Cloud Services User’s Recommendation System Using Random Iterative Fuzzy-Based Trust Computation and Support Vector Regression

Abstract

Cloud computing is now a fundamental type of computing due to technological innovation and it is believed to be a benefit for mid-scale enterprises. The use of cloud computing is increasing daily, which improves service quality but also gives rise to security concerns. Finding trustworthy service can be very challenging, take a great deal of time, or produce subpar services. Due to these difficulties, the client needs a service that is dependable, suitable, time-saving, and trustworthy. As a result, from the end user’s perspective, adopting a cloud service’s trustworthiness becomes crucial. Trust is a measure of how well users’ expectations about a service’s capabilities are realized. In this research, a recommendation system for cloud service customers based on random iterative fuzzy computation (RIFTC) is proposed. RIFTC focuses on the assessment of trust using Quality of Service (QoS) characteristics. RIFTC calculates trust using the machine learning approach Support Vector Regression (SVR). RIFTC can helpfully recommend a cloud service to the end user and anticipate the trust values of cloud services.. Precision (97%), latency (51%), throughput (25.99 mbps), mean absolute error (54%), and re-call (97%) rates are used to assess how well this recommendation system performs. RIFTC’s average F-measure rate is calculated by adjusting the number of users from 200 to 300, and it is 93.46% more accurate on average with less time spent than the current methodologies.

1. Introduction

Cloud computing, which is used in contemporary computing, evolved from grid computing and service-oriented architecture. Web services were essential to the information technology sector before the widespread use of cloud computing [1]. Web services enable compatible machine-to-machine communication over the internet and provide a way to use computer services remotely, similar to how remote procedure calls work. Extensible markup language (XML) service architectures are typically represented using the Web services description language. Applications that use web services include electronic commerce, corporate systems built using Service-Oriented Architecture (SOA), and grid computing. They merely provide a detailed explanation of the products, providing the intriguing semantic properties required for flexible network activation and detection. In cloud computing, cloud servers offer all necessary hardware and software as services to consumers [2]. Figure 1 shows the cloud-based recommendation system.

The filtering approach is a technique that uses a variety of filtering characteristics to display only the commodities that are most advantageous to the recipient out of a wide range of available commodities. In the literature, there are various methods for filtering a RS, including (i) content-based (CB), (ii) CF, (iii) hybrid filtering (HB), (iv) knowledge-based (KB), and (v) context-aware (CA). For the past few years, CF has become the more widely used filtering method among these. CF bases its operations on the fact that user behaviors, purchases, ratings, and preferences are compared, and this data is then used for further analysis.

Users can access several on-demand services thanks to cloud computing, a new idea that represents a collaboration between numerous machines and services across a network. Alternatively, defined, cloud computing is a novel, entirely web-based strategy that offers a top-notch computing system that is readily accessible, scalable, and adaptable for a range of functions. Cloud computing disperses various virtualized resources across numerous networks and locals [3,4]. These virtualized resources fall into 4 different categories: Platform as a Service (PaaS), Expert as a Service (EaaS), Software as a Service (SaaS), and Infrastructure as a Service (IaaS). Customers can access the software via the Internet with SaaS without any difficulty or costs. Along with hosting, hardware, provisioning, and other requirements for operating a cloud, IaaS provides several other services. In this study, the term PaaS refers to an advanced integrated environment where unique applications are developed, tested, and deployed. Individuals can access users with specialized knowledge and skills through Expert Services (EaaS). Recommender systems are one of the most often-used tools for online users looking for opinions and recommendations on various goods or destinations [5,6]. People who have similar thoughts and preferences and have previously preferred the same items can get references about other products through recommender systems. Sentiment analysis engines are created across various social media platforms as a crucial component of the user experience [7,8]. If customers wish to select the finest or most appropriate services from a wide range of available services, the availability of services in the Cloud market and their growth could provide a serious challenge [9,10]. Choosing the best Cloud storage service would be difficult as a result. The usage of recommendation systems (RS) is one of the computer-based intelligence strategies for overcoming the difficulty of choosing the finest services or products from a vast range of possibilities. In the majority of e-business fields, they may be effectively used to deliver individualized services, which is advantageous for both the provider and the customer [11]. The consumer will profit from recommender systems since the systems will offer suggestions for goods or services. This is advantageous for those who are searching for products or services. Additionally, the company will profit from the rise in sales that typically results from providing customers with more options for goods and services that they are likely to find appealing [12]. The process of deploying services in the cloud has attracted much interest from the business and academic communities because it can significantly lower the overall cost of shareholdings for the supplier of services and the interface client by removing initial expenditure and lowering the expense of maintaining hardware and software. The topic of service computing, however, clearly confronts a significant research challenge in terms of efficient service selection and recommendation as the variety and quantity of services grow. In a cloud computing environment, several factors, such as the size of the infrastructure provider, the location of the information center, marketing plan, material administration plan, the level of popularity, and the user’s location can influence a user’s decision to use a particular service [13]. Cloud computing thus functions as a kind of collaborative cloud filtering that bases conclusions on clusters of user behavior [14]. The research problem addressed in this study is the design and development of an effective cloud services user’s recommendation system that can assist users in selecting appropriate cloud services based on their specific requirements and preferences. Despite the availability of a wide range of cloud services, selecting the service that best meets the user’s needs can be challenging, and users often face difficulties in making informed decisions. The proposed recommendation system aims to address this issue by leveraging machine learning algorithms and user data to provide personalized recommendations that are tailored to the user’s unique requirements. The research will focus on developing a system that can accurately predict the user’s needs based on their usage history and feedback and offer recommendations that match their expectations in terms of performance, cost-effectiveness, and overall user experience. The success of the recommendation system will be measured based on user satisfaction, adoption rate, and the overall impact on the cloud services industry.

