12.3.1 Classification

The key objective of this model is to categorize a dataset into more than one predetermined classes. It can be achieved by implementing the mapping which is derived from a vector of values toward a classified parameter. From this classification, the membership of a data instance can be detected for provided predetermined classes. For example, collection of outlet can save customer information into three classes, namely, maximum spending, moderate spending, as well as minimum spending users. The classifying operation is applied for several other domains like economic sectors, bioinformatics, classification of document, multimedia, textual computation, social networks, and so on.

12.3.2 Regression

This method is considered as forecasting model which connect the dataset to quantitative features as well as computes the variable measure. Regression has various applications like using the likelihood where a patient acquires illness on the basis of analysis report and detection of spontaneous team according to the technical information of existing matches. It is generally applied in financial, ecological studies, market tendency, meteorology, as well as epidemiology.

12.3.3 Clustering

The DM operation is applied to find a group of clusters to define the data. Mostly, clustering techniques have been applied in the absence of class which has to be forecasted in an available input data and classified into set of identical instances. There are few applications of clustering which is applicable to find the similar subsets of customers present in database of business groups of objects including the same attributes. It is utilized for several applications like gene computing, intrusion of network, clinical imaging, analyzing the violation, and text mining. In contrary, the classification has predefined classes that should be obtained from data, and seeking for cluster which depends upon the measures of homogeneous data with no help of clients.

12.3.4 Summarization

It is one among the DM process which offers an accurate definition of subset of data. The summarization models are often applied to unstructured data such as classifying text and combine the document by distributing identical behavior. It also uses several tedious functions that have summary criteria as well as identifying the functional associations among variables. Hence, it is mostly employed in the communicative analysis of data and automated production of reports.

12.3.5 Dependency Modeling

It is employed for identifying the methods which explains the important base among variables. This method mainly concentrates in finding the dependencies of each data values. The dependency models can be divided into two phases. Initially, structural level of a method that is based on local variables, whereas quantitative level denotes the efficiency of dependency in a numeric scale. These modeling techniques are applied in retailing, commercial purpose, process management, software designing, and assembly line optimizing process.

12.3.6 Association Rule Discovery

This technique aims in finding the set of items that has been collected from records of a database as well as the correlation between this information is more useful to obtain maximum association which satisfy particular thresholds. It is mainly employed to discover robust patterns from huge datasets under the application of diverse values. Various domains apply this model to understand the intersection massive dataset. At this point, the association rule is named as market basket analysis. Also, there is alternate mandatory application from where the association discovery has been employed such that accessing web page, interacting network application, credit card facilities, medical analysis, bioinformatics, and so on.

12.3.7 Outlier Detection

Here, the outlier is said to be an observation which is more inconsistent by deviating from alternate items of dataset from where it has been originated. Even though the outlier is assumed to be fault or error, it carries more essential data. These models are presented for various domains like environmental analysis, credit card malfunctioning, medicinal trials, network intrusion detection, system process investigation, as well as alternate data analyzing issues.

12.3.8 Prediction

This operation seeks for association among temporary series of actions. The data instances looks like sequences of actions, from where all functions are composed with time of event. There is major problem in detecting the event sequences which is to discover the sequential segments, and set of events that happens frequently. For instance, identifying that a predefined event E1 happens, within a time interval t, also event E2 with maximum probability p. Instances of applications to apply for all alarm detection used in telecommunications, web log designing, client action series, error detection while developing plants, and environmental conditions.

12.4.1 GAs for Feature Selection

In [4], the authors presented a integrative co-evolutionary technique for FS-based GA including three populations, where it mainly aims on FS process and on instance selection, as well as to concentrate on FS and instance selection. Here, the introduced model reports the FS and instance selection in a unique performance that minimizes the processing duration. These types of models undergo further investigations for the provided datasets with thousands of features that contributes unwanted features and noisy data. In addition, they established massive populations dependent GA for the purpose of FS technique, which is comprised with two adjacent populations distributed for interchanging data to improvise the searching potential. Here, local search has been computed on optimal individual for all population that tends to enhance the computation. This presented technique undergo testing with diverse filter as well as wrapper values, that showcased a powerful FS process, also sampled using datasets along with more number of features.

