Rider weed deep residual network-based incremental model for text classification using multidimensional features and MapReduce

Stop word removal

This is a process that removes words with less representative value for the data. Some of the non-representative words include pronouns and articles. When evaluating data, some words are not valuable to text content, and removing such redundant words is imperative. This procedure is termed stop word removal (Dave & Jaswal, 2015). Certain words such as articles, conjunctions, and prepositions, continuously appear and are called stop words. The removal of the stop word, the most imperative technique, is utilized to remove redundant words using vocabulary because the vector space size does not offer any meaning. The stop words indicate the word, which does not hold any data. It is a process used to eliminate stop words from a large set of reviews. The elimination of the stop word is used to save space and accelerate and improve processing.

Stemming

The stemming procedure is utilized to convert words to stem form. In massive amounts of data, several words are utilized that convey a similar meaning. Therefore, the critical method used to minimize words to root is stemming. Stemming is a method of linguistic normalization wherein little words are reduced. Moreover, it is the procedure used to retrieve information for describing the mechanism of reducing redundant words to their root form and word stem. For instance, the words connections, connection, connecting, and connected are all reduced to connect (Dave & Jaswal, 2015). (5)${\displaystyle M_l=\left\{Q_i,1\le \eta \le P_k\right\}}$

where Q_i symbolizes total words present in text data from the database.

The pre-processed outcome generated from pre-processing is expressed as M_l, which is subjected as an input to feature extraction phase.

Extraction of SentiWordNet features

The SentiWordNet features are utilized from pre-processed partitioned data by removing keywords from reviews. Here, the SentiWordNet (Ghosh & Kar, 2013) is employed as a lexical resource to extract the SentiWordNet features. The SentiWordNet assigns each WordNet text one of three numerical sentiment scores: positive, negative, or neutral. Here, different words indicated different polarities that indicated various word senses. The SentiWordNet consisted of different linguistic features: verbs, adverbs, adjectives, and n-gram features. SentiWordNet is a process used to evaluate the score of a specific word using text data. Here, the SentiWordNet was employed to determine the polarity of the offered review and for discovering positivity and negativity. Hence, SentiWordNet is modeled as F₁.

Extraction of contextual features

The context-based features (Ranjan & Prasad, 2018) were generated from pre-processed partitioned data that described relevant words by dividing them using non-relevant reviews for effective classification. This requires finding key terms that have context and semantic meaning in order to establish a proper context. The key term is considered a preliminary indicator for relevant review while context terms act as a validator that can be used to evaluate if the key term is an indicator. Here, the training dataset contained keywords with pertinent words. The context-based features assisted in selecting the relevant and non-relevant reviews.

We considered representing a training dataset that had relevant and non-relevant reviews. Using this method, assume x_s represents the key term and x_c indicates the context term.

- Detection of key terms:

The language model that employs each term and the metric are expressed as: (6)${\displaystyle C=\frac{L_{\mathrm{rel}}}{L_{\mathrm{on}\text{\_}\mathrm{rel}}}}$

where, L_rel symbolizes the language model for N_rel and L_non-rel signifies the language model for N_{non_rel}.

- Discovery of the context term:

After discovering key terms, the process of context term discovery, which is similar to separately detecting each term, begins. The steps employed in determining the context term are given as:

(i) Computing all instances of the key term employed among relevant and non-relevant reviews.

(ii) By employing sliding window size, the conditions are mined as context terms. Hence, the size of the window is employed as a context span.

(iii) The pertinent terms generated are employed as a text, modeled as d_r, and non-relevant terms are denoted as d_nr. The set of pertinent text is modeled as R_{d_r}, and the non-relevant set is referred to as R_{d_nr}.

(iv) After that, the score evaluated for each distinctive term is expressed as: (7)${\displaystyle C\left(x_C\right)=|L_{R_{d\text{\_}r}}\left(x_C\right)-L_{R_{d\text{\_}nr}}\left(x_C\right){|}_S}$

where L_{Rd_r}(x_C) symbolizes the language model for an excerpt with relevant review set. The term L_{Rd_nr}(x_C) indicates a language model for an excerpt with a non-relevant review set and S represents the size of the window. If the measure is a definite threshold then that score is adapted as a context term x_S. Generated context-based features are modeled as F₂.

