A method and system for manufacturing cloud service recommendation generation

Abstract

The application provides a manufacturing cloud service recommendation generation method, comprising: obtaining manufacturing service descriptive information and label information of a cloud platform; using an LDA topic model to vectorize and model the manufacturing service descriptive information, obtaining p topic information of the manufacturing service, and obtaining p*q-dimensional service topic word vectors through Word2vec mapping expansion; inputting all the word vectors into CNN to learn the deep features of the manufacturing service; further calculating the manufacturing service similarity matrix through the deep feature similarity; using an improved K-Means clustering to cluster the manufacturing service label information to obtain the manufacturing service category, further obtaining a user category matrix model and a service category matrix model, and obtaining a coefficient to construct a user preference matrix; and combining the manufacturing service similarity matrix and the user preference matrix to obtain a manufacturing service recommendation list. The application solves the problems of incomplete manufacturing service information feature extraction and sparse user service matrix in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloud manufacturing, and particularly relates to a manufacturing cloud service recommendation generation method and system. BACKGROUND

[0002] Cloud manufacturing is a new intelligent manufacturing mode that uses ubiquitous network and cloud manufacturing service platform to organize online manufacturing resources according to user demand and provides various on-demand manufacturing services. Manufacturing cloud service (referred to as cloud service) is a basic element of a cloud manufacturing platform, and is a virtualized and serviced manufacturing resource and manufacturing capability on the cloud manufacturing platform. Manufacturing cloud service encapsulates and accesses cloud platforms through virtualization, Internet of Things and other technologies, forms a manufacturing cloud, and through the manufacturing cloud, users can obtain personalized services as needed, quickly respond to changing market demand at low cost, so that enterprises become more agile and intelligent, and further promote specialization and flexible business cooperation, and activate manufacturing resources in a wide range.

[0003] In recent years, the application of cloud manufacturing has started in China. The China Manufacturing Capacity Sharing Development Report provided by the National Information Center shows that the scale of the cloud manufacturing and manufacturing capacity sharing market has reached 412 billion yuan, and the number of enterprises providing services through the cloud manufacturing resource sharing platform has exceeded 200,000. However, with the continuous growth of the number of manufacturing cloud services, the information overload problem inevitably occurs in the manufacturing cloud service platform. Under this background, the manufacturing service recommendation technology helps users quickly find the required personalized services in the massive services by ordering the potential use relationship of the users to the services, shortens the time and space of the users to the required services, and assists the cloud platform operator to improve the user experience, which has important application value for solving the information overload problem of the cloud platform and has obtained extensive attention.

[0004] Existing service recommendation research mainly focuses on function-based service recommendation and non-function-based service recommendation, and the functions of cloud services mainly come from user demand and cloud service description. There has been some research on service feature extraction and user preference modeling in the field of service recommendation, but there are still the following deficiencies: (1) Difficulty in representing deep theme features of manufacturing services. The theme information of manufacturing services is mainly extracted by algorithms to represent the most representative manufacturing service description information. Existing researches mainly focus on the extraction of shallow description information features, and lack of research on the extraction of deep implicit features of manufacturing services. (2) The sparsity of the traditional behavior-based user service matrix affects the efficient construction of the user preference model. User preferences are mainly reflected by constructing a user service matrix, and the manufacturing cloud platform generally includes a large number of users and manufacturing services, which makes the constructed user service matrix too sparse and cannot efficiently represent user preferences.

[0005] Therefore, it is necessary to propose a new manufacturing cloud service recommendation research method, which can solve the problems of incomplete manufacturing service information feature extraction and sparse user service matrix in the prior art. SUMMARY

[0006] The technical problem to be solved by the embodiments of the present application is to provide a manufacturing cloud service recommendation generation method and system, which solves the problems of incomplete manufacturing service information feature extraction and sparse user service matrix in the prior art.

[0007] To solve the above technical problems, the embodiments of the present application provide a manufacturing cloud service recommendation generation method, which comprises the following steps:

[0008] S1, obtaining each manufacturing service descriptive information and label information of a cloud platform;

[0009] S2, using a preset Dirichlet distribution LDA topic model to vectorize and model each manufacturing service descriptive information, obtaining p potential low-dimensional topic vectors corresponding to each manufacturing service respectively, and based on the expansion rule that each potential low-dimensional topic vector is mapped to a q-dimensional word vector by Word2vec, obtaining p potential low-dimensional topic vector expanded p x q-dimensional service topic word word vectors of each manufacturing service, further inputting into a preset convolutional neural network to learn manufacturing service deep features, obtaining an output vector of manufacturing service deep features, and according to the obtained output vector of manufacturing service deep features, obtaining a manufacturing service similarity matrix through deep feature similarity calculation;

[0010] S3, using an improved K-Means clustering algorithm to cluster each manufacturing service label information to obtain a manufacturing service category, and according to the obtained manufacturing service category, determining a user category matrix model and a service category matrix model, and further obtaining a user category matrix model coefficient and a service category matrix model coefficient to construct a user preference matrix;

[0011] S4, combining the manufacturing service similarity matrix and the user preference matrix to obtain a manufacturing service recommendation list.

[0012] The step S2 specifically comprises:

[0013] After pre-processing each manufacturing service descriptive information, inputting into an LDA topic model to obtain p potential low-dimensional topic vectors corresponding to each manufacturing service respectively;

[0014] The word vector model of the manufacturing service descriptive information corpus is trained by Word2vec, and each potential low-dimensional topic vector is expanded into a q-dimensional service topic word vector by using the trained word vector model, so as to obtain p potential low-dimensional topic vectors of each manufacturing service.