The proposed cloud user recommendation system utilizes Random Iterative Fuzzy based Trust Computation (RIFTC) and Support Vector Regression (SVR) to evaluate and enhance the “Quality of Service (QoS)” through personalized and relevant recommendations based on user feedback and behavior.

The remaining sections are structured as follows: Part 2 covers related works. Part 3 details the suggested method, and Part 4 discusses the results. Part 5 offers suggestions for further research and concludes the analysis.

2. Related Works

Mezni and Abdeljaoued [15] suggested deploying a collaborative filtering-based fuzzy formal concept analysis (Fuzzy FCA)-based recommendation system for cloud services. Fuzzy FCA is mathematically sound and is based on the lattice theory. It was suggested that the recommendations could be improved by setting a detailed representation of the cloud environment in the structure model [16] including users, services, ratings, etc. (users who are equivalent to an active user, their ratings of one another, the best services, etc.). Rahhali et al. [17] presented a model of a recommender system for an e-learning platform that will encourage students to choose the courses that best suit their needs. The infrastructure of the system is built on the cloud, notably Google cloud services [18]. Aznoli and Navimipour [19] reviewed cloud service recommendations and provided various suggestions.

Liu and Chen [20] created a unique Clustering-based trust-aware approach. To deliver a more customized service, QoS forecast and trustworthy cloud service proposals for the active user in Cloud Manufacturing (CMfg), they integrate the “clustering-based method” with trust-aware technology. Ding et al. [13] provided a mechanism called time-aware Service Recommendation (taSR). They also created a unique “similarity-enhanced" Collaborative Filtering (CF) method to gather the duration characteristic of client usage similarities and reduce data sparsity in the current Point in Time (PIT) framework. The concept of stochastic frontier incorporated moving estimates will then be used to forecast the QoS values in the ensuing PIT under QoS instantaneity. They assess it by contrasting the suggested course of action with the most recent best practices. Deebak and Al-Turjman [21] proposed a design for community-based recommender systems that are mindful of trust. The preferences of similar users in the target user’s neighborhoods who are dependable neighbors are combined to represent the users’ preferences. User similarity is computed with rating confidence added. Combining the ratings of connected community members yields the prediction for an item that has not yet been rated. According to experimental results on actual data sets, the suggested method performs better in terms of accuracy than alternative complements.

The proposed cloud services user’s recommendation system, which utilizes Random Iterative Fuzzy-Based Trust Computation (RIFTC) and Support Vector Regression (SVR), offers several advantages over other methods for generating personalized recommendations. RIFTC and SVR algorithms help to identify and filter out unreliable or irrelevant data, which leads to more accurate and relevant recommendations.

