A manufacturing resource recommendation system based on a knowledge graph and collaborative filtering

Abstract

A manufacturing resource recommendation system based on knowledge graph and collaborative filtering includes 6 modules, namely: data preprocessing module, collaborative filtering module, technology content calculation module, knowledge graph module, similarity fusion module and recommendation module. For the recommendation system of manufacturing resources, fully combines the characteristics of manufacturing resources, and introduces the resource relationship knowledge graph, breaks the resource isolation, reflects the real and reliable manufacturing capacity, and can effectively improve the recommendation accuracy of the system, enhance the recommendation richness and recommendation interpretability, and has good innovation significance.

Description

TECHNICAL FIELD

[0001] The present application is directed to manufacturing resource characteristics, creating a manufacturing resource recommendation method based on knowledge graph and collaborative filtering, belonging to the field of recommendation algorithm. BACKGROUND

[0002] In user-oriented systems, an important technology is the recommendation system, especially in today's era, users face massive data and information every day, and the information overload problem is serious. In some cases, users can be more clearly aware of their search goals and needs, but due to the diversity of choices and thousands of choices for each type of information, users often do not have clear goals. At this time, the recommendation system is needed to help users make choices, which helps various platforms to mine users' potential needs and improve revenue. Recommendation systems are often used in movie, book, e-commerce, resource service, song, etc. platforms.

[0003] The common method of recommendation system can be divided into three categories according to whether the knowledge graph is introduced: traditional method, method based on knowledge graph, and hybrid method, that is, the combination of knowledge graph and traditional method.

[0004] The traditional method of recommendation system can be divided into two categories: content-based recommendation and collaborative filtering-based recommendation. The application scenarios of the two methods are distinguished by the number of users. In the case of single user or small number of users, the content-based recommendation method is generally used. According to the user's browsing content, the characteristics are analyzed, and the content with similar characteristics is recommended to the user. When the platform has a certain number of users, recommendations can be made according to the similarity between users. The so-called collaboration refers to the representation of the similarity between multiple users. Collaborative filtering is divided into user-based collaborative filtering and item-based collaborative filtering. In short, the usual approach is to calculate the similarity of U2, U3, …, with respect to U1 for user U1, and if you want to calculate the interest level of U1 to item V i , then use the known interest of U2, U3, … to V i weighted average. With the continuous accumulation of platform users, the recommendation algorithm can be transitioned from content-based recommendation to collaborative filtering-based recommendation.

[0005] However, there are some obvious disadvantages of the collaborative filtering method. First, the sparsity problem of the user-item interaction matrix. Because of the existing platform, such as Taobao, not only has a large number of registered users, but also has billions of goods, and the number of goods contacted by each person is extremely limited, so that the constructed rating prediction matrix or click rate prediction matrix is very sparse. Second, the cold start problem, that is, the recommendation algorithm is difficult to execute when the new user just registers and has not interacted with the system. In order to solve this problem, auxiliary information is often added on the basis of collaborative filtering. For example, social network information, user attribute information, and context information.

[0006] Manufacturing resources include numerical control equipment, stamping equipment, industrial robots, detection equipment and 3D printing equipment, etc. The production activity elements in the product life cycle. Its characteristics are diversity and complexity, as well as the heterogeneity and isolation brought by different providing entities. At present, the manufacturing resource system adopts the traditional recommendation method, relies on the interaction information between users and resources for recommendation, ignores the above characteristics of manufacturing resources, and there is still a great isolation between resources, and the recommendation precision is low, the richness is poor, and the good recommendation effect cannot be achieved. In view of the above problems, the present application fully combines the characteristics of manufacturing resources and designs the recommendation system. SUMMARY

[0007] The purpose of the present application is to overcome the deficiencies of the prior art, and disclose a manufacturing resource recommendation system based on knowledge graph and collaborative filtering. The system adopts an algorithm framework based on the combination of knowledge graph and collaborative filtering:

[0008] First, the manufacturing resources such as numerical control equipment, stamping equipment and 3D printing equipment belong to bulk equipment. Considering the stability of the user's capital investment amount on this kind of resource, the basic similarity of the resource is obtained by fusing the price similarity on the basis of the similarity of collaborative filtering.

