Personalized Product Service Scheme Recommendation Based on Trust and Cloud Model

Considering the low satisfaction rate and low efficiency of product service plans, a personalized product service plan recommendation method adopting the degree of trust and cloud model is proposed. The recommendation algorithm mainly includes calculating the similarity and the prediction results of the scheme. First, by fully considering the user’s subjective characteristics, the user’s trust is used to improve the traditional similarity. Second, considering data sparsity and discreteness, the cloud drop distance similarity calculation method is introduced in the process of calculating the trust similarity, and a new similarity is generated via the weighted synthesis to predict and fill in the gaps in the data. Then, when the new user does not have the cold start problem caused by the historical score record, the neural network method can be used to classify the users based on the user characteristics. The method is proposed and introduced to predict sparse user interest features and obtain similar user sets based on feature classification. The corresponding program offers recommendations. Finally, the effectiveness and rationality of the proposed method are verified by using the recommendations for a machine tool product for a manufacturing enterprise as an example.


I. INTRODUCTION
With the increasingly fierce competition in the market and the gradual increase of users' individualized demands, the manufacturing industry has begun to shift from manufacturing products to the ''product + service'' operation mode. The proportion of product services in the whole product life cycle is increasing, which reflects the core competitiveness of the products to a large extent. The product service system (PSS) has gained development opportunities in the manufacturing industry. The PSS provides an integrated solution for the integration of products and services [1]. Through information technology, user needs, product design, product production and product services are integrated organically. Because the user demand is often vague, uncertain and incomplete in the early stage, the deviation between the designer and the user when understanding the demand and designing the PSS will lead to the problems of a low satisfaction rate with and low The associate editor coordinating the review of this manuscript and approving it for publication was Ming Luo . efficiency of the final product and service plan [2]. Therefore, to reduce the impact of the subjective understanding deviations of users and designers on the results of the scheme generation, this paper proposes to clean the metadata and then use the cleaned data to generate the scheme. In addition, how to quickly and efficiently recommend a satisfactory service plan has become a difficult problem for enterprises and the recommendation method has been widely used and continuously improved in many fields. This paper proposes to apply it to the product service plan generation and recommendation process. Recommendation methods generally have data sparsity problems caused by insufficient user rating data and cold start problems caused by unrated users. Considering the subjectivity of users in the recommendation process and the inaccurate recommendation results caused by the discreteness and sparseness of user data, this paper proposes to introduce the trust among users in the process of calculating user similarity and the cloud droplet distance similarity to mitigate the impacts of the data sparseness and discreteness problems. The method fused the comprehensive values of the user trust and cloud droplet distance similarity to predict the missing values of scheme scores. For the cold start problem caused by the lack of historical records for new users, to make reasonable recommendations, a neural network method is proposed to classify new users based on their similar interests to obtain appropriate recommendation schemes.