Extraction of thematic features

The pre-processed partitioned data d_r,l isused to find thematic features. Here, the count of the thematic word (Tas & Kiyani, 2007) in a sentence is imperative as the frequently occurring words are most likely connected to the topic in the data. Thematic words are words that grab key topics defined in a provided document. In thematic features, the top 10 most frequent words are employed as thematic. Thus, the thematic feature F₃ is modeled as: (8)${\displaystyle F_1=\frac{T}{\sum _{t=0}^{\ell }{T}_t}}$

where T expresses the count of thematic words in a sentence, and it is expressed as $\left\{T_1,T_2,.\dots ,T_{\ell }\right\}$.

The feature vectors considering the contextual, thematic, and SentiWordNet features are expressed as: (9)${\displaystyle F=\left\{F_1,F_2,F_3\right\}}$

where F₁ symbolizes SentiWordNet features, F₂ signifies contextual features, and F₃ refers to thematic features.

Architecture of the deep residual network

We employed a deep residual network (Chen et al., 2019) in order to make a productive decision regarding which text classification to perform. The DRN is comprised of different layers: residual blocks, convolutional (Conv) layers, linear classifier, and average pooling layers. Figure 2 shows the structural design of a deep residual network with residual blocks, Conv layers, linear classifier, and average pooling layers for text classification.

-Conv layer:

The two-dimensional Conv layer reduces free attributes in training and offers reimbursement for allocating weight. The cover layer processes the input image with the filter sequence known as the kernel using a local connection. The cover layer utilizes a mathematical process for sliding the filter with the input matrix and computes the dot product of the kernel. The evaluation process of the Conv layer is represented as: (11)${\displaystyle B2d\left(O\right)=\sum _{a=0}^{E-1}\sum _{s=0}^{E-1}{X}_{a,s}\cdot O_{\left(u+a\right),\left(v+s\right)}}$ (12)${\displaystyle B1d\left(O\right)=\sum _{Z=0}^{C_{\mathrm{in}}-1}{G}_Z\ast O}$

where O expresses the CNN feature of the input image, u and v refer to the recording coordinates, G signifies the E × E kernel matrix termed as a learnable parameter, and a and s are the position indices of the kernel matrix. Hence, G_Z expresses the size of the kernel for the Zth input neuron and ∗ expresses the cross-correlation operator.

Pooling layer: This layer is associated with the Conv layer and is especially utilized to reduce the feature map’s spatial size. The average pooling is selected as a function of each slice and the depth of the feature map. (13)${\displaystyle a_{\mathrm{out}}=\frac{a_{\mathrm{in}}-Z_a}{\lambda }+1}$ (14)${\displaystyle s_{\mathrm{out}}=\frac{s_{\mathrm{in}}-Z_s}{\lambda }+1}$

where a_in symbolizes the input matrix width, s_in signifies the height of the input matrix, a_out and s_out represent the respective value of output, and Z_a and Z_s symbolize the width and height of the kernel size.

-Activation function: The nonlinear activation function is adapted for learning nonlinear and complicated features so it is utilized to improve the non-linearity of extracted features. Rectified linear unit (ReLU) is utilized for processing data. The ReLU function is formulated as: (15)${\displaystyle \mathrm{ReLU}\left(O\right)=\left\{\begin{array}{c}0;K<0\hfill \\ K;K\ge 0\hfill \end{array}\right.}$

where K symbolizes a feature.

-Batch normalization: Here, the training set was divided into various small sets known as mini-batches to train the model. It attains a balance between evaluation and convergence complexity. The input layers are normalized by scaling activations to maximize reliability and training speed.

-Residual blocks: This indicates the shortcut connection amongst the Conv layers. The input is unswervingly allocated to output only if input and output are of equal size. (16)${\displaystyle \kappa =\Re \left(O\right)+O}$ (17)${\displaystyle \kappa =\Re \left(O\right)+{ƛ}_{M}O}$

where O and κ signify input and output residual blocks, κ symbolizes mapping relation, ƛ_M expresses dimension matching factor, and $\Re \left(.\right)$ signifies activation function.