3. Proposed Methodology

The internet is used by a new type of computer architecture known as the Cloud to distribute resources like processing energy, memory, software, and internet solutions as architecture. The whole user interface is contained within a single window in a Web browser for the majority of cloud computing applications. The goal of various initiatives is to improve the user experience of Internet apps. One tactic is to implement the concept of cloud computing such that a browser has all the functionality of an operating system. Cloud computing is almost fully developed and keeps improving, offering significant possibilities to reduce expenses by implementing innovative options for restructuring and combining the provisioning of internet technology to be more flexible, easier, affordable, open, trustworthy, and safe. CSRS uses QoS and performance score evaluation of services offered by Cloud suppliers to determine the most favorable services from the Cloud environment. In the world of IT, recommender systems are receiving increased attention from researchers (Figure 2).

3.1. Random Iterative Fuzzy Based Trust Computation (RIFTC)

Consider a network of X binary neurons and use $a\in \left\{\mathrm{0,1}\right\}$ to represent the ith neuron’s activity [22]. The vector $a=\left\{a_1,a_{2\dots \dots \dots .}{a}_N\right\}$ specifies the status of the network. Assume that Equation (1) is the stochastic difference equation that predicts that the system will evolve in discrete time.

$$a_i\left(t+1\right)=⊝\left[{\sum }_{j=1}^N{J}_{ij}{\omega }_{ij}\left(t\right)a_j\left(t\right)+{\varphi }_i\left(t\right)-V_i\right]$$

(1)

$⊝$ is the step function in Equation (2):

$$⊝\left(x\right)=\left\{\begin{array}{c} 1 if x>0\\ 0 if x<0\end{array}\right.$$

(2)

The product is that of $j_{it}{\omega }_{it}$, where $j_{it}$ is fixed and ${\omega }_{it}$ is a random variable produced from the time-independent probability distribution $P_{ij}\left({\omega }_{ij}\right)$. Equation (4) is used to represent the connection weights. ${\omega }_{ij}$ is randomly chosen from its distribution at each time step so that

$${〈{\omega }_{ij}\left(t\right)〉}_{\omega }={\overline{\omega }}_{ij}$$

(3)

$${\langle [{\omega }_{ij}\left(t\right)-{\overline{\omega }}_{ij}]\left[{\omega }_{i\prime j\prime }\left(t\prime \right)-\right]\rangle }_{\omega }={\delta }_{ii\prime }{\delta }_{jj\prime }{\delta }_{tt\prime }{\sigma }_{{\omega }_{ij}}^2$$

(4)

where ${\overline{\omega }}_{ij}$ and ${{\sigma }^2}_{{\omega }_{ij}}$ stand for the distribution’s mean and variance, respectively. In a similar vein, ${\varphi }_i\left(t\right)$ is a random external field produced from $P_{ij}\left({\omega }_{ij}\right)$, in Equation (6)

$${〈{\varphi }_i\left(t\right)〉}_{\varnothing }={\overline{\varphi }}_i$$

(5)

$${\langle [{\varphi }_i\left(t\right)-{\overline{\varphi }}_i][{\varphi }_{i\prime }\left(t\prime \right)-{\overline{\varphi }}_{i\prime }]\rangle }_{\varnothing }={\delta }_{i{i}^{\prime }}{\delta }_{t{t}^{\prime }}{\sigma }_{{\varnothing }_i}^2$$

(6)

The deterministic thresholds,$V_I$, are modified randomly by ${\varphi }_I$.

One interpretation of Equation (7) $f_{\alpha i}$ is a random iterative map.

$${\displaystyle f_{{\alpha }_i}}\left(a\right)=⊝[{\sum }_{j=1}{J}_{ij}{\omega }_{ij}{\alpha }_j+{\varphi }_i-V_i]$$

(7)

where ${\alpha }_i$ represents the totality of the connection weights and thresholds of the $i$^th neuron, ${\mu }_i$ adds the probability measure to the map space in Equation (8)

$${=}_i=\left\{f_i|f_i:{\left\{\mathrm{0,1}\right\}}^N\to {\left\{\mathrm{0,1}\right\}}_i\right\}$$

(8)

According to Equations (9) and (10), define a measure $\mu$ on the product space.

$${\mu }_i\left(f_i\right)=\int {\prod }_{j=1}d{\omega }_{ij}{\rho }_{ij}\left({\omega }_{ij}\right)\int d{{\varphi }_i}_i{\rho }_i\left({\varphi }_i\right)\delta \left(f_i-f_{{\alpha }_i}\right)$$