II. RELATED WORK
Currently, the recommendation technologies can be divided into content-based recommendations, collaborative filteringbased recommendations, and domain knowledge-based recommendations. In the process of using a recommendation method, there are some problems such as data sparsity and cold starts. Data sparsity generally refers to the user's project rating being lower or insufficient. Cold start refers to when there is no scoring record due to new users in the historical database. For example, based on the social network among users, document [3] used the context aware Point-of-Interest (POI) recommendation model based on matrix decomposition and used the social network data of users and the POI category information to establish a general matrix decomposition model to improve the accuracy of recommendation results. Documents [4], [5] used the random walk method to travel a social network. According to the distance between users in the travel process, it predicted the similarity between users and filled in the missing values in a matrix, which fixed the data sparsity. In view of the fact that the matrix decomposition method does not consider independent users and projects and other factors, a nonnegative matrix decomposition algorithm based on user and project biases was proposed by introducing user and project biases in document [6]. Considering the local optimal solution problem that may be generated by the random initialization matrix, an improved method using singular value decomposition (SVD) was proposed to reduce the dimension of the data. In document [7], a factor decomposition machine based on Lasso was proposed to control the data dimension of factor decomposition, which was used to process high-dimensional and sparse data and reduce the impact of data sparsity. In terms of data sparsity, the hidden factors between matrices were used to predict or reduce the dimension of the sparse data to reduce the impact of missing data on the recommendation results. In document [8], a fuzzy clustering method was used to calculate the similarity between users and obtain the nearest neighbor set to fill the gaps in the data. In document [9], considering the influence of different numbers of ratings and the asymmetric relationship between projects, a suitable neighbor set was selected to fill in the missing data based on the probability distribution. In document [10], based on the influence of subjective scoring on the recommendation results, the Pearson similarity was introduced to improve the accuracy of the user attribute similarity calculation and the K-means clustering algorithm and the collaborative filtering algorithm were integrated to reduce the influence of data sparsity. In document [11], the probability matrix decomposition algorithm was used to measure the relationship between users or products and obtain similar neighboring users or products to achieve accurate recommendations. In documents [8]- [11], clustering, the probability distribution, the similarity and other calculations were used to obtain the nearest neighbor set, which was used to fill in the missing data. As one of the most mature methods, collaborative filtering was improved by combining it with many methods, such as the BP neural network, the conditional probability, the bipartite graph, the clustering algorithm, the genetic algorithm, etc. in documents [12]- [16] to reduce the impact of data sparsity on recommendations. Considering the influence of subjective user ratings on recommendation results, document [17] took into account that the Quality of Service (QoS) incorporated the contribution of unreliable users and used a clustering algorithm and the collaborative filtering method integrating the degree of trust to rebuild the trust network of clustered users and make personalized QoS predictions and cloud service recommendations for active users. In document [18], users had subjective preferences for various project categories, which meant that the deviation of the original data may be amplified or reversed by the potential recommendation algorithms. Appropriate recommendation methods were selected by comparing different algorithms reflecting the sorting quality and deviation. In this paper, the recommendation method was applied to the generation process of a product service scheme. Aiming at the data sparsity problem occurring in the early stage, the trust relationship of user subjectivity was introduced as a constraint, which meant that the degree of trust was introduced into the calculation of the user similarity and the results of similar calculations and predictions were constrained. In documents [19]- [20], considering the potential interest preferences between new users and new projects, a clustering method was used to obtain the nearest neighbor set of new users and the projects of the nearest neighbor set were recommended to new users. In document [21], to solve the cold start problem, the interaction information between the attribute preference information of the project and the user's rating behavior was not fully utilized. The singular value decomposition recommendation algorithm based on the attribute preferences for the project was introduced to obtain the user's preference matrix for the project attributes, and the attribute characteristic factors of the project attributes and the user's preferred characteristic factors of the project attributes were added to the matrix decomposition. Document [22] proposed a cold start recommendation model based on voice personality characteristics to obtain the user's personalized information. It collected the user's personality characteristics through any normal voice of the new user and provided the user with personalized recommendations by combining the cross domain personality knowledge and recommendation system. In document [23], an improved association rule algorithm based on the FP Tree was proposed to measure and model user preferences in a personalized information recommendation service to mine user behavioral patterns.