-Linear classifier: After completion of the Conv layer, linear classifier performs a procedure to discover noisy pixels using input features. It is a combination of the softmax function and a fully connected layer. (18)${\displaystyle \kappa =ƛ\kappa +\upsilon }$

where ƛ expresses weight matrix and υ represents bias. Figure 2 shows the structural design of the deep residual network. Here, the output, represented as κ, assists in classifying the texts.

Training of the deep residual network with the Adam algorithm

The deep residual network training is performed using the Adam technique which assists in discovering the best weights for tuning the deep residual network for classifying text. Adam (Kingma & Ba, 2014) represents a first-order stochastic gradient-based optimization extensively adapted to a fitness function that changes for attributes. The major implication of the method is computational efficiency and fewer memory needs. Moreover, the problems associated with the non-stationary objectives and the subsistence of noisy gradients are handled effectively. In this study, the magnitudes of the updated parameters were invariant in contrast to the rescaling of gradient, and step size was handled with a hyperparameter that worked with sparse gradients. In addition, Adam is effective in performing step size annealing. The classification of text employs a deep residual network for texts. The steps of Adam are given as:

Step 1: Initialization

The first step represents bias correction initialization wherein ${\hat{q}}_l$ signifies the corrected bias of the first moment estimate and ${\hat{m}}_l$ represents the corrected bias of the second moment estimate.

Step 2: Discovery of error

The bias error is computed to choose the optimum weight for training the deep residual network. Here, the error was termed as an error function that led to an optimal global solution. The function is termed as a minimization function and is expressed as: (19)${\displaystyle \mathrm{Err}=\frac{1}{f}\sum _{l=1}^f{\left(O_l-\kappa \right)}^2}$

where f signifies total data, κ symbolizes output generated with the deep residual network classifier, and O_l indicates the expected value.

Step 3: Discovery of updated bias

Adam is used to improving convergence behavior and optimization. This technique generates smooth variation with effectual computational efficiency and lower memory requirements. As per Adam (Kingma & Ba, 2014), the bias is expressed as: (20)${\displaystyle {\theta }_l={\theta }_{l-1}-\frac{\alpha {\hat{q}}_l}{\sqrt{{\hat{m}}_l+\varepsilon }}}$

where α refers to step size, ${\hat{q}}_l$ expresses corrected bias, ${\hat{m}}_l$ indicates bias-corrected second-moment estimate, ɛ represents the constant, and θ_l−1 signifies the parameter at a prior time instant (l − 1). The corrected bias of the first-order moment is expressed as: (21)${\displaystyle {\hat{q}}_l=\frac{q_l}{\left(1-{\eta }_1^l\right)}}$ (22)${\displaystyle {\hat{q}}_l={\eta }_1{q}_{l-1}+\left(1-{\eta }_1\right)G_l^1.}$

The corrected bias of the second-order moment is represented as: (23)${\displaystyle {\hat{m}}_l=\frac{m_l}{\left(1-{\eta }_2^l\right)}}$ (24)${\displaystyle {\hat{m}}_l={\eta }_2{m}_{l-1}+\left(1-{\eta }_2\right)H_l^2}$ (25)${\displaystyle H_l={\nabla }_{\theta }\mathrm{loss}\left({\theta }_{l-1}\right).}$

Step 4: Determination of the best solution: The best solution is determined with error, and a better solution is employed for classifying text.

Step 5: Termination: The optimum weights are produced repeatedly until the utmost iterations are attained. Table 1 describes the pseudocode of the Adam technique.

Architecture of deep residual network

The model of the deep residual network is described in ‘Classification of Texts with Adam-Based Deep Residual Network’.

Training of deep residual network with proposed RIWO

The training of the deep residual networks was performed with the developed RIWO, which was devised by integrating IWO and ROA. Here, the ROA (Binu & Kariyappa, 2018) was motivated by the behavior of the rider groups, which travel and compete to attain a common target position. In this model, the riders were chosen from the total number of riders for each group. We concluded that this method produces enhanced classification accuracy. Furthermore, the ROA is effective and follows the steps of fictional computing for addressing optimization problems but with less convergence. IWO (Sang, Duan & Li, 2018) is motivated by colonizing characteristics of weed plants. The technique showed a fast convergence rate and elevated the accuracy. Hence, we integrated IWO and ROA to enhance complete algorithmic performance. The steps in the method are expressed as:

Step (1) Initialization of population

The preliminary step is algorithm initialization, which is performed using four-rider groups provided by A and represented as: (26)${\displaystyle A=\left\{A_1,A_2,\dots ,A_{\mu },\dots ,A_{\vartheta }\right\}}$

where A_μ signifies μth rider and ϑ isthe total riders.