(9)

$$==\left\{f\left|f:{\left\{\mathrm{0,1}\right\}}^N\to {\left\{\mathrm{0,1}\right\}}^N\right.,f=\left(f_1,f_2,\dots ,f_N\right)\right\}$$

(10)

Using the result of the measures ${\mu }_i$ in Equation (11),

$$\mu \left(f\right)=\prod _{i=1}{\mu }_i\left(f_i\right)$$

(11)

the random iterative map might then be used to rewrite Equation (12):

$$a\left(t+1\right)=F^t\left(a\left(t\right)\right)=F^t{F}^{t-1}\dots F^0\left(a\left(t\right)\right)$$

(12)

With $F^t$ being randomly selected from the probability distribution p at each time step. Entropy and other general characteristics of random maps on networks are covered elsewhere. We will refer to a network as a random iterative network if Equation (1), or (12), adequately describes its dynamics.

The suggested framework would use fuzzy rules to determine trust values, with subjective and objective trust values serving as fuzzy inputs and fuzzy outputs, respectively.

I.

Fuzzy inputs have three states with a 0 to 1 range: low, medium, and high.

II.

Fuzzy outputs have a range of 0 to 1 and three states: low, medium, and high.

The following, Equation (13), are some ambiguous rules:

$$\begin{gathered} If\left({GOT}_j^t\left(T\right)=low\right) and {GST}_j^t\left(T\right)=l then TV \left(T\right) =low, \\ If \left({GOT}_j^t\left(T\right) = low\right) and \left({GST}_j^t\left(T\right) = \right.\left.medium\right) then TV \left(T\right) = low, \\ If \left({GOT}_j^t\left(T\right) = medium\right) and \left({GST}_j^t\left(T\right) = medium\right) then TV \left(T\right) = medium, \\ If \left({GOT}_j^t\left(T\right) = ]high\right) and \left({GST}_j^t\left(T\right) = medium\right) then TV \left(T\right) = medium, \end{gathered}$$

(13)

Agent for updating trust values will be performed. The task of the agent upgrading trust values updates the trust repository. Fuzzy rules are used to map three states—trust, distrust, and uncertainty—to generate the trust score. If the calculated trust score is less, the ${CSP}_j$ trust score maps to distrust. The ${CSP}_j$ trust score maps to trust if the calculated trust score is high; if not, it maps to uncertainty.

3.2. Support Vector Regression

Support Vector Regression (SVR) is a supervised machine-learning technique for addressing recurrence issues [23,24]. Regression analysis may be employed to examine the connection between one or more explanatory factors. To train a regression product that translates from intake estimator factors to output observable response parameters, SVR establishes an optimization model [25]. SVR offers certain benefits since it effectively handles high-dimensional data and strikes a balance between model complexity and forecast error. SVR modifies the Support Vector Machine (SVM) classification approach.

SVR manages regression concerns to permit real-valued function estimation while SVM classification generates binary output (for instance, continuous score on a clinical scale) [26,27]. A small to-do classification-defined hyperplane is used by SVM, a sparse kernel machine. As a result, SVR optimization is stated in regard to support factors (a compressed set of training datasets), with the enhancement service relying only on the number of support vectors and not the size of the input data [28].

SVR is superior to other regression techniques in other ways. SVR can successfully handle a nonlinear regression problem since we were able to map the original feature onto a kernel space where data can be discriminated linearly.

Unlike traditional data regression techniques, SVR develops a method to define an element’s significance in establishing the correlation between input and output. For example, a linear regression (such as finite difference analysis) implies assertions about a linear dispersion of source data without giving relevant parameters or a credibility range when the actual connection between inputs and results is complex. SVR, in contrast, is a machine learning approach that seeks to increase prediction reliability by generating a probability range for a variable’s significance to represent the link between inputs and outcomes. Only springs between data cases outside the tube are attached in SVR, and these springs attach to the tube rather than the decision boundary. As a result, the data items inside the tube have no bearing on the final answer (or to be more precise, shifting them slightly doesn’t affect the solution).