In document [24], the time interval between the user evaluation time and the project release time was introduced and the time impact perception similarity measure was used as the user time weight value to obtain the user's preference for new projects and give recommendations. In document [25], considering that traditional similarity models and QoS prediction methods rarely use time information, time information was integrated into the similarity measurement to obtain the final QoS prediction and recommendation. The above documents introduced personalized features and time features to alleviate the cold start problem caused by new users. Meanwhile, document [26] helped users to find valuable information based on location-based services and the characteristics of a locationbased service platform by focusing on the low recommendation efficiency of less active areas or new users and the inability to maximize the information problems due to hidden data. This paper proposed a recommendation algorithm that combined the decomposition of a collaborative probability matrix and the iterative decision tree to alleviate the cold start problem of new users. In document [27], the RaPare method was used to explore the differences between cold start and hot start users/items, fine-grained calibration was carried out for the potential profile of cold start users/items and the method to solve the cold start problem was obtained and recommended. In the implementation process, the recommendation method was affected not only by the data sparsity and cold start problems but also by other factors, such as the time characteristics and the similarity calculation formula. Considering that the improved collaborative filtering method had been widely used in e-commerce recommendation systems, manufacturing product service recommendations also began to use neural networks to solve the cold start problem of new users. For example, in document [28], the collaborative filtering algorithm based on the community similarity performed poorly in the cold start scenario of new users and new products and the matrix decomposition method was used to find the potential interest space of users. Then, a neural network was trained to learn the mapping relationship from user attribute data to potential interest vectors, and historical data and attribute data were combined to form the user's predicted values of goods and give appropriate recommendations. Document [29] proposed a hybrid QoS recommendation model based on user situational awareness in a cloud environment. Aiming at the cold start problem of new users and new projects in traditional QoS recommendation methods, the missing QoS attribute preference values were predicted and recommended based on user and item data using collaborative filtering. In document [30], [31], aiming at the data sparsity of the multi criteria decision collaborative filtering and recommendation system in e-commerce, a multi criteria decision collaborative filtering and recommendation system based on a neural network was proposed. In it, an adaptive neural fuzzy inference system was used to predict the relationship between the total score of options and each criterion score of options. The similarity measurement mechanism was used to alleviate the data sparsity and improve the robustness and reliability of the similarity measurement for cold start users. Because the neural network provided a breakthrough in the process of solving the cold start problem, based on the preferences of new user characteristics, this paper proposed to use an activation function to classify user characteristic preferences and select the scheme corresponding to the nearest neighbor set of a new user to provide a recommendation.
In addition to being widely used in e-commerce, recommendation methods were also applied in imaging, product services and other fields. For example, in document [32], since the traditional image feature representation cannot capture product styles, highly related style feature modeling was introduced into the visual recommendation model and the style feature was integrated into collaborative learning to incorporate user preferences and improve the validity of the sensory recommendation method. In document [33], a Hierarchical Dirichlet Processes model was proposed based on service description text and service tag information. The personalized PageRank algorithm based on the service tags was used to rank and recommend cloud services in each cluster to alleviate the difficulty of personalized cloud services.

III. DESIGN OF A PERSONALIZED PRODUCT SERVICE RECOMMENDATION SCHEME A. FRAMEWORK FOR A PERSONALIZED PRODUCT SERVICE RECOMMENDATION SCHEME
As an important recommendation technology, the collaborative filtering algorithm has been widely used in various fields and achieved good results. Aiming at the data sparsity and cold start problems caused by new users during the implementation of collaborative algorithms, this paper proposes a recommendation method based on trust and a cloud model to improve similarity. The personalized product service proposal recommendation framework is shown in figure 1. First, according to the influence of a user's direct subjective comments on the recommendation results, the weighted sum of the direct trust and indirect trust is proposed to measure the similarity of a user's trust Second, considering that the cloud model can realize the qualitative and quantitative transformation of data, three characteristics of the cloud model are proposed and the forward cloud method is used to obtain several cloud drops to find the degree of cloud drop distance similarity. Finally, the weighted sum of the trust similarity and cloud droplet distance similarity are used to predict and fill in the sparse data to reduce the data sparsity. Aiming at the cold start problem caused by new users, the neural network is used to predict the nearest neighbor set to which new users belonged, and the nearest neighbor set is recommended.  droplet dispersion but also helps to solve fuzzy and uncertain mathematical problems by instituting the mutual mapping process between qualitative and quantitative relationships. In this paper, the cloud similarity is used to calculate the score similarity of schemes to predict the scores of schemes. The concepts of the cloud and cloud droplets and the calculation methods of the related indicators are presented as follows.