[0015] The p×q-dimensional service topic word vector of each manufacturing service is input into the convolutional neural network to learn the deep features of the manufacturing service, so as to obtain a one-dimensional feature vector of the descriptive information of each manufacturing service, and form an output vector of the deep features of the manufacturing service according to the one-dimensional feature vector of the descriptive information of each manufacturing service.

[0016] The similarity between the vectors in the output vector of the deep features of the manufacturing service is calculated by cosine similarity calculation, so as to obtain the similarity between each two manufacturing services, and the similarity matrix of the manufacturing service is obtained based on the similarity between each two manufacturing services.

[0017] The similarity coefficient between each manufacturing service is obtained by the formula ; wherein sim(a i ,a j ) is the similarity of manufacturing services a i and a j ; c ij is the jth-dimensional vector of manufacturing service i in the manufacturing service feature vector ; and N is the dimension, which is equal to the total number n of manufacturing services.

[0018] The similarity matrix between the manufacturing services is n×n.

[0019] The step S3 specifically comprises:

[0020] The manufacturing service categories are obtained by clustering the manufacturing service label information by using the improved K-Means clustering algorithm.

[0021] The user category matrix model and the service category matrix model are determined according to the obtained manufacturing service categories, and the user category matrix model coefficient and the service category matrix model coefficient are obtained further according to the user category matrix model and the service category matrix model.

[0022] The user preference matrix is constructed according to the obtained user category matrix model coefficient and service category matrix model coefficient.

[0023] The step of obtaining the manufacturing service categories by clustering the manufacturing service label information by using the improved K-Means clustering algorithm specifically comprises:

[0024] S31. Randomly select k labels from n labels as initial cluster centers C = (μ1, μ2, ..., μ...) k )∈R n ;

[0025] S32. The cosine similarity between two points for each label is calculated as shown in the following formula (1):

[0026]

[0027] Assign it to its class c (i) The formula is as follows, and c (i) ∈[1,k], as shown in equation (2):

[0028]

[0029] S33. For each class j, redetermine the cluster center of that class, as shown in equation (3) below:

[0030]

[0031] S34. Repeat steps S32 and S33 until convergence. Define the function shown in equation (4) to ensure convergence. At this point, the clustering of the label information can be obtained:

[0032]

[0033] Here, J represents the sum of squared distances from each label to its cluster center; K-means is to minimize J and output the μ and c corresponding to the minimum J.

[0034] The user preference matrix is ​​composed of user preference coefficients. Constructed; among them,

[0035]

[0036] Indicates user U p For manufacturing services a i Preference coefficients; A is the service category matrix model; U is the user category matrix model; α i For the user category matrix model coefficients, and n is for manufacturing services a i The number of tags included; β p For the user category matrix model coefficients, and Indicating in category C j User U p The number of manufacturing services being focused on. Indicates user U p The manufacturing services that are of interest ai the total number of manufacturing services a i represents the i-th manufacturing service.

[0037] The step S4 specifically includes:

[0038] From the user preference matrix, the first l manufacturing services with large user preference coefficients are selected as the user initial recommendation list.

[0039] The l user preference coefficients are multiplied by the similarity in the manufacturing service similarity matrix corresponding thereto, to obtain manufacturing service recommendation coefficients, and the obtained recommendation coefficients are arranged in descending order to obtain the first z services with larger similarity to the l services to form a manufacturing service recommendation list a i .

[0040] The embodiment of the application also provides a manufacturing cloud service recommendation generation system, comprising:

[0041] A manufacturing service information acquisition unit is configured to acquire manufacturing service descriptive information and label information of a cloud platform.

[0042] A manufacturing service information feature extraction unit is configured to use a preset Dirichlet distribution LDA topic model to vectorize and model the manufacturing service descriptive information, to obtain p potential low-dimensional topic vectors corresponding to each manufacturing service, and based on an expansion rule that each potential low-dimensional topic vector is mapped to a q-dimensional word vector by Word2vec, to obtain p×q-dimensional service topic word vectors expanded from the p potential low-dimensional topic vectors of each manufacturing service, and further input the service topic word vectors into a preset convolutional neural network to learn manufacturing service deep features and obtain an output vector of the manufacturing service deep features, and based on the output vector of the manufacturing service deep features, a manufacturing service similarity matrix is obtained through deep feature similarity calculation.

[0043] A user preference perception information construction unit is configured to use an improved K-Means clustering algorithm to cluster the manufacturing service label information to obtain manufacturing service categories, and based on the obtained manufacturing service categories, to determine a user category matrix model and a service category matrix model, and further obtain user category matrix model coefficients and service category matrix model coefficients, to construct a user preference matrix.

[0044] A manufacturing service recommendation list generation unit is configured to combine the manufacturing service similarity matrix and the user preference matrix to obtain a manufacturing service recommendation list.

[0045] The embodiment of the application has the following beneficial effects:

[0046] In one aspect, the manufacturing cloud service deep feature extraction method based on the convolutional neural network (CNN) is proposed by combining the latent Dirichlet allocation (LDA) topic model and the word vector model, so as to realize the acquisition of the required potential associated information in service recommendation; in another aspect, the concept of category is introduced by the label clustering method to perceive the user behavior, the user preference model is efficiently established, the problem of the sparse user service matrix is solved, and the effect of the manufacturing service recommendation is improved. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings obtained according to these drawings without creative labor are still within the scope of the present application.

