Ep-Satty-MCDM based business decision-making model using behaviour and review data

Abstract

Reaching the maximum number of customers at the right time to increase sales and profitability of the business is the primary goal of Electronic Commerce (E-Commerce). However, owing to the low-influenced product, the profitability of e-commerce has been drastically affected in recent years. Therefore, this work proposes an Empirical Probability assigned Satty’s method integrated Multicriteria Decision-Making model (EP-Satty-MCDM)-based business decision-making model for improving the sales of low-influenced products by advertising with celebrities. Primarily, the authenticated user securely downloaded the encrypted data using Armstrong number private key generated-Trident Curve-Cryptography (Arm-TCC) in the web application. After that, the data is cleansed and the attributes of review and behavior data are extracted. Then, by utilizing the Interval-valued Atanassov intuitionistic fuzzy-based Mann-Whitney U test (IAF-MWU), the correlation between the review and behavior data is evaluated. The correlated features under each user are mapped under the product, and semantic ontology is constructed, where the data is again mapped with the product’s subsections. Afterward, the domains are extracted. Thereafter, to identify whether the product is high-influenced or low-influenced, the obtained ratings from ontology and extracted domains are inputted into the Boosting Regression Tree-Recurrent Neural Network (BRT-RNN). Then, for the decision-making, the positively forecasted celebrities with their garment and low-influenced products are given as the input to EP-Satty-MCDM. The experimental outcomes exhibited that the proposed technique withstands maximum accuracy when contrasted with the existing methodologies.

Keywords

Ontology boosting regression tree-recurrent neural network (BRT-RNN)empirical probability assigned Satty’s method integrated Multicriteria decision making model (EP-Satty-MCDM)semantic domain ontology decision-making model

1. Introduction

In the past few years, online businesses have been consistently growing worldwide in which customers spend more time browsing, voicing their opinions on brands and their products, and comparing products along with their prices across websites [1]. Globally, e-commerce sales are rising quickly, which allows businesses and consumers to distribute, buy, sell, and market goods and services via advertising in electronic systems [2, 3]. Making the consumers to practice comfortability owing to changes in behavior patterns and also for practicing requirements owing to technological advances and information flows is the major drawback in such models [4]. Because of this dynamic nature of business, the business process efficiency in highly competitive industries is affected when they tend to evolve through growth, transformation, or expansion [5, 6]. Thus, for coping with such evolvements and for ensuring the speed and consistency of the community’s requirements, companies must transform into more modern technologies as well as information packaging in which decision-making plays a vital role in ensuring the company’s success [7, 8] Also, business leaders look for implementing new technologies so that they could efficiently manage risk and optimize their business performance [9].

The business workers’ lives are made easier to expedite the decision-making process by the emergence of data mining mechanisms and business intelligence systems [10]. They examine the change in human interaction owing to social network building that occurs generally via online as well as social media marketing themes [11, 12]. Artificial Intelligence (AI) is gaining popularity in supporting operational decision-making across numerous business domains like Amazon and Target, which utilizes Machine Learning (ML) for recommending products along with product categories grounded on insights from other customers’ purchase decision processes [13, 14]. Nevertheless, the most accurate decision could not be produced by the decision-maker devoid of a detailed analysis centered on the enormous data received [15]. Moreover, they concentrated less on moderating the consumers’ buying behavior when the product fails to attain a satisfactory number of transactions and/or unacceptable customer feedback. Hence, for predicting sales on those low-influenced products, business models must build and implement a system that analyses huge data.

1.1 Problem statement

The prevailing business decision-making approaches are inefficient in addressing some important issues like,

Earlier, the models did not concentrate on purchasing behavior, knowledge discovery, and image mining altogether for developing product recommendation models.

The hedonic value of the social websites was considered very less as this might increase the moderating effect on buying behavior significantly.

Considering purchasing behavior and leaving other factors, namely reviews given and contextual analysis, et cetera could affect decision-making in business models.

The privacy of the recommender people was not considered, which may lead to business competitors raising an attack as well as sabotaging for recommender people.

Over-specialization or diversity problems occur in recommendation systems as they were trained narrowly by sorting vast amounts of content and selecting them.

Therefore, the proposed recommendation system has the following contributions,

The model effectively integrated purchasing behavior, knowledge discovery, and image mining altogether to make the most accurate decision with detailed analysis.

For the moderating effect on buying behavior, the information collection also contributed to the social media celebrity images.

Important factors that influence product sales were considered along with behavioral data to improve effectiveness in decision-making.

The Arm-TCC-based policy matching is introduced to minimize attacks and preserve the privacy of recommender people.

To alleviate the diversity problems in the recommendation system, the model conducts two different types of classification using BRT-RNN.

The remaining paper is organized to analyze the existing literature in Section 2, explain the proposed technique in Section 3, assess the proposed model in Section 4, and wind up the paper in Section 5.

2. Related literature survey

Niu et al., concentrated on increasing organizational effectiveness by improving the decision-making analysis. Regarding business intelligence, the model utilized data analysis via Optimized Data Management-based Big Data Analytics (ODM-BDA) and improved the risk-taking ability and plan failure through the backtracking method. The framework’s reliability was proven by the simulation analysis. Owing to the low quality of data, the model accuracy was lower [16].

Lessmann et al., implemented profit analytics in marketing. By utilizing a Profit-Conscious Ensemble Selection (PCES) technique, which integrated business and statistical objectives to select responsive customers, the model planned to select a more profitable target group for the targeted profit of the marketing campaign. The outcomes revealed that the framework was more advantageous than several benchmarks. Nevertheless, the black box problem in PCES was not resolved, which limited the model in understanding the factors influencing customer characteristics [17].

Kumar et al., enriched the recommendation generation’s performance in the E-Commerce system. The model improved customer retention performance by generating the purchase pattern for customers with similar interests and identifying the high-frequency pattern for producing recommendations to the user. Even though the model supported recommendation generation in a higher ratio, they did not consider the recommender people’s privacy [18].