Definition 1 (Cloud and Droplet): Let U be a quantitative domain expressed numerically and C be a qualitative concept on U. If the quantitative value x ∈ U is a random realization of the qualitative concept C, the determinacy of x to C: µ (x) ∈ [0, 1] is a random number with a stable tendency µ : U → [0, 1], ∀x ∈ U , and x → µ (x). The distribution of x on the domain U becomes a cloud, which is denoted as C (x), and each x becomes a cloud drop.
The conceptual features contained in the cloud model can be expressed by three digital features of the cloud, including the Expected value (Ex), the Entropy (En) and the Hyper Entropy (He), as shown in (1) [34].
Among them, the variable N represents the number of cloud droplets and xi represents cloud droplet i.
Definition 2: The random variable x satisfies x ∼ N (E x , E 2 n ) and E n ∼ N (E n , He 2 ) if the determinacy of the qualitative concept C satisfies formula (2).
Considering that users' subjective evaluations have the characteristics of vagueness and randomness, the subjective experience is used to divide the data into several levels of trust. The qualitative and quantitative relationship transformation can be achieved using the cloud model. A cloud model is proposed to convert the quantitative data of user plans into cloud drops to form a cloud of trust.
User trust refers to the degree of trust of one user to another user, which is mainly divided into the degree of direct trust and the degree of indirect trust. The degree of direct trust indicates the direct trust relationship between user u and user v. Furthermore, the degree of indirect trust indicates the many indirect connections between user u and user v that can establish the degree of user trust. The degree of trust is expressed by the relationships between users on social networks, and the strength of trust is represented by Trust. dTrust means direct trust. For any users u i and u j , there is a direct connection between them. For example, u i recommends something to u j and succeeds, which means that the first user directly trusts the second. dTrust means indirect trust. For u i and u j , there are other users in the middle that allow the two to be connected, which is indirect user trust. The direct trust calculation method is shown in formula (3) [35].
In formula (3), I (u j ) represents the score item set of user u i , and I (u j ) represents the score item set of user u j . dTrust represents the score item set of users u i and u j , which measures the proportion of the total number of user cumulative scores.
An indirect trust relationship is expressed by mTrust, which is the continuous connection between user u t and user u t+1 , including the indirect relationship between u i and u j that form an indirect relationship. The equation for calculating this is shown in formula (4).
Considering that the trust propagation channel is long and the trust is declining in the indirect user transmission process; this paper stipulates that the information transmission between two users is limited to three users. The two user transmission paths are not limited to being one way. Therefore, the overall indirect trust of a user is calculated as shown in formula (5).
The comprehensive degree of trust is the weighted comprehensive value of the degree of direct trust and the degree of indirect trust, which acts as the trust similarity between users, as shown in formula (6).
l(u i u j ) is the weight coefficients of users u i and u j 's degree of direct trust. g(u i u j ) represents the trust probability of user u j in the social network where user u i is located. Take the social network in figure 2 as an example. When the user indegree between user u i and user u j is s ij , the average trust indegree in the social network is avg. When s ij > avg, g(u i u j ) is set to 1, and vice versa; and g(u i u j ) = s ij /avg.
Universe U can be expressed as one-dimensional and multidimensional forms. Cloud droplets are random variables whose degree of membership is realized in the whole space that realizes the qualitative and quantitative transformation. The relevant concepts of the cloud characteristics are also reflected. Cloud droplets that are not simply fuzzy and random are random variables generated by membership constraints. The degree of qualitative concepts of cloud droplets can be represented by the degree of membership. In other words, the concept of cloud droplet appearance is proportional to the degree of cloud droplet determination. The cloud model processing includes the forward cloud and reverse cloud generators. In this paper, the forward cloud algorithm is used to transform the qualitative vector feature C(Ex, En, He) into cloud droplets describing individual quantitative values. With three eigenvalues of the cloud (0.5, 0.05, 0.01) and n = 1000 as the input values, the cloud image shown in figure 3 is output by the forward cloud generator. The expected value Ex, which is the point most representative of the qualitative concept, represents the expected value of the cloud droplet distribution in the universe. En represents the uncertainty of the qualitative concept and reflects the degree of cloud dispersion. He is the degree of uncertainty of the entropy, which describes the thickness of the cloud.