[0048] Figure 1 The flow chart of the manufacturing cloud service recommendation generation method provided by the embodiment of the present application;

[0049] Figure 2 The structure diagram of the LDA topic model in the manufacturing cloud service recommendation generation method provided by the embodiment of the present application;

[0050] Figure 3 The structure diagram of the convolutional neural network model in the manufacturing cloud service recommendation generation method provided by the embodiment of the present application;

[0051] Figure 4 The formation process diagram of the manufacturing cloud service recommendation list in the manufacturing cloud service recommendation generation method provided by the embodiment of the present application;

[0052] Figure 5 The effect comparison diagram of the manufacturing cloud service respectively clustered by the LDA, PLSA and Our model algorithms in the application scenario of the manufacturing cloud service recommendation generation method provided by the embodiment of the present application;

[0053] Figure 6 The F1 value result comparison diagram of the manufacturing cloud service respectively clustered by the LDA, PLSA and Our model algorithms in the application scenario of the manufacturing cloud service recommendation generation method provided by the embodiment of the present application;

[0054] Figure 7 The index normalization discount cumulative gain comparison diagram of the manufacturing cloud service respectively clustered by the UPCC, IPCC, WSRec, SVD and Our model algorithms in the application scenario of the manufacturing cloud service recommendation generation method provided by the embodiment of the present application;

[0055] Figure 8 A manufacturing cloud service recommendation generation structure schematic diagram is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0056] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings.

[0057] As Figure 1 shown, a manufacturing cloud service recommendation generation method is provided in an embodiment of the present application, and the method comprises the following steps:

[0058] Step S1, obtaining manufacturing service descriptive information and label information of a cloud platform;

[0059] Step S2, using a preset Dirichlet distribution LDA topic model to vectorize and model the manufacturing service descriptive information, obtaining p potential low-dimensional topic vectors corresponding to each manufacturing service, and based on an expansion rule that each potential low-dimensional topic vector is mapped to a q-dimensional word vector by Word2vec, obtaining p×q-dimensional service topic word vectors expanded from the p potential low-dimensional topic vectors of each manufacturing service, further inputting into a preset convolutional neural network to learn manufacturing service deep features, obtaining an output vector of manufacturing service deep features, and according to the obtained output vector of manufacturing service deep features, obtaining a manufacturing service similarity matrix through deep feature similarity calculation;

[0060] Step S3, using an improved K-Means clustering algorithm to cluster the manufacturing service label information to obtain manufacturing service categories, and according to the obtained manufacturing service categories, determining a user category matrix model and a service category matrix model, and further obtaining a user category matrix model coefficient and a service category matrix model coefficient to construct a user preference matrix;

[0061] Step S4, combining the manufacturing service similarity matrix and the user preference matrix to obtain a manufacturing service recommendation list.

[0062] The specific process is that in step S1, a manufacturing service set A={a1, a2, …, an} is obtained, descriptive information D={d1, d2, …, dn} is obtained, and label information K=(k1, k2, …, kn) is obtained. i ,…,a n} is obtained, descriptive information D i ={d i1 ,d i2 ,…,d im} is obtained, and label information K j =(k j1 ,k j2 ,…,k jm ) is obtained.

[0063] In step S2, first, the manufacturing service descriptive information is preprocessed and input into the LDA topic model to obtain p latent low-dimensional topic vectors corresponding to each manufacturing service; wherein, as shown in Figure 2 each topic k = 1, 2, …, T in the LDA topic model corresponds to a probability distribution of a word i.e. the distribution rule of the word under the topic; each manufacturing service a i ∈A in the model corresponds to a topic probability distribution θ m ~ Dirichlet (α).

[0064] At this time, the generation process of each word d i ∈D of the manufacturing service descriptive information is to first select a topic z m,n ~ multi (θ m ), and then generate the word under the topic wherein, T is the number of topics; is the multinomial distribution of the word under the kth topic; θ m is the multinomial distribution of the topic under the mth manufacturing service; z m,n is the probability of the nth word in the descriptive information of the mth manufacturing service; ω m,n is the nth word in the descriptive information of the mth manufacturing service; β and α are the hyperparameters of the Dirichlet distribution to which and θ m are subjected.

[0065] The parameters and θ m are solved by the Gibbs sampling algorithm, and θ m is a 1 × T vector, denoted as , which represents the topic distribution of the descriptive information d i of the manufacturing service a i , i.e. the manufacturing service topic vector, and ω m,n ∈D is the T topic word distribution, denoted as wherein, the topic distribution vector θ m satisfies the constraint , thereby obtaining the topic vector representation of the manufacturing service descriptive information.

[0066] Second, the word vector model of the manufacturing service descriptive information corpus is trained by Word2vec, and each latent low-dimensional topic vector is expanded into a q-dimensional service topic word vector by using the trained word vector model, so as to obtain p latent low-dimensional topic vectors of each manufacturing service expanded into p × q-dimensional service topic word vectors.

[0067] Word2vec is a word vector calculation tool widely used in the field of natural language processing (NLP). It can convert a word into a word vector containing context information, represented as: v(d) ∈ R m , where d ∈ D, d represents a word in the manufacturing service descriptive information, D represents the manufacturing service descriptive information corpus, v(d) represents the word vector of d, and m is the dimension of the word vector. In the trained word vector model, the word vectors of words with similar meanings are closer in the word vector space, and they have similar semantic and grammatical relationships.