Hossen et al., enhanced the business performance in the online hotel booking system by the assessment of customer satisfaction. For this, by utilizing Gated Recurrent Unit (GRU) as well as Long Short-Term Memory (LSTM), the sentiment about the customer opinion was identified. The deep learning models were trained on the customer review data. As per the findings, the business site had been improved through the analysis of customer opinion. The model only utilized a limited set of features, which led to a focus on the most popular items only and increased overstocking of other products [19].

Zong et al., propounded an Operational Research-centric Intelligent Decision Support System (OR-IDSS) for e-commerce decision-making. The fuzzy logic theory was utilized for promoting price negotiation in e-commerce organizations by the recommendation of choices for the product’s online selection. The experimental results exposed that when contrasted with the prevailing techniques, the model had enhanced performance. The model can break down gradually as the fuzzy logic system fits to specific problem domain [20].

Sun et al., evaluated the sustainability vector for e-commerce business development. Correlation to determine the company performance in economic and environmental aspects, clustering to group similar functioning groups, and regression for determining the business model’s key areas to be enhanced were conducted by the model. As per the results, augmenting the sustainability vector brought a company nearby to the business sustainability benchmark. The system lacked consideration of the companies’ public information, which lowered the E-commerce model’s accuracy [21].

Li et al., supported user’s purchasing decisions utilizing the Review of the Helpfulness-based Recommendation Methodology (RHRM). The review semantics were learned by the hybrid neural network for the review helpfulness classification; also, for the item recommendation generation, the past interactions in user preferences were utilized. The experiments exposed that reviewing helpfulness information was advantageous in the recommender system. Nevertheless, the model’s performance was degraded by the ignorance of the purchase history of items [22].

Wazarkar and Keshavamurthy proffered an image mining methodology for decision-making in fashion industries. The model gathered social fashion images and grouped them utilizing the soft clustering technique for extracting attributes on the grouped data, which were then applied to the forecasting model. Even though the system was beneficial in fashion industries for uplifting their business, the exclusion of judgment and subjective factors resulted in incorrect business decisions [23].

Karthik and Ganapathy recommended a product recommendation system to predict relevant products under the user’s interests. The product user target category’s sentiment score was calculated by the model and the ontology assignment made the decision using the fuzzy rules as well as an ontology-centric recommendation system. When contrasted with the prevailing mechanisms, superior performance was shown by the recommendation system’s experimental outcomes. Yet, the computation complexity of the model was high since the fuzzy-based systems required regular updates of rules [24].

Pan et al., employed a text mining technology to conduct reliability analysis on online reviews. The failure-related customer knowledge was extracted by the model from the user’s online reviews for classifying the failure for every single component and analyzing the reliability. The results specified that the consideration of failure distribution effects had a higher impact on product reliability. Nevertheless, the model lacked proper mapping between the failure mode and failure component, which affected the reliability analysis [25].

3. Proposed methodology for business decision-making model

The proposed technique aims to maximize sales by advertising lower-influenced products with social media celebrities. The proposed system’s block diagram is displayed in Fig. 1.

Figure 1.

Block Diagram of the proposed business decision-making model.

3.1 E-commerce Web applications

An e-commerce application allows users to browse product catalogs, add items to a cart, shop online, create wish lists, and complete purchases. Vast numbers of e-commerce web applications exist. Generally, many web applications exist. Among that, one web application $\left(W\right)$ is chosen to proceed with the proposed decision model.

Problem: A problem considered in the business decision-making model was that products with low influence caused lesser market sales. Thus, the decision-making model is prominent and is proposed here to quickly make the decision to improve the sales of the low-influenced product. The amount of data is gathered for decision-making to increase their lower-influenced product’s sales of $W$ . The data $\left(\Re\right)$ are initialized as,

$\displaystyle\Re=\left[{\Re^{1},\Re^{2},\Re^{3},\ldots\Re^{\Delta}}\right]$ (1)

Here, the $\Delta^{th}$ number of data is represented as $\Re^{\Delta}$ . The data will be in the format of decrypted format (unsecuritized format), which is securely obtained using the following process.

(a) Data encryption: By performing malicious activities, the unsecuritized data in $W$ may be attacked by business competitors. Hence, it is assumed in the proposed work that the web application’s data is encrypted and securely stored in the web applications. The proposed paper uses Arm-TCC for encrypting $\Re$ . The prevailing Elliptic Curve Cryptography (ECC) is computationally faster. Nevertheless, it uses a three-degree curve that was standard and publically known key operations. Thus, the keys are easily hacked by the business competitor. To solve this issue, the existing ECC is modified with a trident curve and Armstrong number. The trident curve derives the better parameters by adjusting the curve using more iteration. From these parameters, private keys are efficiently selected using the Armstrong number.

The cryptographic key algorithm (Arm-TCC) defines the trident curve $\left(C\right)$ as follows:

$\displaystyle C=\textit{QR}+aQ^{3}+bQ^{2}+cQ$ (2)

Where, $Q$ and $R$ signify the axis of the field, and $a$ , $b$ , and $c$ are the integers. From this derived curve, the third value $\left(\xi\right)$ is generated by recomputing Eq. (2) till getting the maximum curve value $\left(\beta\right)$ . Here, the third value is symbolized as the number of iterations for deriving the Eq. (2). For this third value, the best order is predicted randomly by iteration to change as the Armstrong number. The best order $\left({W^{\textit{pri}}}\right)$ should satisfy the following criteria to get the Armstrong number $\left(\alpha\right)$ . The criteria are expressed as,

$\displaystyle\alpha_{1}^{W^{\textit{pri}}}+\alpha_{2}^{W^{\textit{pri}}}+% \alpha_{3}^{W^{\textit{pri}}}+\alpha_{4}^{W^{\textit{pri}}}=\alpha$ (3)

The random number’s random initialization $\left({W^{\textit{pri}}}\right)$ is represented as,