In this paper, three eigenvalues of the cloud are transformed into cloud droplets by the forward cloud method. In addition, the distance between two clouds is obtained using the membership degree as the distance similarity of cloud droplets. The distance similarity of cloud droplets is shown in formula (7).
Because the user's subjective evaluation is the eigenvalue of the cloud model, it has certain subjective uncertainty.
To solve this problem, a forward cloud method is proposed to transform the eigenvalues into cloud droplets, and the randomness of the cloud droplets is used to reduce the influence of subjective eigenvalues. First, the cloud feature of the existing user's rating is obtained. However, considering that the user's rating includes quantitative and qualitative values, it is proposed that the qualitative value be converted into a quantitative value according to certain criteria, and then the user's or project's rating be converted into a cloud feature vector of the user's overall rating by using the reverse cloud algorithm. The specific calculation process is as follows.
Step 1: For N cloud droplets {x 1 , x 2 , . . . , x N }, the sample meanx and the sample mean S2 shown in formula (1) are calculated according to the following formula.
Step 2: Formula (1) is used to calculate the three cloud features to get E x and E n .
Step 3: The forward cloud method is used to transform the cloud eigenvalues into several cloud droplets, each of which is one cloud droplet in the number domain with a degree of certainty.
Step 5: The distance similarity is used to solve the cloud model similarity. VOLUME 8, 2020 Users with similar features also have high similarities among features. Considering the subjective factor of user trust and the objective factor of user ratings, the similarity and prediction formulas are synthesized to fill in the sparse rating data. The formulas are shown in (8) and (9), respectively.
In the formulas, R v denotes user u v 's score for a scheme. sim(uv) denotes users u and v's similarity. R u,i is user u u 's score for scheme i. N (i) is the nearest neighbor set of user u v and user u v 's predictive score for scheme i is P v,i .

C. NEW USER RECOMMENDATION BASED ON ANEURAL NETWORK
A neural network is a parallel and distributed network structure that consists of a large number of neurons. There are many connection paths among input data. Each connection path corresponds to a connection coefficient. The transmission of enhanced or suppressed signals between interconnected neurons is realized by adjusting the interconnected weight coefficients. The output of the regulation mechanism converges to the correct target value. The BP neural network algorithm can be divided into two stages: forward signal propagation and error back propagation. First, information is input from the input layer and passed through the hidden layer to the output layer. The connection weight remains unchanged in the transmission process. In addition, there is a difference between the expected results and the output results. Meanwhile, the difference is returned and the weights between neurons are adjusted according to the error during the return process until the standard deviation between the actual output and the expected output is at its minimum. The recommendation algorithm can recommend corresponding schemes to existing users according to their historical behavioral characteristics and interest preferences. However, it is difficult to recommend an accurate scheme to new users without any historical ratings. A neural network can adjust the user similarity in a recommendation algorithm and add other scheme attributes to optimize the training process. A neural network method based on user interest similarity is proposed to alleviate the impact of the cold start problems of new users on the recommendation results. The data in this article include the users' rating of the scheme, which is from 1 to 5 and obtained from the questionnaire. The grade of the feature of the scheme is from 1-5, and the score of the new and old users' interest features is from 1 to 3 Then, we obtain the similarity scheme of the new users' interest for recommendations by calculating the activation function. The neural network used in this paper had a three-layer structure, including an input layer, a hidden layer, and an output layer. First, it takes the user-scheme scores in Table 4 as the input of the neural network The user's subjective interest similarity is adjusted to adjust the weight of the network layer to obtain the data of the hidden layer, and then the sigmoid function value of the scheme-feature corresponding score is taken as the hidden layer's input weights Finally, it outputs the user matching scheme score values under different neighbor sets. A plan with a Top-N rating is chosen for new users. The relevant training steps of the neural network are as follows.