Step (2) Determination of error:

The computation of errors is already described in Eq. (19).

Step (3) Update the riders’ position:

The rider position in each set is updated for determining the leader. Thus, the updated rider position using a feature of each rider is defined below. The updated position of each rider is expressed as:

As per ROA (Binu & Kariyappa, 2018), the updated overtaker position is used to increase the rate of success by determining the position of the overtaker and is represented as: (27)${\displaystyle A_{n+1}^o\left(g,h\right)=A_n^o\left(g,h\right)+\left[{\partial }_n^{\ast }\left(g\right)\ast A^L\left(L,h\right)\right]}$

where ${\partial }_n^{\ast }\left(g\right)$ signifies the direction indicator.

The attacker has a propensity to grab the position of the leaders and is given by: (28)${\displaystyle A_{n+1}^a\left(g,u\right)=A^L\left(L,u\right)+\left[\mathrm{Cos}{K}_{g,u}^n\ast A^L\left(L,u\right)\right]+r_g^n.}$

The bypass riders contain a familiar path, and its update is expressed as: (29)${\displaystyle A_{n+1}^b\left(g,u\right)=\lambda \left[A_n\left(\chi ,u\right)\ast \delta \left(u\right)+A_n\left(\xi ,u\right)\ast \left[1-\delta \left(u\right)\right]\right]}$

where λ symbolizes the random number, χ signifies the arbitrary number between 1 and P, ξ denotes an arbitrary number between 1 and P, and δ expresses an arbitrary number between 0 and 1.

The follower poses a propensity to update the position using the leading rider position to attain the target and is given by: (30)${\displaystyle A_{n+1}^F\left(g,h\right)=A^L\left(L,h\right)+\left[\mathrm{Cos}\left(K_{g,h}^n\ast A^L\left(L,h\right)\ast r_g^n\right)\right]}$

where h is the coordinate selector, A^L indicates the leading rider position, L represents the leading rider index, $K_{g,h}^n$ represents the steering angle of gth rider in hth coordinate, and $r_g^n$ is the distance. (31)${\displaystyle A_{n+1}^F\left(g,h\right)=A^L\left(L,h\right)\left[1+\mathrm{Cos}\left(K_{g,h}^n\ast r_g^n\right)\right].}$

The IWO assists in generating the best solutions. Per IWO (Sang, Duan & Li, 2018), the equation is represented as: (32)${\displaystyle A_{n+1}^F\left(g,h\right)=\sigma \left(n\right)A_n^F+A_{\mathrm{best}}-A_n^F}$

where $A_{n+1}^F$ symbolizes the new weed position in iteration n + 1, $A_n^F$ signifies the current weed position, A_best refers to the best weed found in the whole population, and σ(n) represents the current standard deviation. (33)${\displaystyle A_{\mathrm{best}}=A_{n+1}^F\left(g,h\right)-\sigma \left(n\right)A_n^F+A_n^F.}$

Substitute Eq. (33) in Eq. (31), (34)${\displaystyle A_{n+1}^F\left(g,h\right)=\left(A_{n+1}^F\left(g,h\right)-\sigma \left(n\right)A_n^F+A_n^F\right)\left[1+\mathrm{Cos}\left(K_{g,h}^n\right){\ast r}_g^n\right]}$ (35)${\displaystyle A_{n+1}^F\left(g,h\right)=A_{n+1}^F\left(g,h\right)\left[1+\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right]-\left(\sigma \left(n\right)A_n^F+A_n^F\right)\left[1+\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right]}$ (36)${\displaystyle A_{n+1}^F\left(g,h\right)-A_{n+1}^F\left(g,h\right)\left[1+\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right]=\left[A_n^F-\sigma \left(n\right)A_n^F\right]\left[1+\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right]}$ (37)${\displaystyle A_{n+1}^F\left(g,h\right)\left[1-1-\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right]=\left[A_n^F-\sigma \left(n\right)A_n^F\right]\left[1+\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right]}$ (38)${\displaystyle A_{n+1}^F\left(g,h\right)\left[-\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right]=\left[A_n^F-\sigma \left(n\right)A_n^F\right]\left[1+\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right].}$