$\displaystyle W^{\textit{pri}}=\left\{{W_{1}^{\textit{pri}},W_{2}^{\textit{pri% }},W_{3}^{\textit{pri}},\ldots W_{\infty}^{\textit{pri}}}\right\}$ (4)

Here, $\alpha_{1},\alpha_{2},\alpha_{3},\alpha_{4}$ are the digits in, and $W_{\infty}^{\textit{pri}}$ symbolizes the infinite number. $W^{\textit{pri}}$ is considered here as the private key of the admin. The private key is a securely sharable one between the admin and the user. Grounded on this private key, the public key of admin is generated as,

$\displaystyle W^{\textit{pub}}=\beta\ast W^{\textit{pri}}*\xi$ (5)

Thereafter, $\Re$ is encrypted. During the encryption process, the private key is added by the encryption formula. The signatures (data encryption) $\left({E_{1},E_{2}}\right)$ are created as,

$\displaystyle E_{1}=(\beta+\left({W^{\textit{pub}}}\right))+W^{\textit{pri}}$ (6) $\displaystyle E_{2}=(\Re+\left({W^{\textit{pub}}}\right)\ast r)+\kappa^{% \textit{sec ret}}$ (7)

Here, $r$ is therandom number. The pseudo-code for the proposed Arm-TCC is,

Input: Web application’s user Data

\left(\Re\right)

Output: Encrypted Data

\left({E_{1},E_{2}}\right)

and Decrypted Data

\left(\Re\right)

Begin

Initialize

Q

R

a

b

For obtain of maximum curve value, do

While

(\textit{iter}<\max_{\textit{iter}})

Define curve

C=\textit{QR}+\textit{aQ}^{3}+\textit{bQ}^{2}+\textit{cQ}

End while

End for

Estimate curve value

Define private key

If Satisfy the criteria

\alpha_{1}^{W^{\textit{pri}}}+\alpha_{2}^{W^{\textit{pri}}}+\alpha_{3}^{W^{% \textit{pri}}}+\alpha_{4}^{W^{\textit{pri}}}=\alpha

Consider

W^{\textit{pri}}

as private number

Else

Apply next random number

End if

Compute public key

W^{\textit{pub}}=\beta\ast W^{\textit{pri}}*\xi

Decrypt the data

E_{1}=(\beta+\left({W^{\textit{pub}}}\right))+W^{\textit{pri}}

E_{2}=(\Re+\left({W^{\textit{pub}}}\right)\ast r)+\kappa^{\textit{sec ret}}

Return

\left({E_{1},E_{2}}\right)

End

3.2 Authentication

Subsequent to the data encryption, the authentication is performed to collect the data securely from $W$ for processing the decision-making model. The privacy of the user is verified with a privacy policy for this authentication. For that privacy policy, privacy rules are generated, which comprise the purpose of taking the data, the name and identity of the person, the way it was logged in, previous history of browsing, and others. The features included in the rules are commonly symbolized as $\Gamma^{\textit{policy}}$ . The features of users should not violate the privacy policy like searching for any violated content, improper login, etcetera. The verification process of user’s privacy $\left({\Gamma^{\textit{user}}}\right)$ with privacy policy $\left({\Gamma^{\textit{policy}}}\right)$ is formulated as,

$\displaystyle V=\left\{{\begin{array}[]{ll}A^{\textit{user}},&\textit{if}\ % \Gamma^{\textit{policy}}=\Gamma^{\textit{user}}\\ \textit{declined},&\textit{if}\ \Gamma^{\textit{policy}}\neq\Gamma^{\textit{% user}}\\ \end{array}}\right.$ (8)

Here, $A^{\textit{user}}$ epitomizes the authenticated user to access the data from web applications. If the privacy of the user is satisfied with the privacy policy, then the user will be authenticated to access the data from web applications. If both are not satisfied, then the process for collecting data is declined.

3.3 Data collection

For the obtained authenticated user $\left({A^{\textit{user}}}\right)$ , the decryption key is given here by the admin to access their data. The private key is subtracted from the encrypted data during the decryption, and it is expressed as,

$\displaystyle\Re=(E_{2}-\left({W^{\textit{pub}}}\right)+E_{1})-W^{\textit{pri}}$ (9)

Here, $\Re$ signifies the decrypted data. To identify the low-influenced product $W$ , the overall web application data is collected after applying a decryption key by the authenticated person for downloading the data.

3.4 Ontology construction

Here, the collected data $\left(\Re\right)$ , which belongs to many users, is in an unstructured and random format. It makes the system more complex to analyze the data. Thus, the correlation between each data under each user is constructed by the ontology. Data cleansing, attribute extraction, attribute correlation, and data mapping are performed in this ontology construction.

3.4.1 Data cleansing

Data cleansing is performed after collecting the data $\left(\Re\right)$ . The data’s quality is improved by the cleansing phase, which also makes it more suitable for the specific data mining task. In the data cleansing, errors, inconsistencies, and duplications are found and removed. The data cleansing $\left({\Re^{\textit{clean}}}\right)$ is mathematically expressed as,

$\displaystyle\Re^{\textit{clean}}=\left[{\Re-\Re^{\textit{dup}}}\right]-\varsigma$ (10)

Here, $\Re^{\textit{dup}}$ implies the duplicated data, and $\varsigma$ is the error to be removed.

3.4.2 Attributes extraction

Following the data cleansing, the attributes of $\Re^{\textit{clean}}$ are extracted to reduce the training time during the product classification. From the review data, the user’s ID, brand, categories, name, reviews did purchase, manufacturer, review date, review user Province, reviews do recommend, review rate, review text, review title, review user city, review username, and user sentiment are extracted, and these attributes are signified as $\Re^{R}$ .