For breaching the tube constraint from above and below, they introduce various limitations in Equations (14) and (15).

$$minimize-w,\xi ,\widehat{\xi }\frac{1}{2}{\left|\left|w\right|\right|}^2+\frac{C}{2}\sum _{i=1}({\xi }_i^2+{\xi }_i^2)$$

(14)

$$\mathrm{subjected}\mathrm{to}{w}^T{\varphi }_i+b-y_i\le \epsilon +{\xi }_i\forall i;y_i-w^T{\varphi }_i-b\le \epsilon +{\xi }_i$$

(15)

The basic Lagrangian method develops in Equation (16)

$$\begin{gathered} L_P=\frac{1}{2}{\left|\left|w\right|\right|}^2+\frac{C}{2}{\sum }_{i=1}\left({\xi }_i^2+{\xi }_i^2\right)+{\sum }_{i=1}{\propto }_i\left(w^T{\varphi }_i+b-y_i-\epsilon -{\xi }_i\right)++{\sum }_{i=1}{\widehat{\propto }}_i\left(y_i- \\ {w}^T{\varphi }_i+b-y_i-\epsilon -{\xi }_i\right) \end{gathered}$$

(16)

Remark I: SVM may have included the restrictions that ${\xi }_i\ge 0$ and $\overline{{\xi }_i}\ge 0$. It is obvious, though, that the ultimate solution will already meet that criterion, therefore there is no need to limit the optimization to the ideal outcome. To understand this, consider the case where some ${\xi }_i$ is negative. Setting ${\xi }_i=0$ is preferred in this situation since it is less expensive and does not break any limitations.

They also observe that, based on the foregoing logic, they will always have at least one of the zeros. i.e., both are zero inside the tube and one is zero outside. This indicates that the answer is $\xi .\overline{\xi =0}$ at the solution.

Remark II: SVR sees that, unlike the SVM scenario, they don’t scale = 1. The scope of the problem is now determined by $y_i$, thus we haven’t over-parameterized it. They now take the initiative concerning w, b, $\xi$, and $\overline{\xi }$ to determine the KKT conditions shown below (there are others, of course) in Equations (17) and (18) as follows:

$$w={\sum }_{i=1}\left({\widehat{\propto }}_i-{\propto }_i\right){\varphi }_i$$

(17)

$${\xi }_i={\propto }_i/C\widehat{{\xi }_l=}{\widehat{\propto }}_i/C$$

(18)

Reconnecting this and leveraging the fact that now they also have ${\alpha }_i,\stackrel{-}{{\alpha }_l=0}$ leads to the discovery of the dual problem in Equation (19)

$$\begin{gathered} -\frac{1}{2}\left({\widehat{\propto }}_i-{\propto }_i\right)\left({\widehat{\propto }}_j-{\propto }_j\right)\left(K_{ij}+\frac{1}{C}{\mathsf{\delta }}_{ij}\right)+{\sum }_{i=1}\left({\widehat{\propto }}_i-{\propto }_i\right)y_i-{\sum }_{i=1}\left({\widehat{\propto }}_i-{\propto }_i\right)ϵ \\ \sum _{i=1}\left({\widehat{I\propto }}_i-{\propto }_i\right)y_i=0 \\ {\propto }_i\ge 0{\widehat{\propto }}_i=0\hspace{1em}\forall i \end{gathered}$$

(19)

They can infer the sparsity of the solution from the complimentary slackness conditions in Equations (20)–(22)

$${\propto }_i\left(w^T{\varphi }_i+b-y_i-\epsilon -{\xi }_i\right)=0$$

(20)

$${\widehat{\propto }}_i\left(y_i-w^T{\varphi }_i-b-\epsilon -{\xi }_i\right)=0$$

(21)

$${\xi }_i\widehat{{\xi }_i}=0{\propto }_i{\widehat{\propto }}_i=0$$

(22)

where they manually added the last conditions because they didn’t seem to follow from the formulation. If the constraint $\overline{{\xi }_i}$= $Y_i-W^T{\mathsf{\Phi }}_i-b—$will be satisfied if $\overline{{\xi }_i}=0$, then the case is taking on the smallest value possible. This suggests that I will assume a positive value and that the further $\widehat{{\alpha }_i}$ is from the tube, the larger $\widehat{{\alpha }_i}$ will be (as in a compensating force). Note that I in this instance equals 0. If both ${\xi }_i$ > 0 and ${\alpha }_i$ > 0, the situation is equivalent. When a data case is within the tube, the ${\alpha }_i$ and $\widehat{{\alpha }_i}$ are always 0, which causes sparsity $\left|{\beta }_i\right|.$ Equation (23) is as follows:

$$\begin{gathered} {maximize}_{\beta }-\frac{1}{2}{\sum }_{ij}{\beta }_i{\beta }_j\left(K_{ij}+\frac{1}{C}{\mathsf{\delta }}_{ij}\right)+{\sum }_{i=1}{\beta }_i{y}_i-{\sum }_{i=1}\left|{\beta }_i\right|\in \\ {\sum }_{i=1}{\beta }_i=0 \end{gathered}$$