Input: Neighbor user set score of the scheme Output: New users' ratings for different schemes Step 1: Set the number of input and output layer nodes, and initialize the weight W and threshold Q of each layer.
Step 2: Input the score value of the nearest user scheme and the user interest feature rating into the neural network.
Step 3: The neural network is used to train the hidden layer and the output layer, and the scheme score and scheme feature values of the new user neighbor set are respectively used as the input and weight of the neural network algorithm. The weight of the hidden layer is the value of the sigmoid function obtained from the eigenvalues of the existing scheme.
Step 4: Determine whether the error of the sample set met the requirement, and finish learning if it met the requirement; otherwise, continue learning until it meets the requirement.
Step 5: Using the above trained network, the similarity of a new user scheme classification is obtained, and the score of a new user scheme is output.
Let us take a specific example for illustration. Assume that the user {u 1 , u 2 , u 3 , . . . , u m } has the same solution A = {a 1 , a 2 , a 3 , . . . , a t }, and the user 's rating of the solution is R = {r 11 , r 21 , r 31 , . . . , r m1 }. The similarities of the interests between the new user and the existing user are {sim 1 , sim 2 , sim 3 , . . . . sim n } and are used as the weight of the input layer. The hidden layer's weight is the value of the scheme feature f ={f 11 , f 12 , f 13 , . . . , f 21 , . . . , f m1 . . .}, and the output value is the corresponding scheme score value of each neighbor user of the new user. The activation function in the neural network process is the sigmoid function (x) = 1 1 + e −x , where the hidden layer of the input value of x is expressed as the feature value of the scheme, and the user's score for each scheme is the output. The scheme with the highest value is selected as the new user. The TOP-N scheme is selected for recommendation based on the scores of each plan, as shown in Figure 4.

IV. CASE ANALYSIS A. CASE BACKGROUND
With China's manufacturing 2025, industry 4 and Internet + put forward policies, increasing more manufacturing and  service industries are also advancing with the times. In the product and solution design and generation processes, personalized recommendation methods have been introduced to make corresponding recommendations to improve the efficiency of enterprises. Taking the machine tool product of a manufacturing enterprise as an example, the recommendation method proposed in this paper is used to make a suitable scheme for users. The machine tool indicators are the nominal pressure f 1 , the table length f 2 , the throat depth f 3 , the slider stroke f 4 , the maximum opening height f 5 , and the main motor power f 6 The nominal pressure unit is KN; the length, depth, stroke and height unit is millimeters (mm); main motor power is in kw; and the index value is quantified as 1-7. The specific parameters are shown in Table 1 According to the existing parameter indicators of the enterprise and the composition of different types of products, different product solutions are formed as A = {a 1 , a 2 , a 3 , . . . , a n }, as shown in Table 2 B. RECOMMENDATION PROCESS OF A PERSONALIZED PRODUCT SERVICE SCHEME In view of the fact that new users did not have any historical records in the original database, the user interest similarity is adopted in this paper to obtain a new user's nearest neighbor set whose corresponding scheme can be recommended. First, the questionnaire is used to obtain the scores of the existing users and new user u 9 on the user's interest characteristics, which are shown in Table 3 The user's interest features mainly include the following: strong adaptability to the processing object p 1 , high processing accuracy p 2 , high production efficiency p 3 , high automation p 4 , and high reliability p 5 . These features are rated as low, medium and high from 1 to 3, respectively.