The authors in [5] projected a model to report the FS-based issue by using GAs for feature clustering, and GA has been applied to make optimized cluster middle values in clustering model which is essential to cluster the features as diverse clusters. Features present in every cluster undergoes ranking on the basis of distance measures to the intermediate cluster. The FS process could be attained by selecting top most features as a representative from every cluster. The deployed model shows an effective datasets including more amounts of features. They established a GA-relied FS technique by considering the domain knowledge about economical distress forecasting so that features are divided as various groups as well as GA have been employed for feature subsets that contains best candidate features from every cluster. The above process might consist of identical issue with respect to avoid feature communications. Subsequently, GA is applied in two-stage method, which uses a filter value for ranking the features and selects a first ranked feature which is then employed for GA based FS.

Bio-encoding method in GA which comprises of every chromosome along with pair of strings is applies in [6]. The primary string is a binary-encoded, which denotes the FS process while secondary string is encoded in the form of real-numbers by implying the feature weight. By integrating with Adaboost learning technique, the bio-encoding approach reaches a best performance when compared with conventional binary encoding. A novel representation, which includes FS as well as parameter optimization of assertive classifying method like SVM, was used by the authors. The length is defined as total count of features as well as parameters. They designed a three-level representation in GA and Multilayer Perceptron (MLP) which is useful in FS operation that signifies the election of features, also pruning of neurons, and the structure of MLP. The above mentioned instances recommend that combination of FS as well as optimizing the classification technique is very efficient to enhance the classifying operation as data and classification models has to be optimized.

12.4.2 GP for Feature Selection

Several works have been applied genetic programming (GP) for identifying optimized feature subset and concurrently equipped in the form of classifier. In this, a wrapper FS method which depends upon the multi-tree GP that selects optimal feature subset at the same time as well as a trained classification under the application of chosen features was modeled. Additionally, two novel crossover processes are established to improve the GP performance in selecting features. According to the two crossover operations developed in [7], it deployed an alternate crossover tool that operates in random selection of sub-tree from the primary parent and to search for effective location from alternate parent. It displayed a successful capability of GP to perform as simultaneous FS and in training a classifier. It has the same work that applies GAs in concurrent FS and to produce optimized classification process; however, it has the major variations of GP that applied for searching process to select features and in the form of classifier for classification technique (embedded FS) as GA is employed only for exploring purpose.

The authors also presented a mutual data metric technique for ranking single features as well as to eliminate the vulnerable unwanted features in a primary stage, where GP have been used for selecting a subset of residual features. In order to acquire various merits of diverse values, various filter values has been employed in ranking features and group of features have been selected on the basis of individual metric. By concatenating the above features, it might be applied as input for GP to improve the FS operation. Unfortunately, there is a potential shortcoming is that to eliminate the essential features with no consideration of communication with alternate features. In [7], the authors developed alternate type of single feature ranging technology, that consist of frequency embedded feature which has a unique score and ranked accordingly. Consequently, FS has been accomplished by applying most top ranged features which is induced for classification. It is the path of estimating the unique features to consider the adjacent ones that is capable of removing the shortcoming of individuals to rank features, respectively.

A GP relied multi-objective filter FS model has been presented here for binary classification issues. In contrast, many of the filter approaches can estimate the relation of individual feature of class labels, and the presented technique is capable of finding the correlation among hidden association from feature subset as well as the target classes, also to attain effective classification process. GP consists of minimum number of process in selecting features for multi-objective condition. It is more attractive to examine the upcoming process as GP shows a potential reporting of FS as well as multi-objective issues.