The final updated equation of the proposed RIWO is expressed as: (39)${\displaystyle A_{n+1}^F\left(g,h\right)=-\frac{\left(A_n^F-A_n^F\sigma \left(n\right)\right)\left(1+\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n\right)}{\mathrm{Cos}\left(K_{g,h}^n\right)\ast r_g^n}.}$

Step (4) Re-evaluation of the error:

After completing the update process, the error of each rider is computed. The position of the rider in the leading position is replaced using the position of the new generated rider so that the error of the new rider is smaller.

Step (5) Update of the rider parameter:

The rider attribute update is imperative to determine an effectual optimal solution using the error.

Step (6) Riding off time:

The steps were iterated repeatedly until we attained off time N_OFF, in which the leader was determined. The pseudocode of the developed RIWO is shown in Table 2.

The output produced from the developed RIWO-based deep residual network is κ, which helps classify the text data since dynamic learning helps classify the dynamic data. Here, we employed fuzzy bounding to remodel the classifier if there was a high chance of a previous data error.

Fuzzy theory

An error is evaluated whenever incremental data is added to the model and weights are updated without using the previous weights. If the error evaluated by the present instance is less than the error of the previous instance then the weights are updated based on the proposed RIWO algorithm. Otherwise, the classifiers are remodeled by setting a boundary for weight using fuzzy theory (Ranjan & Prasad, 2018) and optimal weight is chosen using the proposed RIWO algorithm. On arrival of data d_i+1, the error e_i+1 will be computed and compared with that of the previous data d_i. If e_i < e_i+1, then prediction with training based on RIWO is made. Otherwise, the fuzzy bounding-based learning will be done by bounding the weights, which is given as: (40)${\displaystyle {\omega }^B={\omega }^t\pm F^s}$

where ω^t is weight at the current iteration and F^s signifies a fuzzy score. For the dynamic data, the features {F} were extracted. Here, the membership degree is given as: (41)${\displaystyle \text{Membership Degree}=\left|\right|{\omega }^{t-2}-{\omega }^{t-1}\left|\right|}$

where ω^t−2 represents weights at iteration t − 2 and ω^t−1 signifies weights at iteration t − 1. When the highest iteration is attained, the process is stopped.

The 20 Newsgroups database

The 20 Newsgroups dataset (Crawford, 2020) was curated by Ken Lang for newsreaders to extract the Netnews. The dataset was established by collecting 20,000 newsgroup data points split across 20 different newsgroups. The database is popular for analyzing text applications used to handle machine-learning methods such as clustering and text classification. The dataset is organized into 20 different newsgroups covering different topics.

Reuters database

The Reuters-21578 Text Categorization Collection Dataset was curated by David D. Lewis (NLTK Data, 2020). The dataset is comprised of documents collected from Reuters newswires starting in 1987. The documents are arranged and indexed based on categories. There were 21,578 instances in the dataset with five attributes. The number of websites attained by the dataset was 163,417.

Accuracy

Accuracy is described as the measure of data that is precisely preserved and is expressed as: (42)${\displaystyle \mathrm{Acc}=\frac{P+Q}{P+Q+H+F}}$

where P signifies true positive, Q symbolizes true negative, H denotes true false positive, and F is a false negative.

TPR

The TPR refers to the ratio of the count of true positives with respect to the total number of positives. (43)${\displaystyle \mathrm{TPR}=\frac{P}{P+H}}$

where P refers to true positives and H is the false negatives.

TNR

The TNR refers to the ratio of negatives that are correctly detected. (44)${\displaystyle \mathrm{TNR}=\frac{Q}{Q+F}}$

where Q is true negative and F signifies false positive.