The Pearson similarity is used to calculate the similarity between user u 9 and the other users to obtain the neighbor set. The similarities between u 9 and u 1 − u 8 are 0.598, 0.896, 0.423, 0.327, 0.896, 0.134, 0.873, and 0.071, respectively. According to the different thresholds, u 2 , u 5 and u 7 are classified as high; u 1 and u 3 are classified as medium; and u 4 , u 6 and u 8 are classified as low. The similarity of neighboring users is taken as the weight of the input layer, and the input value is the user scheme score. The score is expressed as 1-5 from low to high, respectively. The users' scores on the first 13 programs are obtained through questionnaires, as shown in Table 4.
Before giving a reasonable recommendation to new users, the sparse rating data of users affect the recommendation results of new users. Before that, the cloud droplet distance similarity and user trust similarity proposed in this paper are used to reasonably predict user ratings. First, the cloud droplet distance similarity of the user rating is calculated, which means that three cloud eigenvalues of each user rating are calculated. As shown in Table 5, the forward cloud generator is used to convert the features into N cloud droplets, where N = 1000, and the distance between cloud droplets and the distance between trust clouds is calculated.   According to the social network graph of users, the trust degree among users is obtained. The graph of the known social network is shown in figure 5 below. Taking user u 1 as an example, u 1 is directly related to u 2 and u 8 and indirectly related to u 6 and u 4. Formulas (3) -(6) are used to calculate the trust similarity. We know that u 1 and u 8 , u 2 and u 4 , u 3 and u 5 , u 4 and u 2 , u 5 and u 7 , u 6 and u 4 , u 7 and u 5 , and u 8 and u 5 are highly similar. Combining the cloud droplet similarity of u 1 and u 2 in Table 5, the comprehensive similarities of 0.59, 0.64, 0.63, 0.40, 0.63, 0.64, and 0.57 can be obtained. Prediction formula (9) is used to fill in the sparse data in Table 3, and the weight λ is set as 0.3. The filled in results are represented by the bold figures in Table 5.
After solving the data sparsity problem, a scheme recommendation method based on a neural network is proposed to solve the cold start problem of new users. First, the user-scheme score in Table 3 is used as the input of the neural network. In addition, the user interest feature similarity is used as the weight to get the data of the hidden layer. Second, the scheme-feature correspondence score is used as the weight input of the hidden layer. Finally, the scores of the user matching schemes under different neighborhood sets are output.
Step 2: Calculate the new user's program score under a high threshold, that is, multiply the similarity by the neighboring user's program score, as shown in Table 6. Step 3: The filtered users in Table 6 use the sigmoid function to calculate the product of the feature score of the solution and each corresponding solution as the criterion to judge and classify the new user and obtain the output value of the nearest neighbor set of solutions. The highest score in the plan is recommended, as shown in Table 7. For example, the high threshold u 7 corresponds to scheme a 1 ; u 5 corresponds to solutions a 5 , a 7 , and a 9 ; u 1 corresponds to solutions a 4 , a 10 , and a 13 ; and u 2 corresponds to solutions a 3 , a 6 , a 8 , a 11 and a 12 for new user u 9 . The scores of a 1 and a 3 to a 13

C. COMPARATIVE ANALYSES
In this paper, the similarity calculation combined with the trust and cloud model method is used to predict and fill in the ungraded missing data, which can make up for inaccurate recommendations caused by data sparsity. Considering that new users have no historical ratings, a neural network method is proposed to predict and classify the new users based on their similar interests to obtain the recommendation scheme corresponding to the nearest neighbor user set. To prove the efficiency of this method, data sets with 200-1200 entries are randomly obtained from the data set of the enterprise. Using a platform with windows 10 and python 3.0, the method of this paper, and the recommendation method based on the cloud model, we randomly obtain different data sets ranging from 200-1200 entries from the enterprise's data set. The recommendation method based on trust and user characteristics compares the recommendation accuracy and algorithm execution time. The experimental results are shown in figures 6 and 7, respectively. As shown in Figure 6 above, the recommendation accuracy of this paper increased as the size of the data set increases, which showed that the recommendation accuracy of this method and the recommendation accuracy of the trust-based method are improving. When the data set is small, the result is slightly better than the recommendation of this algorithm. When the data set is small, the user's subjective recommendation effect is better. This can be explained as the new user being more willing to believe the old user's push recommendation scheme. However, as the size of the data set increases, the accuracy of this method is higher. As the size of the data set increases, the cloud model-based recommendation can transform the qualitative evaluation into a quantitative evaluation; therefore, the recommendation accuracy is improving, but it is still worse than the other two methods in general. In conclusion, the recommendation accuracy of this method is better than those of the other two methods. The subjective interest characteristics of users and the cloud model method are integrated so that the recommendation accuracy is better here than elsewhere and the recommendation effect is verified.