(23)

Now, SVR alters the variables to resemble the SVM more closely and perform ridge regression in this optimization issue. Introduce ${\beta }_i=\widehat{{\alpha }_i}-{\alpha }_i$ and use $\widehat{{\alpha }_i}{\alpha }_i=0$ to write $\widehat{{\alpha }_i}+{\alpha }_i=\left|{\beta }_i\right|$

where the Equation (24) constraint is caused by the inclusion of a bias term $b$

$$y=w^T\varphi \left(X\right)+b={\sum }_i{\beta }_{i}K\left(x_i,x\right)+b$$

(24)

The slackness conditions also allow us to determine a value for b. (similar to the SVM case). Additionally, as usual, the forecast for new data cases is provided.

Consider the scenario in which the penalty is linear rather than quadratic in Equation (25).

$$\begin{gathered} {Minimize}_{W,\xi \widehat{,\xi }}\frac{1}{2}{\left|\left|w\right|\right|}^2+\frac{C}{2}\sum _{i=1}({\xi }_i^2+{\widehat{\xi }}_i^2) \\ {w}^T{\varphi }_i+b-y_i\le \epsilon +{\xi }_i\forall i \\ {y}_i-w^T{\varphi }_i-b\le \epsilon +{\xi }_i\forall i \\ {\xi }_i\ge 0,{\widehat{\xi }}_i^2\ge 0\hspace{1em}\forall i \end{gathered}$$

(25)

Resulting in the dual issue in Equations (26) and (27)

$${maximize}_{\beta }-\frac{1}{2}{\sum }_{ij}{\beta }_i{\beta }_j\left(K_{ij}+\frac{1}{C}{\mathsf{\delta }}_{ij}\right)+{\sum }_{i=1}{\beta }_i{y}_i-{\sum }_{i=1}\left|{\beta }_i\right|\in$$

(26)

$$\begin{gathered} \mathrm{Subjected}\mathrm{to}{\sum }_i{\beta }_i=0; \\ -C\le {\beta }_i\le +C\hspace{1em}\forall i \end{gathered}$$

(27)

where SVR observes that, as is expected when transitioning from $L_2$ norm to $L_1$ norm, the quadratic penalty on the size of $\beta$ is replaced by a box restriction.

Take into account that the quadratic programs we specified are convex optimization problems having a singular optimum solution that can be readily found using numerical approaches. It is usually said that this represents a considerable improvement over the early stages of neural networks, which were constrained by multiple local optimal solutions.

4. Results

The experiments were conducted on a Hp Pavilion dv6 laptop with an Intel core i7-2630QM processor, 4 GB of RAM and a 750 GB Hitachi hard drive. The first set focused on the computation time of the recommendations, whereas the second set focused on the quality of the generated recommendations.

The dataset consists of 522 movies, 705,309 user reviews, and ratings (from 1 to 5). The suggested movie is set up in a multi-cloud environment with various virtual machines that can access CPU, Memory, and Storage at various times for improved resource efficiency [29]. The MATLAB tool is used for the user recommendation system.

A measurement of errors between paired observations corresponding to the same occurrence is called the mean absolute error. Mean Absolute Error (MAE) is a model evaluation statistic used with regression models. The MAE provides the mean of the absolute difference between the goal value and model predictions. Divided by the absolute difference between all of the observations’ estimated and actual values, the MAE is determined from the array’s total number of observations (Equation (28)).

$$\mathrm{M}\mathrm{A}\mathrm{E}={\sum }_{i=1}^n\left|y_{i-}{x}_i\right|$$

(28)

where N presents the total number of criteria $y_i$ and $x_i$ are the actual and predicted rating values of the $i$ th user for the $j$th service, respectively. MAE will be calculated with each density variation to evaluate the quality (i.e., score) of the recommended services.