12.4.3 PSO for Feature Selection

Here, continuous PSO as well as binary PSO have been applied for filter and wrapper techniques with unique objective and multi-objective FS. The presentation of every particle within PSO has been applied in selecting features which is a typical operation where the dimension is same as total count of features present in the dataset. A bit-string could be a binary value where binary PSO in continuous PSO. While applying the binary representation, “1” refers the FS, whereas “0” symbolizes no selection. In case of employing the continuous representation, a threshold θ can be applied to compute the selection of specified feature, which has the value greater than θ, and then, adjacent features are selected. Else non-selected features are available. In [8], the authors presented the application of PSO and statistical clustering that helps to cluster the identical features into similar cluster for FS that has novel representation to combine the statistical feature clustering data at the time searching operation is carried out in PSO. The novel representation reveals that the features derived from similar cluster that has been organized together as well as individual feature have been chosen from all clusters. The projected technique shows the potential of minimizing the count of features. The technique was enhanced by enabling the selection process for many features from the identical cluster which helps to improvise the classifying operation.

By acquiring the knowledge from neighbors’ experience, by interacting under the application of gbest, as well as by learning from every individual’s self-experience by pbest are assumed to be the main suggestions of PSO. In [8], the authors deployed a gbest resetting model by adding zero features to suggest the swarm for searching tiny feature subsets. They assumed the count of features while upgrading pbest as well as gbest at the time of exploring PSO, that tends to decrease the amount of features than the conventional update of pbest and gbest techniques with no alleviation in classifying operation. All estimation present in local search has been improved by measuring the fitness relied on features which is often modified from required to non-required and vice versa. Therefore, the presented model again reduces the feature count and enhances the classifying performance than existing models as well as a reputed PSO. Any PSO with more number of swarms helps to share the experience which can be used in FS operation but tends in greater computation expense.

12.4.4 ACO for Feature Selection

In previous models, ACO as well as SVM have been employed for wrapper FS that is useful in face recognition, which consist of actual features that are obtained by applying the strategy of PCA from the images which is present in pre-processing phase. In [8], the authors projected the application of restricted pheromone measures in ACO that is applied for FS process as well as the deployed technique has been upgraded with pheromone trails of edges by linking diverse features of best-so-far solution. The final outcome displayed that the newly deployed technique attains good performance when compared with SA, a GA, and TS dependent models with respect to classification process and amount of features. Techniques applied in [8] use a ACO to select features concurrently as well as to parameter optimization of SVM, that consist of weighting model used in computing the probability of ant choosing specific feature. They also integrated ACO with DE for FS, and DE has been applied in searching the optimized feature subset on the basis of solutions derived by ACO. A classical FS technique has been presented for ACO, so that ACO starts with minimal number of core features. In [9], the authors referred an applicable ACO model including two colonies for FS task, where the initial feature allocated the feature value which is more essential and secondary colony is useful to choose unique features. They related the process of ACO along with GA-relied FS model that initiates ensemble classification. Hence, the simulation outcome depicts that ACO performs better in case of single classifiers are tiny at the time of GA performing optimal process while having maximum values.

Several ACO-based wrapper techniques are applied as fitness evaluation criteria and the fitness of ants have been measured with the help of entire classification process, as the operation of single features consider the further enhancing operation. Also, the fitness functions have been added for classifying process as well as count of features. Then, by expanding the process on individual objective ACO as well as a fuzzy classification that is applied for FS process, they deployed a multiobjective wrapper technique where the ACO focuses in reducing classification error as well as count of features. In [9], the authors introduced a novel multi-objective ACO which can be applied for filter FS operation that obtains elitism principle to improve the converging process that applied non-dominated solutions to include pheromone which results in reinforcing the exploitation, as well as to be used in a crowding relational operator to balance the diversity of solutions. The derived final outcome shows a better performance when compared with individual objective methodologies, since it is more attractive to analyze the application of multi-objective ACO in FS process.

12.5.1 Neural Predictive Classifier

The neural predictive classifier exploits the human neuron on a system. Billions of neurons are interconnected in human brain for conveying the information’s with other layer using electric and chemical signals. A multilayer network called neural predictive classifier which is made up of input layer neurons, hidden neuron, and output neurons.

Back propagation neural network is a supervised learning that is used to classify the medical data conditions on big data. The main aim of using neural predictive classifier is to reduce error rate during the classification. Back propagation is performed with the aid of three different layers. The construction of the neural network for classification of medical data on big data is shown in Figure 12.4.