Likewise, the behavior data, namely work status, no. of clicks, gender, views, marital status, wishlist, frequency of purchase, amount of purchase, age, et cetera are extracted, and these attributes are notated as $\Re^{B}$ . The overall extracted attributes $\left(F\right)$ are written as,

$\displaystyle F=\Re^{R}\oplus\Re^{B}$ (11)

3.4.3 Attributes correlation and data mapping

User attributes $\left(F\right)$ placed randomly in the two datasets (groups) make the system more complex to classify. Attributes Correlation means the activities of the user to purchase the product, such as when the product was allotted to the wishlist, when the product was added to the cart, when it was purchased, etc. Grounded on the time and date, the activities of each user are correlated continuously. Therefore, by utilizing IAF-MWU, the relationship between the attributes is correlated. Mann-Whitney U test (MWU) can find significance even between both groups that are independent of each other. Moreover, for analyzing differences betwixt the medians of 2 datasets, MWU is utilized. However, the null value hypothesis undergoes an if else condition randomly that increases the length of the calculation. Therefore, it uses a modified Interval-valued Atanassov intuitionistic fuzzy set.

The MWU is performed in 2 parts. The MWU’s main part is represented by the first part, whereas the second part is the computation of the median of every single group. The Analyze menu (attribute’s different category) opens first when the database to be processed is activated. The analyzed menu options $\left(M\right)$ are expressed as,

$\displaystyle M=\left[{M^{1},M^{2},M^{3},\ldots M^{h}}\right]$ (12)

Here, $M^{h}$ specifies the $h^{th}$ menu. Afterward, two Independent Samples are selected. They are notated as $F^{i}$ and $F^{j}$ . Following this, the continuous dependent variable $\left({F^{\textit{dep}}}\right)$ on the extracted attributes is shifted in the field test variable. The shifting function is formulated as,

$\displaystyle F^{\textit{shift}}=f\left({F\left({F^{i}*F^{j}}\right)+F^{% \textit{dep}}}\right)$ (13)

For encrypting the variable’s values, it is required to choose the Define Groups field. Numbers 1 and 2 are entered in the boxes adjacent to each group, correspondingly. The MWU field in the section Test Type is then validated before selecting the command Continue. Lastly, Options are chosen $\left({o^{\textit{sel}}}\right)$ , attributes are checked for command continue $\left(C\right)$ , and OK is entered. The final operation in the test field is mathematically expressed further:

If $\left({o^{\textit{sel}}}\right)$ is chosen and the attributes are not satisfied $\left({D^{n-\textit{satis}}}\right)$ with satisfied continuity, then the equation will be,

$\displaystyle V=o^{\textit{sel}}|D^{n-\textit{satis}}=G^{1}\left(C\right)+% \textit{rand}$ (14)

Here, rand epitomizes the random number to avoid the attributes’ insufficiencies, and $G^{1}$ exemplifies the group. After completing testing, the median of each group is calculated using Interval-valued Atanassov intuitionistic fuzzy as,

$\displaystyle\mathchar 22\mskip-10.0mu \lambda^{\textit{med}}=\left\{{<F,\left% [{\eta^{-}\left(F\right),\eta^{+}\left(F\right)}\right],\left[{v^{-}\left(F% \right),v^{+}\left(F\right)}\right]>F^{i},F^{j}}\right\}$ (15)

Here, the membership degree is exemplified as $\left[{\eta^{-}\left(F\right),\eta^{+}\left(F\right)}\right]$ , and the non-membership degree is notated as $\left[{v^{-}\left(F\right),v^{+}\left(F\right)}\right]$ . Afterward, the dependent variable is shifted $\left({F^{\textit{dep}}}\right)^{*}$ in the field Dependent List. The shifting function $\left({F^{\textit{shift}}}\right)^{*}$ is expressed as,

$\displaystyle\left({F^{\textit{shift}}}\right)^{*}=f\left({F\left({F^{i}*F^{j}% }\right)+\left({F^{\textit{dep}}}\right)^{*}}\right)$ (16)

Thereafter, the Median switches from the Statistics field to the field of Cell Statistics. Also, the Mean and Standard Deviation switch from the field of Cell Statistics to the Statistics field. At last, the commands Continue and OK are given. Finally, the obtained correlated attributes of one user are expressed as,

$\displaystyle A_{k}^{\textit{cor}}=\left[{A_{1}^{\textit{cor}}\oplus A_{2}^{% \textit{cor}}\oplus A_{3}^{\textit{cor}}\oplus\ldots A_{K}^{\textit{cor}}}\right]$ (17)

Here, $A_{K}^{\textit{cor}}$ signifies the $K^{th}$ correlated attribute of the user. $K$ and $k$ exemplify the number of correlated attributes. The obtained correlated attributes are notated as $\Psi$ . Following the identification of the correlation between the attributes placed under each user, the correlated attributes of each user are mapped under the product. For instance, the list of each user’s action (correlated attributes) is based on the purchasing product, and these attributes are matched with their purchased product $\left(P\right)$ . The mapped data $\left({\Psi^{\textit{map}}}\right)$ is expressed as,

$\displaystyle\Psi^{\textit{map}}=\frac{1}{K}\sum\limits_{k=1}^{K}{\left(P% \right)}\left(\Psi\right)^{T}$ (18)

From this constructed ontology, the review ratings are taken as the output, and it is represented as $r$ .

3.5 Semantic ontology construction and domain extraction

Following the construction of ontology centered on the product’s characteristics (as the parent node), the semantic ontology is constructed here. By developing the child node (product’s sub-categories, such as size, cloth, design, et cetera), the semantic ontology over $\Psi^{\textit{map}}$ is constructed. The $\Psi^{\textit{map}}$ comprises several attributes and it is added to the subcategories of each product. They are expressed as,

$\displaystyle S=\Psi^{\textit{map}}*\sum{B^{Y}}$ (19)

Here, $B^{Y}$ signifies the $Y$ number of sub-categories, and $S$ epitomizes the overall attributes. The graph $\left({G^{\Im}}\right)$ is drawn manually as,

$\displaystyle g^{S}=\left({S,\kappa}\right)$ (20)

Here, the link between the two nodes is considered an edge point $\left(\kappa\right)$ . This ontology comprises numerous child nodes. Each node encloses detailed information about the product with its subcategories, User ID, purchasing and reviewing details, etc. Thus, it helps in easily extracting the details to classify the influenced product. It does not perform with constructed ontology owing to the designed characteristics of the classifier. Hence, the domain details are extracted and given as the input into the classifier. The extracted domains are specified as $D$ .