Analysis with the Reuters dataset

The assessment of techniques with the Reuters dataset considering TPR, TNR, and accuracy parameters is described. The assessment is done with mapper = 3 and mapper = 4 and varying the chunk size.

(a) Assessment with mapper = 3

Figure 3 shows an assessment of techniques measuring accuracy, TPR, and TNR in the Reuter datasets with mapper = 3. The assessment of techniques with the TPR measure is depicted in Fig. 3A. For chunk size = 3, the TPR evaluated by LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.747, 0.757, 0.771, 0.776, 0.780, 0.785, 0.790, and 0.803, respectively. Likewise, for chunk size = 6, the TPR evaluated using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.800, 0.812, 0.818, 0.818, 0.818, 0.819, 0.819, and 0.830, respectively. The assessment of techniques measuring TNR is depicted in Fig. 3B. For chunk size = 3, the TNR evaluated using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.831, 0.842, 0.853, 0.861, 0.870, 0.880, 0.886, and 0.913, respectively. For chunk size = 6, the TNR evaluated using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.846, 0.850, 0.865, 0.869, 0.878, 0.885, 0.896, and 0.925, respectively. The assessment of the methods measuring accuracy is depicted in Fig. 3C. For chunk size = 3, the accuracy evaluated using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.8306, 0.842, 0.852, 0.861, 0.870, 0.880, 0.886, and 0.913, respectively. Likewise, for chunk size = 6, the accuracy evaluated using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.846, 0.850, 0.865, 0.869, 0.878, 0.885, 0.896, and 0.925, respectively. The performance improvement of LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, BPLion+LFNN with respect to the proposed RIWO-based deep residual network considering accuracy were 8.540%, 8.108%, 6.486%, 6.054%, 5.081%, 4.324%, and 3.135%, respectively.

(b) Assessment with mapper = 4

We assessed the techniques by measuring accuracy, TPR, and TNR, considering the Reuters dataset and using mapper = 4 (Fig. 4). The assessment of techniques using TPR is displayed in Fig. 4A. For chunk size = 3, the TPR evaluated using LSS-CNN was 0.754, RNN was 0.768, SLKNN+MLKNN was 0.792, SVM was 0.796, NN was 0.800, LSTM was 0.806, BPLion+LFNN was 0.810, and the proposed RIWO-based deep residual network was 0.828. Likewise, for chunk size = 6, the TPR evaluated using LSS-CNN was 0.810, RNN was 0.820, SLKNN+MLKNN was 0.824, SVM was 0.824, NN was 0.825, LSTM was 0.826, BPLion+LFNN was 0.826, and the proposed RIWO-based deep residual network was 0.850. The assessment of techniques measuring TNR is depicted in Fig. 4B. For chunk size = 3, the TNR evaluated using LSS-CNN was 0.839, RNN was 0.860, SLKNN+MLKNN was 0.863, SVM was 0.870, NN was 0.875, LSTM was 0.881, BPLion+LFNN was 0.896, and the proposed RIWO-based deep residual network was 0.925. Likewise, for chunk size = 6, the TNR evaluated by LSS-CNN was 0.855, RNN was 0.856, SLKNN+MLKNN was 0.876, SVM was 0.878, NN was 0.885, LSTM was 0.893, BPLion+LFNN was 0.900, and the proposed RIWO-based deep residual network was 0.940. The assessment of the methods measuring accuracy is displayed in Fig. 4C. For chunk size = 3, the accuracy evaluated by LSS-CNN was 0.837, RNN was 0.843, SLKNN+MLKNN was 0.846, SVM was 0.850, NN was 0.855, LSTM was 0.858, BPLion+LFNN was 0.862, and the proposed RIWO-based deep residual network was 0.880. For chunk size = 6, the accuracy evaluated by LSS-CNN was 0.833, RNN was 0.849, SLKNN+MLKNN was 0.852, SVM was 0.857, NN was 0.859, LSTM was 0.863, BPLion+LFNN was 0.868, and the proposed RIWO-based deep residual network was 0.887. The performance improvement of LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, and BPLion+LFNN with respect to the proposed RIWO-based deep residual network and considering accuracy was 6.087%, 4.284%, 3.945%, 3.38%, 3.156%, 2.705%, and 2.142%, respectively.