As shown in figure 7, as the size of the data set increases, the different recommendation execution times increase. With the increase in the size of the data set, the execution time of this method approaches those of the other methods. With the increase in the size of the data set, the implementation times of the three methods based on the cloud model are the least, and the implementation method in this paper has a slightly higher implementation time than that of the cloud model. Due to the combination of the cloud model and user trust, the implementation time of the recommendation method based on trust is the longest because of considering more subjective characteristics and the complicated similarity calculation process. This method is proposed based on the cloud model, the degree of trust and the neural network prediction. Therefore, the execution time of this method is a little longer, but as the size of the data set increases, the gap between the two methods is narrowed.

V. CONCLUSION
The maturity of Internet technology and the growth of the e-commerce industry require increasingly more information technology. As a common method in the field of information, recommendation technology is increasingly more important in improving the efficiency of enterprises. Recommendation technology is immature in the manufacturing industry. To improve the satisfaction rate and efficiency of users for product service schemes, recommendation technology is introduced in this paper. An optimization method is proposed to improve the data sparsity and cold start of new users. The effectiveness and rationality of the proposed method are verified by case analysis and an algorithm comparison. The main contributions of this paper are as follows.
(1) Aiming at the data sparsity problem of user schemes, a similarity calculation method is proposed to predict the user's scores for the schemes. Considering the subjectivity of user ratings, a comprehensive similarity between user trust and the cloud drop distance is proposed to reduce the impact of sparse data on recommendation results.
(2) Considering the cold start problem caused by new users, a neural network method based on user feature scores is proposed to output a classified neighbor set with similar user characteristics. Highly similar users are selected according to the threshold and their corresponding solutions are recommended.
Since the output and threshold judgment of the neural network will consume computing time, future work will consider the execution time of the optimization recommendation method.
XIAOMING DU received the bachelor's and master's degrees from the Jiangsu University of Science and Technology, in 2007 and 2011, respectively, where he is currently pursuing the Ph.D. degree. He has published several articles in international journals and conferences, such as the China Science paper Science Technology and Engineering Research and Exploration in Laboratory. His research interests include data mining and information management.
SHI-LUN GE received the Ph.D. degree in management science and engineering from the Nanjing University of Technology, Nanjing, China, in 2003. He is currently a Professor and a Ph.D. Supervisor with the Jiangsu University of Science and Technology. He has presided over and completed three projects on the surface of the National Natural Science Foundation of China, more than 20 projects commissioned by provinces, ministries, and enterprises. He has been engaged in the teaching and research of management information systems, enterprise information model, and manufacturing informatization for a long time.
NIAN-XIN WANG received the Ph.D. degree in management science and engineering from Southeast University, Nanjing, China, in 2009. He is currently a Professor and a Ph.D. Supervisor with the Jiangsu University of Science and Technology. He has published several articles in international journals and conferences, such as journal of management information systems communications of the association for information systems expert systems with applications management decision. He is a member of the International Information Systems Association.
ZHEN YANG received the master's degree from the Shanghai University of Technology, in 2016. During her master's degree, she has published personally recommended related articles in computer integrated manufacturing systems, Chinese mechanical engineering, computer engineering, and applications. Her research interests include data mining and personalized recommendations. VOLUME 8, 2020