3.6 Product classification

Here, for classifying high and low-influenced products of web applications, the evaluated rating $\left(r\right)$ and extracted domains along with their product $\left(D\right)$ are inputted into the trained BRR-RNN. The conventional Recurrent Neural Network (RNN) has a higher ability to handle Variable-Length Sequences and memory of past inputs. However, using a normal regression tree in RNN makes the process slow down. To overcome this issue, the paper uses a Boosting Regression Tree (BRT) rather than using a normal regression tree. BRT significantly learns the data by dividing it into small subsets and provides the possible outcome. Figure 2 displays the architecture of BRR-RNN,

Figure 2.

Structure of proposed SR-RNN.

Here, the rating of each product by users $\left(r\right)$ and extracted domains $\left(D\right)$ are commonly symbolized as $\aleph$ .

Input layer: This layer is responsible to get the inputs and feed them to the hidden layer. This is represented as,

$\displaystyle\aleph_{i}=\{\aleph_{1},\aleph_{2},\aleph_{3},\ldots,\aleph_{z}\}$ (21)

Hidden layer: This layer is activated by the sigmoid activation function and trains the input by aggregating it with weight value and bias value. The Hidden layer $\left({H_{i}}\right)$ is computed as,

$\displaystyle H_{i}=\delta[\omega_{IH}\alpha_{i}+W_{HH}H_{i-1}+b]*\Im_{i}$ (22)

Where, $\varpi_{\textit{IH}}$ is the input $(I)$ -hidden $(H)$ weight value, $\omega_{\textit{HH}}$ is the hidden layer weight value, and $\Im_{i}$ symbolizes BRT, which learns the appropriate data from several domains with the number of subsections by constructing parent and child nodes. The ReLu activation function is expressed as,

$\displaystyle\delta=\max\left({0,\aleph_{i}}\right)$ (23)

By computing the weighted sum and further adding bias to it, the activation function decides whether a neuron should be activated or not. The BRT is evaluated as,

$\displaystyle\Im_{i}=\sum{s^{\aleph_{i}}}$ (24)

Here, $s^{\aleph_{i}}$ symbolizes the score value of the attributes.

Output layer: Afterward, the outcome from the hidden layer is given to the output layer, which classifies the product as high-influenced or low-influenced. This output layer $\left({O_{i}}\right)$ is determined by,

$\displaystyle O_{i}=\sigma[\omega_{HO}H_{i}+b]$ (25)

Where, $\omega_{\textit{HO}}$ implies the hidden-output weight value, and $\sigma$ epitomizes the sigmoid activation function, which is expressed as,

$\displaystyle\sigma=\frac{e^{H_{i}}}{\sum\nolimits_{i=1}^{z}{e^{H_{i}}}}$ (26)

Loss function: Thereafter, by computing the difference between the actual value $(\textit{Av})$ and the predicted value $(\textit{Pv})$ , the loss value is evaluated. The error value is calculated as,

$\displaystyle L=(\textit{Av}-\textit{Pv})^{2}$ (27)

If the model’s loss value $(L)$ or else the error value is zero $(L=0)$ , then it indicates that the model accurately detects the product’s status. The back-propagation takes place by updating the weight values if the loss values $L\neq 0$ . Lastly, this classifier significantly classifies the product as low influenced $l^{in}$ or high influenced $h^{in}$ . The pseudo-code of the proposed BRR-RNN is,

Input: Evaluated rating

\left(r\right)

and Extracted domains

\left(D\right)

Output: Low

l^{in}

or High influenced

h^{in}

product

Begin

Initialize inputs

(\aleph_{i})

, weights

(W_{i})

, bias

(b_{i})

\textit{Max}_{\textit{itr}}

For

(\textit{Max}_{\textit{itr}}=1\,to\,z)

Evaluate the hidden layer operation using,

H_{i}=\delta[\omega_{\textit{IH}}\alpha_{i}+W_{\textit{HH}}H_{i-1}+b]*\Im_{i}

Perform ReLu activation function

\delta=\max\left({0,\aleph_{i}}\right)

Evaluate the output layer operation using,

O_{i}=\sigma[\omega_{\textit{HO}}H_{i}+b]

(\textit{Av}==\textit{Pv})

Accurate classification of products

Else

Perform Backpropagation by updating weights

End if

End for

End

3.7 Social media celebrity image mining

Here, the number of different social media celebrity images $\left(I\right)$ is taken and considered as one of the inputs for the decision-making model. Providing a great suggestion for advertising the low-influenced product with the best celebrity is the main motive behind considering the celebrity image. The decision-making model judges the advertisement of low-influenced products with the best celebrity, which is explained in the final phase. $I$ consists of several images for each celebrity. From this, images of one celebrity are considered for example. The Garment segmentation and pattern matching were analyzed for these images. The segmentation is mathematically represented as,

$\displaystyle t=\varpi_{g}(\lambda_{g}-\lambda_{I})^{2}+\varpi_{B}(\lambda_{B}% -\lambda_{I})^{2}$ (28)

Wherein, $\Delta$ exemplifies the maximum pixel value in the input image, $\varpi_{g}$ and $\varpi_{B}$ are cumulative probabilities of garment and background region, $\lambda_{g}$ and $\lambda_{B}$ are the adaptive mode of garment and background region, $\lambda_{I}$ implies the adaptive mode of the whole image, and $t$ signifies the optimal thresholds. The garment is segmented grounded on the threshold level and it is represented as,

$\displaystyle\vartheta=\left\{{\begin{array}[]{ll}X_{g}&\Omega>\Im\\ X_{\textit{back}}&\Omega<\Im\\ \end{array}}\right.$ (29)