[0068] The manufacturing service descriptive information corpus is input as the context information of the manufacturing service topic, and the fixed-length word sequence Context i = {w i-c ,w i-c+1 ,…,w i-1 ,w i+1 ,…,w i+c-1 ,w i+c}, w i is the current word, and c is the size of the window.

[0069] Using the chain rule and Markov assumption, the probability of each word is approximated by the product of the previous topic words, and the probability of the next topic word is obtained. The Skip-gram model is selected as the basic framework, which samples the words in the context environment during training. The essence is to calculate the cosine similarity between the input vector of the input word and the output vector of the target word. Negative sampling is used to optimize the training process. Unlike updating all weights for each training sample, negative sampling only updates a small part of the weights for each training sample, which reduces the computational complexity in the gradient descent process. For the Skip-gram model, given the current word w i , its context Context(w i ) is predicted, and the objective function is defined as The construction of the conditional probability function p(Context(w i )|w i ) is key, and Skip-gram defines it as

[0070] Therefore, the word vector model of the manufacturing service descriptive information corpus is trained by Word2vec, and the q-dimensional service topic word vector of each manufacturing service is obtained using the trained word vector model, represented as where and n ∈ N +, the word vector of the p x q dimension service theme word of each manufacturing service.

[0071] In the third step, the word vector of the p x q dimension service theme word of each manufacturing service is input into a convolutional neural network to learn the deep features of the manufacturing service, to obtain a one-dimensional feature vector of the descriptive information of each manufacturing service, and to form an output vector of the deep features of the manufacturing service according to the one-dimensional feature vector of the descriptive information of each manufacturing service.

[0072] The multi-layer convolution in the CNN model can obtain the mutual association between the words in the manufacturing service description information, learn the global information of the manufacturing service and the relationship between the contexts, and then obtain the global deep semantic representation of the manufacturing service.

[0073] As shown in Figure 3 , the CNN model includes an original description layer, a theme layer, an embedding layer, a convolutional layer, a pooling layer and an output layer.

[0074] 1) Embedding layer

[0075] The q dimension word vector model is embedded into the p dimension manufacturing service theme word vector to obtain a p x q dimension word vector of the theme word of the manufacturing service description, and for the manufacturing service a1, there is wherein as shown in the following formula:

[0076]

[0077] wherein, is the word vector, is the word vector weight, a i is a single manufacturing service matrix.

[0078] 2) Convolutional layer

[0079] The convolutional layer is used to extract the context features, and the manufacturing service word vector matrix a i extracts the features, and the sliding window size used in the convolution can be represented as shown in the following formula:

[0080]

[0081] wherein, b i is the activation value on a certain convolution kernel; w i,j is the weight; relu is the activation function; a iThe input single manufacturing service matrix is manufactured for the convolution layer. After the above convolution operation, the output of the convolution layer is: B={b1, b2, …, b n}, wherein B is new manufacturing service information features formed by different convolution kernels, as the input of the convolution pooling layer.

[0082] 3) Pooling layer

[0083] The pooling layer is used to extract representative features from the convolution layer, and can also process variable length service information by constructing a fixed length feature vector. The pooling layer adopts maximum pooling, each convolution kernel corresponds to a value, and these values are spliced to obtain a new feature quantity representing the manufacturing service information, denoted as

[0084] 4) Output layer

[0085] In the output layer, the new feature quantity is mapped into the final manufacturing service description information deep feature representation. The original manufacturing service description information is converted into a manufacturing service description information feature vector by using the convolution neural network, and a one-dimensional feature vector of each manufacturing service description information is obtained, that is, each manufacturing service can be represented by a 1×q-dimensional vector, n is the number of manufacturing services contained, as follows:

[0086] C=cnn(w',a i )=relu(p1{relu(p2d z +b1)}+b2);

[0087] Wherein, p1, p2 are mapping matrices, b1, b2 are biases, d z is the output of the pooling layer; a i is the input of the convolution neural network; w' is the parameter of the convolution neural network, and the output vector of n manufacturing service deep features is finally obtained c ij ∈[0,1] is the vector representation of the jth dimension of the ith manufacturing service.

[0088] Fourthly, the similarity between the vectors in the output vector of the manufacturing service deep features is calculated by cosine similarity calculation to obtain the similarity between the manufacturing services, and based on the similarity between the manufacturing services, a manufacturing service similarity matrix is obtained.

[0089] Wherein, the similarity coefficient between the manufacturing services is obtained by the formula Wherein, sim(a i ,a j ) is the similarity of manufacturing services a i and a j ; c ij is the manufacturing service feature vector where N is the dimensionality, which is equal to the total number of manufacturing services n.

[0090] It should be noted that the similarity matrix between manufacturing services is n x n dimension, according to the similarity matrix, after descending order, the other related manufacturing services with high similarity to manufacturing service a i can be obtained as rank(a i ) = {a i1 , a i2 , …, a i10}.

[0091] In step S3, the first step, when the user U p enters the manufacturing service platform, according to the user's interest in manufacturing services, the user U p adds the manufacturing services he is interested in to his watchlist. Therefore, the improved K-Means clustering algorithm is used to cluster the manufacturing service label information to obtain the manufacturing service category, i.e. the category of interest to the user.