Due to the low ratings density of the rating’s formal context the MAE value is gradually reduced. MAE is less than 0.15 in most cases and reaches below 0.11 by the conclusion of the experiment. Results show that, even with a low ratings density, the average quality of the top−10 recommended services is high (0.81 when the ratings formal context density is 10%). Such is the strength of Fuzzy FCA.

Figure 3 and

Table 1. Comparison of Mean Absolute Error with state-of-the-art.

Data Set	TRSC [30]	Cloud Rec [31]	Clustering [18]	QoS Aware Service [32]	RIFTC+SVR
1	2.5	2.2	1.8	1.99	1.5
2	2.6	2.3	1.9	2.1	1.7
3	2.8	2.5	2.1	2.3	1.9
4	2.9	2.67	2.2	2.45	2.1
5	3	2.98	2.35	2.78	2.3

both illustrate how the Mean Absolute Error obtained by the suggested strategy (RIFTC+SVR) was 54% lower than that of the TRSC (99%), Clustering (85%), Cloud Rec (71%), and QoS aware selection (93%), demonstrating how efficiently the suggested system performs compared to current state-of-the-art systems.

A scalar quantity, Throughput, exists. The entire distance traveled by the data in a specific amount of time is its Throughput. Distance traveled by the data is calculated as time divided by distance, which is the Throughput of recommendation. The magnitude serves as its representation, and it lacks direction.

Table 2. Comparison of Throughput (Mbps) with state-of-the-art.

Data Set	TRSC [30]	Cloud Rec [31]	Clustering [18]	QoS Aware Service [32]	RIFTC+SVR
1	5.5	6	9.86	15	16.33
2	6	9.5	11	13.5	19
3	7	7.98	13	12.63	20
4	8.63	11	16.33	14	23.91
5	9	9.8	14.56	11	25.99

shows the results of the throughput testing suggested and the actual tactics and throughput (Mbps). During our analysis of the efficiency results in relation to throughput, we discovered that the throughput of the system increases proportionately with the magnitude of the operation. Figure 4 and Table 2 show that the suggested approach (RIFTC+SVR) outperforms existing approaches TRSC (throughput = 9 Mbps), Clustering (throughput = 9.8 Mbps), Cloud Rec (throughput = 14.56 Mbps) and QoS aware selection (throughput = 11 Mbps) in terms of throughput (25.99 Mbps).

The proximity of two or more measures to one another, regardless of whether they are accurate or not, is referred to as precision. Precision is defined as the attribute of being exact. Precise measurements may be inaccurate.

The measurement precision is the range of the values that were measured. Equation (29), Precision, can be evaluated by determining the range, or difference, between the lowest and highest measured values.

$$Precision=\frac{Total Positive + Total Negative}{Total Positive}$$

(29)

Figure 5 shows how the suggested approach (RIFTC+SVR) produced results with 97% precision. TRSC (42%), Clustering (63%), Cloud Rec (81%), and QoS aware selection (62%), combined, are all higher, providing further evidence of the suggested system’s effective operation (

Table 3. Comparison of Precision with state-of-the-art.

Data Set	TRSC [30]	Cloud Rec [31]	Clustering [18]	QoS Aware Service [32]	RIFTC+SVR
1	63	54	35	15	99
2	82	70	42	27	90
3	25	42	62	32	93
4	71	37	79	42	96
5	42	63	81	62	97

).

Latency, a term frequently used in networking slang to refer to the full round-trip time required for a data packet to transit, is the interval of time between a user action and a web application’s response to that action. The amount of time it takes for a resource to travel between the moments the browser requests it and the moments the browser receives it is one way to gauge latency. Milliseconds, or ms, are widely used to assess latency. The time it takes for a system to reply to queries or requests for data of various volumes and complexity can be used to compare latency with datasets. The findings can then be analyzed to discover any trends or patterns and develop a means to improve performance and reduce latency. This can help to guarantee that an information processing system is capable of handling the requirements of a business or application and giving users quick access to the data they require. Figure 5 demonstrates how the recommended strategy (RIFTC+SVR) resulted in a 51% latency. Clustering (72%), Cloud Rec (98%), TRSC (95%), and QoS aware selection (93%), taken together, are all lower, offering proof of the suggested system’s successful performance (

Table 4. Comparison of Latency with state-of-the-art.