Here, $\vartheta$ symbolizes the output of segmentation, the number of pixels in the input image is notated as $\Omega$ , $X_{g}$ signifies the segmented garment region, and $X_{\textit{back}}$ epitomizes the segmented background region. Thereafter, the segmented garment’s patterns are analyzed. To identify the celebrity who is forecasted as positive or negative with their garments, the analyzed patterns are given as the input to BRT-RNN. The celebrity with the highly analyzed patterns $\left({\rho^{\textit{ana}}}\right)$ is mathematically expressed using BRT-RNN as,

$\displaystyle\rho^{\textit{ana}}=\left\{{\begin{array}[]{ll}\left({\left({X_{g% }}\right)^{\textit{type1}}}\right)^{*},&\textit{if}\,L\left\{{\left({X_{g}}% \right)^{\textit{type1}}>\left({X_{g}}\right)^{\textit{type2}}}\right\}\\ \left({\left({X_{g}}\right)^{\textit{type2}}}\right)^{*},&\textit{if}\,L\left% \{{\left({X_{g}}\right)^{\textit{type1}}<\left({X_{g}}\right)^{\textit{type2}}% }\right\}\\ \end{array}}\right.$ (30)

Here, $L$ symbolizes the number of likes for the celebrity, and $\left({X_{g}}\right)^{\textit{type1}}>\left({X_{g}}\right)^{\textit{type2}}$ signifies the garment types. The celebrities with the higher influenced patterns are identified based on the higher likes for the garments. Similarly, the analysis of the number of celebrities using the number of images is done. Each celebrity with highly influenced garments is represented as,

$\displaystyle\Phi=\left[{\Phi^{1},\Phi^{2},\Phi^{3},\ldots\Phi^{Q}}\right]$ (31)

Here, $\Phi^{Q}$ specifies the $Q^{th}$ Positively forecasted celebrity.

3.8 Business decision-making model

Here, to make the decision for improving sales by matching the celebrity with the product to advertise, the positively forecasted celebrity $\left(\Phi\right)$ and classified low-influenced products $\left({l^{in}}\right)$ are inputted into the EP-Satty-MCDM. The traditional Multicriteria Decision-making model (MCDM) is an efficient decision-making system. Nevertheless, by utilizing a random process, the weight assignment was done, which impeded the system to provide the optimal decision. To resolve this issue, the paper uses Empirical probability assigned to Satty’s method for weight initialization. It optimally estimates the weight value and reduces the iteration of the proposed model.

In multicriteria decision-making, the rating prediction task $\left(\varepsilon\right)$ could be explicated as,

$\displaystyle\varepsilon=\left\{{\varepsilon^{1},\varepsilon^{2},\varepsilon^{% 3},\ldots\varepsilon^{\kappa}}\right\}$ (32)

Where, $\varepsilon^{\kappa}$ signifies the $\kappa^{th}$ prediction task. The rating prediction or else aggregation function is notated as $\tau$ , which tries to use the multi-criteria ratings for predicting the overall rating.

$\displaystyle\varepsilon=\tau\left\{{\varepsilon^{1},\varepsilon^{2},% \varepsilon^{3},\ldots\varepsilon^{\kappa}}\right\}$ (33)

Two steps follow the assignment of the prediction task. The process of estimating the rating on every single criterion is referred to as both rating aggregations and multi-criteria rating predictions. The process of combining the predicted multi-criteria ratings for estimating the overall rating is referred to as rating aggregations. Grounded on these anticipated overall ratings, the recommended items will be created.

(a) Multi-Criteria Rating Predictions: Here, the multicriteria ratings between $\Phi$ and $\iota^{in}$ are predicted independently and dependently. By utilizing the standard biased matrix factorization associated with every single criterion, the individual rating on each criterion is predicted in the dependent and independent techniques. Standard biased matrix factorization $\left({\iota^{in}}\right)$ is represented as,

$\displaystyle\iota^{in}=\eta+\psi^{\Phi}+\psi^{l^{in}}q\bullet u$ (34)

Here, $\eta$ signifies the global average rating of $\Phi$ , $\psi^{\Phi}$ and $\psi^{l^{in}}$ symbolize latent-factor vectors of $\Phi$ and $\iota^{in}$ , $q$ and $u$ are the bias of $\Phi$ and $\iota^{in}$ . Like this, the correlation amongst multiple criteria between $\Phi$ and $\iota^{in}$ will be considered. This might enhance the predictive performance.

Rating Aggregations: Once the multi-criteria ratings are predicted, they are aggregated together for predicting the overall rating for making the decision. The linear aggregation is epitomized as,

$\displaystyle\varepsilon=\chi_{1}*\varepsilon^{1}+\chi_{2}*\varepsilon^{2}+% \ldots+\chi_{\kappa}\varepsilon_{\kappa}+\tau$ (35)

Here, $\chi_{1}$ , $\chi_{2}$ , $\ldots$ , $\chi_{\kappa}$ are the weights of $\Phi$ and $\iota^{in}$ . The weight values $\left(\chi\right)$ are evaluated using Empirical probability assigned Satty’s method as,

$\displaystyle\chi=\frac{\nu^{\textit{ratings}}\left({\Phi\oplus l^{in}}\right)% }{\nu^{\textit{tri}}}$ (36)

Here, the number of events performed is explicated as $\nu^{\textit{ratings}}$ , and the total number of multi-criteria ratings performed is elucidated as $\nu^{\textit{tri}}$ . From the multiple ratings, celebrity along with the product and higher rating is given as the decision for the proposed decision-making model to advertise for enhancing the sales of the business.

4. Results and discussion

Here, the proposed recommendation system’s efficacy in improving the business performance of the E-commerce system is evaluated by conducting numerous experiments in the working platform of PYTHON.

4.1 Performance analysis

This phase provides the experimental results along with performance analysis as well as a comparative analysis of different methods. Grounded on some quality metrics, the results attained for various stages of the system that use the methods, such as BRT-RNN, EP-Satty-MCDM, IAF-MWU, and Arm-TCC are evaluated.