[0092] At this time, the improved K-Means clustering algorithm is used to cluster the manufacturing service labels, and the clustering process is as follows:

[0093] Step S31, randomly select k labels from n labels as initial clustering centers C = (μ1, μ2, …, μ k ) ∈ R n ;

[0094] Step S32, for each label, calculate the cosine similarity between two points as shown in the following formula (1):

[0095]

[0096] and assign it to the class c (i) , the formula is as follows, and c (i) ∈ [1, k], as shown in the following formula (2):

[0097]

[0098] Step S33, for each class j, re-determine the clustering center of the class, as shown in the following formula (3):

[0099]

[0100] Step S34, repeat step S32 and step S33 until convergence, define the function as shown in the following formula (4) to ensure convergence, and thus the clustering of the label information can be obtained:

[0101]

[0102] Here, J represents the sum of squared distances from each label to its cluster center; K-means is to minimize J and output the μ and c corresponding to the minimum J.

[0103] The second step is to determine the user category matrix model and the service category matrix model based on the obtained manufacturing service categories, and further obtain the user category matrix model coefficients and the service category matrix model coefficients based on the user category matrix model and the service category matrix model.

[0104] At this point, clustering yields manufacturing service categories. A manufacturing service can contain multiple labels, and therefore can belong to multiple categories, which can be used... This allows us to obtain the Service Category Matrix (SC) model. Here, n represents the value of manufacturing service a. i The number of tags included, α i These are the coefficients for the service category model.

[0105] Table 1 below shows an example of a service category matrix model:

[0106] Table 1

[0107]

[0108] At this point, the user category model establishes a relationship between users and categories, where user U p For users interested in manufacturing services, the category is the category to which the manufacturing service the user is interested belongs, and the two are linked through a user category matrix model (UC). The relationship between users and manufacturing services is established by a user adding a manufacturing service to their watchlist when they are interested in it. Each manufacturing service has a corresponding tag, thus revealing the relationship between users and these tags. The coefficients of the user category matrix model are primarily determined through... Perform calculations; where, ; Indicating in category C j User U p The number of manufacturing services being focused on. Indicates user U p The manufacturing services that are of interest a i The total number, a i This represents the i-th manufacturing service.

[0109] Table 2 below shows an example of a user category matrix model:

[0110] Table 2

[0111]

[0112] Thirdly, the service category matrix model and the user category matrix model are constructed, and then the user preference coefficient is obtained The user preference matrix is constructed in this way.

[0113] At this time,

[0114] Among them, represents the user preference coefficient of the user U p to the manufacturing service a i ; A is the service category matrix model; and U is the user category matrix model.

[0115] In step S4, firstly, the first l manufacturing services with large user preference coefficients are selected from the user preference matrix as the initial user recommendation list;

[0116] Secondly, the l user preference coefficients are multiplied by the similarities in the similarity matrix of the corresponding manufacturing services to obtain the manufacturing service recommendation coefficients, and the obtained recommendation coefficients are arranged in descending order to obtain the first z services with large similarities to the l services to form the manufacturing service a i recommendation list.

[0117] At this time, the deep feature mining of the manufacturing service descriptive information and the acquisition of the user preference coefficient are completed, and the service recommendation for the user is completed, and the corresponding execution process is as shown in Figure 4 .

[0118] As shown in Figures 5 to 7 , a manufacturing cloud service recommendation generation method in the embodiment of the application is verified and compared, and the specific process is as follows:

[0119] In order to verify the effectiveness of the method, the data related to ProgrammableWeb.com platform is crawled as the verification of the method proposed in the application. ProgrammableWeb.com is the largest service and service combination online repository so far, although it is not a special manufacturing service data set, but the service model, service combination relationship and its topology structure contained in it are consistent with the manufacturing service data set, among which the service model includes name, provider, category, release date, label and service description, and the service combination includes name, creation date, developer, calling service list and service combination description.

[0120] Therefore, the application adopts the ProgrammableWeb.com dataset to verify the effect of the proposed recommendation algorithm. Data in this system is crawled by parsing network source files, including 1255 services, 179 users, a total of 83057 words of various descriptive information, and a total of 5454 words of various tag information. Among them, the descriptive information of the manufacturing service is preprocessed and data cleaned before the experiment, such as converting the original text to lowercase, removing punctuation, tokenizing, removing stop words, extracting stems, and the like. The processed descriptive information will be input to the next algorithm.

[0121] To objectively evaluate the prediction effect of the experimental method, the application selects four commonly used service recommendation methods, UPCC, IPCC, WSRec, and SVD, for comparison test. Among them, the user-based Pearson Correlation Coefficient (UPCC) algorithm in the neighbor-based algorithm is a method of finding similar users based on the user's preference for services, using the Pearson correlation coefficient to recommend services that similar users like to the current user. The Item based Pearson Correlation Coefficient (IPCC) is a neighbor-based recommendation method based on items, using the Pearson correlation coefficient as a measure of similar users to recommend services to users. The collaborative filtering based web service recommender system (WSRec) is a collaborative filtering (hybrid) recommendation method based on users and items, which introduces a confidence weight to balance the prediction values of the user-based and item-based methods. Singular Value Decomposition (SVD) is a method of singular value decomposition of user-item ratings, which learns the implicit dimensions in the items.

[0122] The experimental parameters of the application are set as follows: according to experience, the Dirichlet distribution hyperparameters are set as alpha = 50 / T and beta = 0.01, the upper limit of the number of iterations of the clustering algorithm is N m = 500, and the termination threshold of the clustering algorithm iteration is delta = 0.00001; after parameter optimization, the LDA model topic number T is set to 10, the Gibbs sampling iteration number N iter = 1000, and the number of clustering centers of the clustering algorithm K = 10. The experimental parameters of each comparison method are set to the optimal parameters.