[0009] Second, for manufacturing resources such as industrial robots and detection equipment, the closer the publication time is, the higher the technical content is, and thus the larger the customer demand is. Therefore, the present application calculates the technical content index according to the publication time of the resource.

[0010] Third, manufacturing resources are distributed in different research institutions, enterprises, universities and innovation industrial parks, mainly including 3D printing equipment, numerical control machine tools, forging and pressing equipment, stamping equipment, welding equipment and robots, etc. They are relatively autonomous, have spatial dispersion and independence, and even the same equipment may have different manufacturing capabilities for different resource providers. Through the introduction of the knowledge graph, not only can the matrix sparsity and cold start problems of the collaborative filtering method be solved, but also for manufacturing resources, the knowledge graph can break the independence of the resources through the relationship between the resources and the providing entities, and reflect the manufacturing capability of a certain resource. It improves the accuracy, diversity and interpretability of the recommendation system.

[0011] In summary of the above three points, the application is a recommendation system for manufacturing resources, which fully combines the characteristics of manufacturing resources and introduces the resource relationship knowledge graph, breaks the resource isolation, reflects the real and reliable manufacturing capability, and can effectively improve the recommendation accuracy, enhance the recommendation richness and recommendation interpretability of the system for numerical control equipment, stamping equipment, industrial robots, detection equipment and 3D printing equipment, etc. It has good innovation significance.

[0012] Technical scheme of the application

[0013] A manufacturing resource recommendation system based on knowledge graph and collaborative filtering, characterized by comprising 6 modules, namely: data preprocessing module, collaborative filtering module, technology content calculation module, knowledge graph module, similarity fusion module and recommendation module.

[0014] The data preprocessing module obtains data including: user ratings of resources, resource price list, resource publication time list, resource relationship knowledge graph, etc. Among them, the user rating matrix of the resources, the resource price list will be input into the collaborative filtering module, the resource publication time list will be input into the technology content calculation module, and the resource knowledge graph will be input into the knowledge graph module.

[0015] The input of the collaborative filtering module is the user-resource rating table and the resource price table, and the output is the collaborative filtering similarity, which will be provided to the similarity fusion module. The rating similarity adopts cosine similarity, which is calculated according to the user-resource rating table. In view of the characteristics of manufacturing resource users that the capital investment is relatively stable, in addition to the quality rating similarity, the price similarity also needs to be calculated. The calculation method of the price similarity is simpler than that of the rating similarity, and the data will no longer be a two-dimensional matrix, but a one-dimensional price data.

[0016] The input of the technology content calculation module is a resource publishing time list, and the output is a technology content index of the resource, which is provided to the similarity fusion module. The technology content level of the resource will be divided into 10 levels 1-10, and the level is divided according to the publishing time of the resource. The earlier the publishing time is, the lower the technology content is, and the later the publishing time is, the higher the technology content is. In order to unify the similarity in the following similarity fusion module and similarity, the level numbers 1-10 are converted into the decimal interval 0-1, which is represented as a technology content index.

[0017] The input of the knowledge graph module is a resource relationship knowledge graph, and the output is a manufacturing capability semantic similarity, which is provided to the similarity fusion module. According to the knowledge graph, the representation of each resource entity is obtained by using the many-to-many model TransHR, and the manufacturing capability similarity between each resource is calculated by using the Euclidean distance.

[0018] The input of the similarity fusion module is the result obtained by the collaborative filtering module, the technology content calculation module and the knowledge graph module, and the output is the final similarity obtained by fusion according to the fusion proportion, which is output to the user resource recommendation list module.

[0019] The recommendation list generation module receives the final fusion similarity, calculates the demand score of each manufacturing resource of the user according to the near neighbor principle. The larger the score is, the more likely the user is to purchase the resource. Therefore, the top k resources are selected to form a recommendation list according to the score from high to low. In the cold start problem of the algorithm, if a new resource is added and the score list is empty, the demand score of the user for the resource can be calculated according to the knowledge graph and technology content information. If a new user is added and the interaction with the resource is less, the initial browsing search information of the user is observed for recommendation.