Figure 3.

Performance comparison of proposed EP-Satty-MCDM based on (a) Geometric average and (b) Net score.

Regarding Geometric average and net score, the overall performance of the system by comparing the proposed EP-Satty-MCDM and existing Satty-MCDM, Fuzzy-MCDM, MCDM, and Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) is analyzed in Fig. 3. When contrasted with the prevailing techniques, the proposed technique has the highest performance for the global average (0.229343) and net score (0.99). This indicates that the EP-based weight assignment aided in the analytics of relevant data from large amounts of data for confident decision-making.

Table 1

Performance comparison of proposed BRT-RNN

Methods/metrics	Product classification accuracy (%)	Category influence classification accuracy (%)
Proposed BRT-RNN	98	98
RNN	95	96
DNN	90	92
CNN	88	87
ANN	84	83

In Table 4, the accuracy of the proposed BRT-RNN and existing RNN, Convolution Neural Network (CNN), Deep Neural Network (DNN), and Artificial Neural Network (ANN) is compared for their ability to product classification and category influence classification. While comparing with the existing approaches, the accuracy attained by the proposed model is improved by 3% for product classification and 2% for category influence classification. This indicates that BRT explicitly fits parameters to direct the information flow and speeds up the process to make more accurate predictions.

Figure 4.

Analyzing the performance of the proposed BRT-RNN.

The F-measure, recall, and precision of the proposed BRT-RNN and existing systems are analyzed in Fig. 4. In comparison, the proposed system achieves higher values of precision (97.8%), recall (97%), and F-measure (97.9%) than the existing methods. Therefore, the BRT technique used in the model effectively defines the predictive attributes for accurate predictions. The analysis concludes that for product classification, the model remains more accurate.

Table 2

Analysis of failure prediction

Methods/metrics	Failure proportion
Proposed BRT-RNN	0.87
RNN	1.25
DNN	1.71
CNN	2.51
ANN	2.81

Table 3

Encryption and decryption analysis

Methods/metrics	Encryption time (ms)	Decryption time (ms)
Proposed Arm-TCC	75	60
ECC	83	71
RSA	91	92
AES	126	108
QCA	172	162

Figure 5.

Analysis of forecasting accuracy.

Concerning the error metrics, namely Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Normalized Mean Square Error (NMSE), the performance of proposed and existing approaches in forecasting the influence of category is analyzed in Fig. 5. Attaining minimum values of these metrics indicates improved performance of the model. By the way, the findings in Fig. 4 indicate the maximum performance of the proposed BRT-RNN by achieving minimum values of NMSE (3.3%), MAE (2.34%), MAPE (1.88%), and RMSE (1.52%). Therefore, BRT associated with the proposed technique significantly improved the prediction accuracy with minimum forecasting error.

Table 2 compares the failure proportions attained by the proposed and conventional techniques, which should be lower for better performance. The proposed mechanism has the highest performance than the existing methods. This is owing to the assistance of the BRT technique that effectively interprets and understands the non-linear relationship between the data for accurate predictions. Therefore, to support the improvement of decision-making performance, the predictions on product classification and category influence forecasting remain more accurate.

The encryption time and decryption time analysis of the proposed Arm-TCC and existing Rivest Shamir Adleman (RSA), Quantum dot Cellular Automata (QCA), ECC, and Advanced Encryption Standard (AES)methods are revealed in Table 3. The time taken by the proposed model is lower by 8 ms for encryption and 11 ms for decryption compared to the existing ECC. Hence, the selection of private keys using the Armstrong number requires more resources, time, and effort than the random selection. From the analysis, the proposed Arm-TCC is more highly secure in preserving the privacy of recommender people than the prevailing mechanisms.

Table 4

AUC analysis

Methods/metrics	AUC
Proposed IAF-MWU	0.8315
MWU	0.7921
Chi-square t-test	0.7527
T-test	0.6128
U-test	0.5921

Figure 6.

Execution time analysis.

In Fig. 6, the execution time of the proposed IAF-MWU and the existing approaches estimated are displayed. The proposed technique’s execution time is much minimum by 598 ms than the existing MWU. The minimum execution time taken by the proposed system indicates that the reduction in null hypothesis calculations makes the model solve problems quicker than the existing approaches.

Concerning the Area Under Curve (AUC), the performance of the proposed IAF-MWU and existing MWU, Chi-square t-test, T-test, and U-test methods are analyzed in Table 4. The AUC attained by the proposed model is improved by 3.94%, 7.88%, 21.87%, and 23.94% than the existing methods. Therefore, the analysis states that the null value hypothesis under the IAF technique reduced the number of calculations to make the model more efficient than the existing techniques.

Figure 7.

Comparative analysis.

In Fig. 7, the proposed BRT-RNN-based EP-Satty-MCDM (BR-ERSM) business decision-making model is compared with the existing LSTM, OR-IDSS, CNN-LSTM, and image mining techniques developed by [19, 20, 22, 23] in Section 2. The comparison was done based on the accuracy attained in improving their decision-making ability. Figure 7 exhibits that the accuracy attained by the proposed model is far superior to the prevailing method, which is 98%. This is owing to the detailed analysis of the proposed model by incorporating different types of data. The proposed model can benefit E-commerce systems in a high ratio by identifying failure products and recommending those for the users according to their purchase patterns and influencing capacity. Moreover, the recommendations from the proposed model were generated grounded on different analytical elements, namely data analysis, image mining, and knowledge recovery, which the existing methods failed to consider. Therefore, the proposed BR-ERSM model outperforms the existing methods.