[0123] To compare the performance of the five recommended methods, the data crawled by ProgrammableWeb.com platform is used to compare the performance of the recommended method, and 25 experimenters are invited to mark whether the recommended manufacturing cloud service meets their own preference requirements, and each is required to read the descriptive information and labels to familiarize themselves with the functions of the manufacturing service. For the manufacturing services recommended to the experimenters, they are required to add the service to their watchlist if the user is interested in each manufacturing service, otherwise, they are not interested. The recommended list includes 10 services, and if the probability of the user adding the service to their watchlist is more than 50%, it is considered that the accuracy of this recommended list is high, otherwise, the recommendation is inaccurate.

[0124] An example is selected to further illustrate the characteristics of the data set and the training process and prediction steps of the proposed model. Taking the services Shapeways, Authentise, and Make Printable on ProgrammableWeb.com as an example, the execution process of the model proposed in the application is intuitively displayed to the user when recommending, and the service data crawled from ProgrammableWeb.com mainly includes service name, service description information, and service label. The training steps of the model are as follows:

[0125] (1) The original data of Shapeways, Authentise 3DIAX, and Make Printable includes a text form description, after data set preprocessing and data cleaning, three 30-dimensional description word vectors are obtained, each item of the vector represents a word, and the value of each item represents the number of occurrences of the corresponding word in the description text.

[0126] (2) The topic features of the word vector are extracted by using the LDA topic model, and the preprocessed text of Shapeways, Authentise, and Make Printable is input into the LDA topic model to obtain 10 topic word vectors and weight representations of each service. Among them, the topic vector word and weight representation of Shapeways is as follows: w1=(0.061 printing, 0.061 3D, 0.042 material, 0.042 available, 0.042 user, 0.042 shapeways, 0.023 order, 0.023 provide, 0.023 set, 0.023 plastic), and Authentise and Make Printable are the same.

[0127] (3) Word2vec word vector training and input into the convolutional neural network. 1255 service description information as corpus for word vector training, get each service word vector is 100-dimensional, so the theme word of Shapeways, Authentise 3DIAX and Make Printable can be represented by 10x100-dimensional word vector and input into the convolutional neural network, and get 1x100-dimensional key feature representation after key feature extraction, and get the similarity between the three by cosine similarity, wherein the similarity between Make Printable and Shapeways and Authentise 3DIAX is 0.669 and 0.568 respectively.

[0128] (4) In the user preference modeling stage, according to the services concerned by the user watchlist, the label information of each service is extracted for label clustering to get two categories of 3D and printing. At this time, the user 3D category matrix user-category model coefficient is 0.884, and the service in the service category matrix belonging to the 3D category is Make Printable, and the service-category model coefficient of Make Printable is 0.5, so the user's preference coefficient for Make Printable is 0.442.

[0129] In the final recommendation process, the user's preference coefficient for Make Printable is 0.442, and the similarity between Make Printable and Shapeways and Authentise 3DIAX is 0.669 and 0.568 respectively from the manufacturing service similarity matrix. The similarity between Make Printable and itself is 1. At this time, multiply the preference coefficient and the similarity to get 0.296, 0.251, 0.442. At this time, arrange the coefficients in descending order and recommend the service list to the user as Make Printable, Shapeways, Authentise 3DIAX. In fact, since Make Printable API is mainly used to integrate the function of making 3D model printable into their application, system and workflow, it is similar to the function of Shapeways and Authentise 3DIAX, and it is consistent with the service required by the user. The user indeed adds these three services to his own watchlist, which shows that the recommendation model proposed in this example is effective.

[0130] The manufacturing cloud services are clustered by using LDA, PLSA and Ourmodel algorithm respectively, the iteration number of the algorithm is 2000, the K value is set to 10, and the clustering results of each label are 10 categories. According to the accuracy, recall, F1 value and global accuracy, the experimental comparison is carried out. The results of each service label after LDA, PLSA and improved K-means clustering are as shown in the following table 3, which shows the clustering effect of each method in each category.

[0131] Table 3

[0132]

[0133] As Figure 5 shown, the P value, R value and F1 value of ACU for the three different methods are improved to a certain extent, for example, the P value of ACU is improved from 84.34% to 86.43% on LDA, and the F1 value of ACU is improved by 12.22%. Among the above three clustering algorithms, the PLSA and LDA model clustering extracts the theme word and depends on the word frequency calculation, and the PLSA model extracts the theme word amplitude change slightly. It can be seen that the improved K-means clustering can have a significant impact on the label clustering and represent the manufacturing service category. Overall, the improved K-means clustering algorithm has higher accuracy than LDA and PLSA itself. For the PLSA and LDA model, it is mainly applied to long text clustering, and the present application is mainly used for manufacturing service label clustering, that is, short text. Therefore, the improved K-means clustering has better clustering effect on the present application.

[0134] As Figure 6 shown, for the top ten categories C1-C10, the F1 value of each category can clearly show that the F1 of PLSA is the smallest, and the F1 of the improved K-Means clustering is the largest, indicating that the improved K-Means clustering is more accurate for manufacturing service label clustering, and the use of LDA model for manufacturing label clustering is more accurate than the use of PLSA. This is because the present application is used for manufacturing service label clustering, which is mainly short text clustering. Therefore, the improved K-Means is more suitable for the present application.