Data Set	TRSC [30]	Cloud Rec [31]	Clustering [18]	QoS Aware Service [32]	RIFTC+SVR
1	71	86	86	99	35
5	95	72	98	93	51
3	61	81	89	95	39
4	86	90	97	94	42
5	95	72	98	93	51

).

Recall is a performance indicator for binary classification models that quantifies the percentage of correctly detected positive cases. The recall estimation of a machine learning model may be impacted by the network payload, which is the volume of data carried over a network. To evaluate the model’s capacity to locate optimistic samples, we utilize the recall measure. The recall is computed by dividing the total number of favorable samples by the percentage of significant samples that were properly identified as positive. This ratio is then multiplied by 100. Figure 6 shows that the suggested approach (RIFTC+SVR) produced a 97% recall rate. When combined, Clustering (76%) is lower than Cloud Rec (68%), TRSC (89%), and QoS aware selection (73%). It provides evidence of the successful operation of the suggested system (

Table 5. Comparison of recall with state-of-the art.

Data Set	TRSC [30]	Cloud Rec [31]	Clustering [18]	QoS Aware Service [32]	RIFTC+SVR
1	63	54	35	15	99
2	82	70	42	27	90
3	25	42	62	32	93
4	71	37	79	42	96
5	42	63	81	62	97

, Figure 7).

5. Discussion

In order to generate individualized recommendations, RIFBTC is a machine learning-based method that blends fuzzy logic and trust computation. It iteratively improves its recommendations using user data and feedback, and user trust levels are calculated using a fuzzy-based algorithm. Subsequently, the technique uses Support Vector Regression (SVR) to forecast the user’s rating of a certain service based on their preferences and usage patterns. Nevertheless, FFCA is a formal concept analysis-based methodology that uses mathematical concepts to find connections between consumers and services. It creates recommendations by identifying the most pertinent services that match the user’s preferences and uses formal concepts to represent the user preferences and services. The approach then uses fuzzy logic to determine how relevant the service is to the user. While both approaches have merits and disadvantages, RIFBTC has some benefits over FFCA. More freedom in modeling user preferences is provided by the fuzzy logic-based trust computation method used by RIFBTC, and the iterative design process enables continuous improvement of the recommendation system. Moreover, the use of SVR enables more precise service rating prediction. Yet, the formal concept analysis-based methodology of FFCA enables a more comprehensible and open recommendation system. Mathematical concepts can be used to identify pertinent services in a more precise and organized manner. While RIFBTC’s SVR-based rating prediction is more precise than FFCA’s fuzzy logic-based degree of relevance computation, the latter may be less flexible in representing complicated user preferences. Ultimately, the choice of method will rely on the particular requirements and objectives of the recommendation system, although both RIFBTC and FFCA are workable approaches for cloud services recommendation systems. While FFCA may be more suitable for systems that prioritize interpretability and transparency, RIFBTC may be better suited for systems that prioritize accuracy and flexibility.

6. Conclusions

Cloud computing is a new computer architecture that has developed in the current digital age from distributed systems and service-oriented design. As most cloud operations are expressed in plain language, especially web services, they cannot be chosen and have no uniform description. In this paper, the design of recommender systems frequently makes use of fuzzy logic to manage the ambiguity, imprecision, and vagueness of item features and user behavior. Product suggestions are a widely used function in social networks and e-commerce platforms at present. When extreme precision is required, the task becomes more urgent. Due to this problem, the customer wants a service that is reliable, appropriate, time-saving, and trustworthy. We suggest a recommendation system for cloud service users that uses Random Iterative Fuzzy Based Trust Computation (RIFTC), which focuses on estimating trust using Quality of Service (QoS) characteristics. Support Vector Regression (SVR), a Machine Learning (ML) approach, is used by the RIFTC to calculate trust. For a cloud-based user recommendation system, the trial results are evaluated using precision, latency, throughput, and mean absolute error. The experimental findings demonstrate that our suggested system performs better than the indicated system. The RIFTC and SVR would offer a less complicated, more adaptable system that could accommodate many business applications. However, the current fuzzy system is still in development and has similar evaluator biases constraints. To increase the effectiveness of our recommendation system, we plan to examine cloud service capabilities and incorporate dynamic trust into further work.