Figure 12.5 demonstrates the processing diagram of activation function. The hidden layer is a processed layer that includes two functions, i.e., sum function and activation function. The activation function utilized sigmoid function since it unites nearly linear performance, curved behavior and nearly constant behaviors based on the input value. A sigmoid function is a statistical function that contains an “S” shaped curve as a sigmoid curve. A sigmoid function lies within the finite ranges at negative infinity and positive infinity, while the output can only be a number between −1 and 1. The sigmoid function is mathematically formulated as follows:

(12.3)

From Equation (12.3), sigmoid function σ (x) is formulated. In hidden layer, product of the input similarity data from the input layer with the weight between input layer and hidden layer is summed with the aid of SUM function. The SUM is set of the output nodes from hidden layer which multiplied with the weights consigned to the each input value to get a single number and process through sigmoid function (activation function). Subsequently, a new weight among the hidden layer and the output layer transmitted from hidden layer output to the output layer. The output layer is the processing unit that contains summation function and activation function. In output layer, the multiplication of the activated output of hidden layer and the synaptic weight among the hidden layer and output layer is summed up with the support summation function. Then, it is transmitted to the activation function for choosing final output using threshold value of the activation function.

The resultant output of neural network is compared with the target for computing the error rate which is expressed as follows:

(12.4)

The measured error value then passed backward to the input from the output by means of updating the synaptic weight. This process is repeatedly performed until the error is reduced during the classification of climate data on cloud big data. Thus, the learning rate and momentum rate are initiated for gradient fall to discover the weight that lessens the error. The error rate between the output layer and hidden layer are determined by using following mathematical formula:

(12.5)

(12.6)

From Equations (12.5) and (12.6), error rate is determined and then compared with target value, whereas D_o is represents desired output neuron value and A_v signifies the activation value using sigmoid function. Here, A_v (1 − Av) indicates the derivative of the sigmoid function in hidden layer. Therefore, the synaptic weight is computed is represented as follows:

(12.7)

From Equation (12.7), ΔSWt_y,z demotes the updated synaptic weight between the output layer and the hidden layer, whereas η represents the learning factor that characterizes the relative variations in synaptic weights and input (X). As a result, the updated synaptic weight is described as the error multiplied with the input. Thus, the new weight is characterized as follows:

(12.8)

The above waiting function is utilized for computing the weight value. Hence, the new weight is formulated by using the following:

(12.9)

From Equation (12.9), the new weight is formulated depends on original weight is assigned with updated synaptic weight. Therefore, the SCM-NPC technique used neural predictive classifier to improve the classification accuracy of climate data on big data. The number of neurons (i.e., attributes) in the hidden layer is to find out the optimal number of neurons which supports to provide the better classification of climate data on cloud big data. There are two functions exploited in the hidden neurons, i.e., summation of the synaptic weight and the activation function. The activation function utilized in neural predictive classifier as sigmoid transfer function to handle complex problems during the classification of climate data on cloud big data. At last, the output layer generates the target output value for predicting climate data information on cloud big data.

The algorithmic process of neural predictive classifier for classification of medical data on cloud big data is shown in the following.

As shown in Algorithm 12.1, at first, the neural network is created with random weights. For similarity data obtained from Semi-Convergent Matrix, the error rate is determined for updating the new weight in back propagation of classifier in order to get the activation output. The output of the activation is concluded, and the climate data is present in cloud big data. In SCM-NPC technique, the sigmoid function is used as an activation function. This produces the required output. If the activation produces the output as −1, then the data is classified as the Adverse Drug Reaction (ADR). Otherwise, the data is no ADR. The neural predictive classifier algorithm classifies the climate data from cloud big data through the activation function which resulting in improved classification accuracy.

12.5.2 MapReduce Function on Neural Class

The SCM-NPC technique employs MapReduce function on neural class to perform prediction analytics of ADR. The MapReduce function in SCM-NPC technique process huge volumes of data in a parallel fashion. The SCM-NPC technique employs map-reduce function on neural class to perform predictive analytics of ADC as in Figure 12.6. In map reduce function, the input is taken by map task and revamps the content into Key-Value pair where the key has class, attributes, and value acquired from neural class which mathematically represented as follows:

(12.10)

(12.11)