[0135] As Figure 7 shown, the NDCG@n experimental results of the recommended model and the related comparison method of the present application are shown, the horizontal coordinate takes the recommended list of different lengths, and the vertical coordinate represents the NDCG@n results of each algorithm.

[0136] According Figure 7We can see that different recommendation list length under each algorithm shows different NDCG@n, and NDCG@n gradually decreases with the increase of the length of the recommendation list. When the recommendation list is short, that is, when the recommendation list is 5, only when the user adds more than 3 services to his own watchlist, the recommendation accuracy is considered to be high at this time. However, since the user browses fewer services when the recommendation list is short, the requirement for the service is low, so the service added to the watchlist is far from the user's preference and cannot represent the user's preference. With the increase of the recommendation list, the user browses more manufacturing services, and the requirement for the service is also high, and the service that meets the user's preference will be more strictly required, so the service added to the watchlist will be less. When the service recommendation list increases to 30, the service added to the watchlist needs to be more than 15 to make the recommendation accuracy higher, but this is relatively difficult, so NDCG@n gradually decreases with the increase of the length of the service recommendation list. Therefore, the length of the recommendation list is 10 in the present application, which is moderate and can better recommend services to the user.

[0137] In the model mentioned in the present application, different NDCG@n of each recommendation method is different. The NDCG@n of UPCC, IPCC and WSRec is low because the recommendation is made through the user and item matrix, and the sparsity of the matrix is not reduced, so the user's preference is not well perceived. As can be seen from the above figure, SVD is one of the four methods with higher recommendation accuracy, which is because SVD is a matrix decomposition method that exists to reduce the sparsity of the matrix, so the recommendation accuracy is relatively higher than the other three methods. However, for the model proposed in the present application, the category is introduced to model the user's preference in order to reduce the problem of matrix sparsity, so the recommendation accuracy is higher.

[0138] As Figure 8 shown, a manufacturing cloud service recommendation generation system is provided in the embodiment of the present application, which comprises:

[0139] The manufacturing service information acquisition unit 110 is configured to acquire the descriptive information and label information of each manufacturing service of the cloud platform.

[0140] The manufacturing service information feature extraction unit 120 is configured to use a preset Dirichlet distribution LDA topic model to vectorize and model each manufacturing service descriptive information, obtain p potential low-dimensional topic vectors corresponding to each manufacturing service, and based on an expansion rule that each potential low-dimensional topic vector is mapped to a q-dimensional word vector by Word2vec, obtain p potential low-dimensional topic vector expanded p x q-dimensional service topic word vectors of each manufacturing service, further input into a preset convolutional neural network to learn manufacturing service deep features, obtain an output vector of manufacturing service deep features, and based on the obtained output vector of manufacturing service deep features, obtain a manufacturing service similarity matrix by deep feature similarity calculation.

[0141] The user preference perception information construction unit 130 is configured to use an improved K-Means clustering algorithm to cluster each manufacturing service label information to obtain a manufacturing service category, determine a user category matrix model and a service category matrix model according to the obtained manufacturing service category, and further obtain a user category matrix model coefficient and a service category matrix model coefficient to construct a user preference matrix.

[0142] The manufacturing service recommendation list generation unit 140 is configured to combine the manufacturing service similarity matrix and the user preference matrix to obtain a manufacturing service recommendation list.

[0143] The embodiment of the present application has the following beneficial effects:

[0144] In one aspect, the present application combines an implicit Dirichlet distribution (LDA) topic model and a word vector model, proposes a manufacturing cloud service deep feature extraction method based on a convolutional neural network (CNN), and realizes the acquisition of potential associated information required in service recommendation. In another aspect, the concept of category is introduced by a label clustering method to perceive user behavior, efficiently establish a user preference model, solve the sparsity problem of a user service matrix, and improve the effect of manufacturing service recommendation.

[0145] It is worth noting that in the above system embodiment, each unit included is only divided according to functional logic, but is not limited to the above division, as long as the corresponding function can be realized. In addition, the specific name of each functional unit is only for easy differentiation, and does not limit the protection scope of the present application.

[0146] Those skilled in the art can understand that all or part of the steps of the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc.

[0147] The above merely provides the preferred embodiment of the application, and cannot allude the protection scope of the application, therefore any equivalent changes made according to the claims of the application shall be within the scope of the application.

Claims

1. A manufacturing cloud service recommendation generation method characterized by, The method comprises the following steps: S1, obtaining manufacturing service descriptive information and label information of a cloud platform; S2. Using a pre-defined LDA topic model with Dirichlet distribution, the descriptive information of each manufacturing service is vectorized and modeled to obtain the corresponding information for each manufacturing service. p There are 10 potential low-dimensional topic vectors, and each potential low-dimensional topic vector is mapped to a Word2vec value. q The rules for expanding word vectors in dimensional space yield the results for each manufacturing service. p Expanding a potential low-dimensional topic vector p×q The word vectors of the service topic words are further input into a preset convolutional neural network to learn the deep features of manufacturing services, and the output vector of the deep features of manufacturing services is obtained. Based on the output vector of the deep features of manufacturing services, the similarity matrix of manufacturing services is obtained by calculating based on the deep feature similarity. S3, using an improved K-Means clustering algorithm to cluster the manufacturing service label information to obtain manufacturing service categories, and determining a user category matrix model and a service category matrix model according to the obtained manufacturing service categories, and further obtaining a user category matrix model coefficient and a service category matrix model coefficient to construct a user preference matrix; S4, combining the manufacturing service similarity matrix with the user preference matrix to obtain a manufacturing service recommendation list.

