values of p 1 and n 1 to compute the FOIL’s IG ( p 0 and n 0 values are p 1 and n 1 values to the old rule, respectively). Altogether, all feasible values to every attribute are regarded as conjuncts that rule more. The FOIL’s IG to all of these values are computed, and the value to that IG is maximal, that is, more conjunct added to rule. The rule pruning was complete utilizing the VS as situation. The mapper–reducer function to pruning computes the p and n values to metric ( p − n ) / ( p + n ) . According to the value of metric, the rule was pruned along with the rule set. Afterward, every rule has made the rule set require that validation on test record. The accuracy mapper–reducer function computes the amount of positive as well as negative records supported by all the rules and entire rule set. These values have been utilized for calculating the accuracy of all rules and entire accuracy as well.

3.4 Rule growing phase

The rule has been adjusted as an empty rule, for instance, it covers every record. Then, the conjunct is added further one by one to the rule. The conjuncts with the value of FOIL’s IG measure have been chosen more. The parameter of measure was computed utilizing MapReduce functions, and the <key, value> pairs have the values of p 0 , p 1 , n 0 , and n 1 . The conjunct is further added to rule if it does not cover negative record.

3.5 Rule pruning phase

The rule created in one is before pruned utilizing ( p − n ) / ( p + n ) metric. For calculating the parameters p and n of this metric, the MapReduce model was named as <key, value> pairs. . Stages 1 and 2 are repeated till increasing a novel rule violates the MDL rule.

3.6 Model evaluation phase

Then the rule set is created, and the rules have been utilized for classifying the test record. The MapReduce purpose is named for classifying the record and computing the accuracy of this technique. The <key, value> pairs comprise the value of entire positive and negative records supporting the method. The rule set is returned, and the correctness of the rules and the rule set on the test record are assessed.