Analysis with the 20 Newsgroups dataset

The assessment of techniques using the 20 Newsgroups datasets with TPR, TNR, and accuracy parameters was elaborated. The assessment was done with mapper = 3 and mapper = 4 and by altering the chunk size.

(a) Assessment with mapper = 3

Figure 5 presents the assessment of techniques measuring accuracy, TPR and TNR and considering the 20 Newsgroups dataset with mapper = 3. The assessment of techniques measuring TPR is depicted in Fig. 5A. For chunk size = 3, the maximum TPR of 0.834 was determined using the proposed RIWO-based deep residual network, while the TPR found using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, and BPLion+LFNN were 0.708, 0.759, 0.780, 0.785, 0.792, 0.803, and 0.812, respectively. Likewise, for chunk size = 6, the highest TPR of 0.840 was found using the proposed RIWO-based deep residual network, while the TPR found using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, and BPLion+LFNN were 0.796, 0.815, 0.818, 0.822, 0.825, 0.828, and 0.829, respectively. The assessment of techniques measuring TNR is depicted in Fig. 5B. For chunk size = 3, the TNR computed using the proposed RIWO-based deep residual network was 0.862, while the TNR found using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, and BPLion+LFNN were 0.832, 0.839, 0.843, 0.846, 0.848, 0.850, and 0.851, respecitvely. Likewise, for chunk size = 6, the TNR evaluated using the proposed RIWO-based deep residual network was 0.879, while those computed by LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, and BPLion+LFNN were 0.833, 0.840, 0.849, 0.808, 0.851, 0.852, and 0.854, respectively. The assessment of the methods measuring accuracy is depicted in Fig. 5C. For chunk size = 3, the accuracy evaluated by the proposed RIWO-based deep residual network was 0.850, while the accuracies computed by LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, and BPLion + LFNN were 0.821, 0.822, 0.833, 0.837, 0.839, 0.840, and 0.843, respectively. The performance improvement of LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, and BPLion + LFNN with respect to the proposed RIWO-based deep residual network considering accuracy were 3.411%, 3.294%, 2%, 1.529%, 1.294%, 1.17%, and 0.823%, respectively. Likewise, for chunk size = 6, the accuracy evaluated using the proposed RIWO-based deep residual network was 0.859, while the accuracies computed by LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, and BPLion+LFNN were 0.824, 0.830, 0.839, 0.843, 0.849, 0.852, and 0.856, respectively.

(b) Assessment with mapper = 4

Figure 6 presents the assessment of techniques measuring accuracy, TPR, and TNR and considering the 20 Newsgroups dataset with mapper = 4. The assessment of techniques with TPR measure is depicted in Fig. 6A. For chunk size = 3, the TPR evaluated by LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.721, 0.769, 0.798, 0.801, 0.803, 0.806, 0.810, and 0.845, respectively. Likewise, for chunk size = 6, the TPR evaluated using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.810, 0.827, 0.836, 0.849, 0.840, 0.843, 0.846, and 0.859, respectively. The assessment of techniques measuring TNR is depicted in Fig. 6B. For chunk size = 3, the TNR evaluated by LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, and the proposed RIWO-based deep residual networks were 0.831, 0.831, 0.842, 0.867, 0.848, 0.854, 0.859, and 0.870, respectively. Likewise, for chunk size = 6, the TNR evaluated by LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.836, 0.839, 0.851, 0.857, 0.863, 0.866, 0.871, and 0.910. The assessment of the method measuring accuracy is depicted in Fig. 6C. For chunk size = 3, the accuracy evaluated using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.821, 0.831, 0.842, 0.846, 0.849, 0.854, 0.860, and 0.861, respectively. Likewise, for chunk size = 6, the accuracy evaluated using LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN, and the proposed RIWO-based deep residual network were 0.824, 0.838, 0.849, 0.852, 0.858, 0.863, 0.868, and 0.870, respectively. The performance improvements of LSS-CNN, RNN, SLKNN+MLKNN, SVM, NN, LSTM, BPLion+LFNN with respect to the proposed RIWO-based deep residual network and considering accuracy were 5.287%, 3.678%, 2.413%, 2.068%, 1.379%, 0.804%, and 0.229%, respectively.