2. The manufacturing cloud service recommendation generation method of claim 1, wherein, The step S2 specifically comprises: After the manufacturing service descriptive information is preprocessed, it is input into the LDA topic model to obtain a corresponding p latent low-dimensional topic vector for each manufacturing service. The word vector model of the manufacturing service descriptive information corpus is trained by Word2vec, and each potential low-dimensional topic vector is expanded to a word vector of a service topic with q dimension by using the trained word vector model, so as to obtain the word vector of the service topic of each manufacturing service p p×q dimension by using the trained word vector model, so as to obtain the word vector of the service topic of each manufacturing service​ The manufacturing service description information of each manufacturing service is input into a convolutional neural network to learn deep features of the manufacturing service, and a one-dimensional feature vector of the manufacturing service description information is obtained. p×q The manufacturing service description information of each manufacturing service is input into a convolutional neural network to learn deep features of the manufacturing service, and a one-dimensional feature vector of the manufacturing service description information is obtained. The similarity between two manufacturing services is obtained by calculating the similarity between two vectors in the output vector of the manufacturing service deep feature through cosine similarity calculation, and the manufacturing service similarity matrix is obtained based on the similarity between two manufacturing services.

3. The manufacturing cloud service recommendation generation method of claim 2, wherein, Through formula The similarity coefficients between various manufacturing services are obtained; among them, Serving manufacturing and Similarity; Feature matrix for manufacturing services Manufacturing services i No. j A vector of 3 dimensions; N It is a dimension, the value of which equals the total number of services produced. n .

4. The manufacturing cloud service recommendation generation method of claim 3, wherein, The similarity matrix between the manufacturing services is n×n dimensions.

5. The manufacturing cloud service recommendation generation method of claim 1, wherein, The step S3 specifically comprises: Using an improved K-Means clustering algorithm to cluster the manufacturing service label information to obtain manufacturing service categories; According to the obtained manufacturing service categories, determining a user category matrix model and a service category matrix model, and further obtaining a user category matrix model coefficient and a service category matrix model coefficient according to the user category matrix model and the service category matrix model; According to the obtained user category matrix model coefficient and service category matrix model coefficient, constructing the user preference matrix.

6. The manufacturing cloud service recommendation generation method of claim 5, wherein, The step of using an improved K-Means clustering algorithm to cluster the manufacturing service label information to obtain manufacturing service categories specifically comprises: S31、in n randomly select k tags as initial clustering centers ; S32, for each label, the cosine similarity between two points is calculated as shown in the following formula (1): （1）； and assigning it to the class to which it belongs The formula in the middle is as follows, and as shown in the following equation (2): （2）； S33, for each category j, the cluster center of the category is re-determined as shown in the following formula (3): （3）； S34, repeat steps S32 and S33 until convergence, define the function as shown in the following formula (4) to ensure convergence, and thus obtain the clustering of label information: （4）； wherein, J The function represents the sum of the squared distances of each label to its cluster center; K-means is to adjust J to the minimum, take the minimum J corresponding to µ and c output.

7. The manufacturing cloud service recommendation generation method of claim 5, wherein, The user preference matrix is constructed by user preference coefficients are constructed; wherein, ； Indicates user Manufacturing services Preference coefficient; A U is the service category matrix model; U is the user category matrix model; For the user category matrix model coefficients, and ; n Serving manufacturing The number of tags included; For the user category matrix model coefficients, and ; Indicates in category C j user The number of manufacturing services being focused on. Indicates user Manufacturing services of interest Total quantity Indicates the first i Manufacturing services.

8. The manufacturing cloud service recommendation generation method of claim 7, wherein, The step S4 specifically comprises: From the user preference matrix, select the top l manufacturing services with large user preference coefficients as the initial recommendation list for the user; This l Each user preference coefficient is multiplied by its similarity in the corresponding manufacturing service similarity matrix to obtain the manufacturing service recommendation coefficient. These recommendation coefficients are then sorted in descending order to obtain the recommendation coefficients for the manufacturing service. l The services with high similarity z One service To form manufacturing services Recommended list.

9. A manufacturing cloud service recommendation generation system characterized by, Including: A manufacturing service information acquisition unit is configured to obtain manufacturing service descriptive information and label information of a cloud platform; The manufacturing service information feature extraction unit is used to vectorize and model the descriptive information of each manufacturing service using a pre-defined LDA topic model with Dirichlet distribution, thereby obtaining the corresponding features for each manufacturing service. p There are 10 potential low-dimensional topic vectors, and each potential low-dimensional topic vector is mapped to a Word2vec value. q The rules for expanding word vectors in dimensional space yield the results for each manufacturing service. p Expanding a potential low-dimensional topic vector p×q The word vectors of the service topic words are further input into a preset convolutional neural network to learn the deep features of manufacturing services, and the output vector of the deep features of manufacturing services is obtained. Based on the output vector of the deep features of manufacturing services, the similarity matrix of manufacturing services is obtained by calculating based on the deep feature similarity. A user preference perception information construction unit is configured to use an improved K-Means clustering algorithm to cluster the manufacturing service label information to obtain manufacturing service categories, and determine a user category matrix model and a service category matrix model according to the obtained manufacturing service categories, and further obtain a user category matrix model coefficient and a service category matrix model coefficient to construct a user preference matrix; A manufacturing service recommendation list generation unit is configured to combine the manufacturing service similarity matrix with the user preference matrix to obtain a manufacturing service recommendation list.

Images