5. Conclusion

By integrating data analysis, knowledge discovery, and image mining, this paper proposes a business decision-making model. Research on the issues of increase in moderating effect, the privacy of recommender people, and considering reviews and contextual analysis factors for informed decision-making aids the entire system to suit business platforms. The performance of the proposed EP-Satty-MCDM decision-making model and BRT-RNN-based classification techniques are compared with some prevailing techniques in the experimental evaluation. The proposed EP-Satty-MCDM yielded a Net score of 0.99, and the BRT-RNN method attained a classification accuracy of 98%. Hence, it is deduced from the outcomes that the proposed business decision-making model on an E-commerce system works well, and it is more efficient and faster for the optimized business performance. The image mining technique utilized in the proposed model concentrated only on the celebrity images grounded on their garments and influence but not on the advertisements they were involved in. Therefore, the work might be extended in the future to analyze the market value, advertising influence, and usage influence to improve the business model.

References

Smedberg

and Bandaru

, Interactive knowledge discovery and knowledge visualization for decision support in multi-objective optimization, European Journal of Operational Research 306 (2023), 1311–1329.

Wilfling

Havenith

Raccuglia

and Hodder

, Can you see the feel? The absence of tactile cues in clothing e-commerce impairs consumer decision making, International Journal of Fashion Design, Technology and Education 16 (2022), 224–233.

Bahari

Basalamah

Ashoer

and Syahnur

M.H.

, Internet Business: How Perception of Benefits, Risks, and Ease in Decision Making, Jurnal Manajemen Bisnis 12 (2021), 1–11.

Coussement

and Benoit

, Interpretable data science for decision making, Decision Support Systems 150 (2021), 1–6.

Al-Anqoudi

Al-Hamdani

Al-Badawi

and Hedjam

, Using machine learning in business process re-engineering, Big Data and Cognitive Computing5 (2021), 1–19.

Chaudhuri

Gupta

Vamsi

and Bose

, On the platform but will they buy? Predicting customers’ purchase behavior using deep learning, Decision Support Systems 149 (2021), 1–10.

Mahoto

N.A.

Iftikhar

Shaikh

Asiri

Alghamdi

and Rajab

, An intelligent business model for product price prediction using machine learning approach, Intelligent Automation and Soft Computing 30 (2021), 147–159.

Rahman

M.S.A.

Jamaludin

N.A.A.

Zainol

and Sembok

T.M.T.

, Machine Learning Algorithm Model for Improving Business Decisions Making in Upstream Oil Gas, in: 2021 International Congress of Advanced Technology and Engineering, ICOTEN, Taiz, Yemen, 2021, pp. 1–6.

Tamang

M.D.

Kumar Shukla

Anwar

and Punhani

, Improving Business Intelligence through Machine Learning Algorithms, in: Proceedings of 2021 2nd International Conference on Intelligent Engineering and Management, ICIEM London, United Kingdom, 2021, pp. 63–68.

10.

Dean

P.W.

, Artificial Intelligence and Machine Learning In Business, International Journal on Integrated Education E-ISSN 2 (2019), 64–68.

11.

Sujith

A.V.L.N.

Qureshi

N.I.

Dornadula

V.H.R.

Rath

Prakash

K.B.

and Singh

S.K.

, A Comparative Analysis of Business Machine Learning in Making Effective Financial Decisions Using Structural Equation Model (SEM), Journal of Food Quality (2022), 1–7.

12.

Huenermund

Kaminski

J.C.

and Schmitt

, Causal Machine Learning and Business Decision Making, Academy of Management Proceedings 2021 (2021), 1–53.

13.

Rajagopal

N.K.

Qureshi

N.I.

Durga

Asis

Ramirez E.H.

Huerta Soto

R.M.

Gupta

S.K.

and Deepak

, Future of Business Culture: An Artificial Intelligence-Driven Digital Framework for Organization Decision-Making Process, Complexity (2022), 1–14.

14.

Ren

, Optimization of Enterprise Financial Management and Decision-Making Systems Based on Big Data, Journal of Mathematics (2022), 1–11.

15.

Zhao

and Keikhosrokiani

, Sales prediction and product recommendation model through user behavior analytics, Computers, Materials and Continua 70 (2022), 3855–3874.

16.

Niu

Ying

Yang

Bao

and Sivaparthipan

C.B.

, Organizational business intelligence and decision making using big data analytics, Information Processing and Management 58 (2021), 1–13.

17.

Lessmann

Haupt

Coussement

and De Bock

K.W.

, Targeting customers for profit: An ensemble learning framework to support marketing decision-making, Information Sciences 557 (2021), 286–301.

18.

Kumar

M.R.

Venkatesh

and Rahman

A.M.J.M.Z.

, Data mining and machine learning in retail business: developing efficiencies for better customer retention, Journal of Ambient Intelligence and Humanized Computing (2021), 1–13.

19.

Hossen

M.S.

Jony

A.H.

Tabassum

Islam

M.T.

Rahman

M.M.

and Khatun

, Hotel review analysis for the prediction of business using deep learning approach, in: Proceedings – International Conference on Artificial Intelligence and Smart Systems, ICAIS, 2021, Coimbatore, India, pp. 1489–1494.

20.

Zong

Yuan

Montenegro-Marin

C.E.

and Kadry

S.N.

, Or-based intelligent decision support system for e-commerce, Journal of Theoretical and Applied Electronic Commerce Research 16 (2021), 1150–1164.

21.

Sun

Grondys

Hajiyev

and Zhukov

, Improving the e-commerce business model in a sustainable environment, Sustainability (Switzerland) 13 (2021), 1–22.

22.

Lee

and Kim

, A hybrid cnn-based review helpfulness filtering model for improving e-commerce recommendation service, Applied Sciences (Switzerland) 11 (2021), 1–20.

23.

Wazarkar

and Keshavamurthy

B.N.

, Social image mining for fashion analysis and forecasting, Applied Soft Computing Journal 95 (2020), 1–16.

24.

Karthik

R.V.

and Ganapathy

, A fuzzy recommendation system for predicting the customers interests using sentiment analysis and ontology in e-commerce, Applied Soft Computing 108 (2021), 1–18.

25.

Pan

Wang

You

Zhang

and Yang

, Assessing the reliability of electronic products using customer knowledge discovery, Reliability Engineering and System Safety 199 (2020), 1–1.