[0020] Further, the data preprocessing module is divided into three sub-modules, i.e. a resource demand feature sub-module, a technology content feature sub-module and a manufacturing capability feature sub-module.

[0021] The data processing required by the resource demand feature module and the technology content feature module is basic processing. In the resource demand feature sub-module, the input is user-resource score and resource price data, and the output is a user-resource co-occurrence matrix R m×n and a resource price list p = {p1, p2, …, p n}. The output of the technology content feature sub-module is a resource publishing time list s = {s1, s2, …, s n}.

[0022] In the manufacturing capability feature submodule, the input information of the manufacturing resource includes title, introduction, and manufacturing resource provider information. First, a resource knowledge graph ontology is constructed, including product ontology, part ontology, project ontology and enterprise ontology. And the relationship types are defined, specifically: the "ownership" relationship between the enterprise ontology and the equipment, and the "application" relationship between the raw materials and the equipment. Then the instances layer of the knowledge graph is filled in, and the 3D printing equipment, robots, stamping equipment manufacturing resources are matched with the ontology, and the relationships between the equipment and other instances are described, including: the experimental base has a "ownership" relationship with the 3D printer, and the printing raw materials have an "application" relationship with the 3D printer. Finally, the manufacturing resource relationship knowledge graph is completed, and is output in the form of triples, that is (h, r, t). Wherein, h and t represent 3D printing equipment, robots, stamping equipment manufacturing resources and enterprise, project entities, and r is the relationship between them. Through the knowledge graph, not only the semantic similarity between entities can be expressed, but also the isolation between various resources can be broken, and the manufacturing capability relationship can be reflected. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The main flowchart of the operation of the manufacturing resource recommendation system based on the knowledge graph and collaborative filtering.

[0024] Figure 2 For Figure 1 The specific implementation idea of the data preprocessing module in the system is shown in the figure.

[0025] Figure 3 The data flow diagram between the modules of the present application. DETAILED DESCRIPTION

[0026] This part will further illustrate the present application according to the drawings.

[0027] As Figure 1 shown, the main flow of the system operation is: first, the required data set is obtained through the data preprocessing module, and is provided to each calculation module. The three modules of collaborative filtering, knowledge graph and technology content calculation are calculated independently. Finally, the data calculated by the three modules is fused and provided to the recommendation module to generate the demand score of the user for the manufacturing resource, and finally a recommendation list is obtained.

[0028] As Figure 2 shown, the specific implementation of the data preprocessing module is: combined with the three features of the manufacturing resource in the recommendation method, the module is divided into three submodules, namely resource demand feature submodule, technology content feature submodule and manufacturing capability feature submodule.

[0029] The data processing required by the resource demand feature module and the technical content feature module is basic processing. In the resource demand feature submodule, the input is user-resource score and resource price data, and the output is user-resource co-occurrence matrix R m×n and resource price list p = {p1, p2, …, p n}. The technical content feature submodule outputs resource release time list s = {s1, s2, …, s n}.

[0030] In the manufacturing capability feature submodule, the input is detailed information of manufacturing resources such as title, introduction, etc., and manufacturing resource provider information. First, the resource knowledge graph ontology is constructed, such as product ontology, part ontology, project ontology and enterprise ontology. And, the relationship types are defined, such as the “owning” relationship between the enterprise ontology and the equipment, the “applied to” relationship between the raw materials and the equipment, etc. Then, by filling the instance layer of the knowledge graph, the 3D printing equipment, robots, stamping equipment and other manufacturing resources are corresponded to the ontology, and the relationships existing between the equipment and other instances are described, such as the “owning” relationship between the experimental base and the 3D printer, and the “applied to” relationship between the printing raw materials and the 3D printer. Finally, the manufacturing resource relationship knowledge graph is completed, and is output in the form of triples, i.e. (h, r, t). Among them, h and t represent 3D printing equipment, robots, stamping equipment and other manufacturing resources, as well as enterprises, projects and other entities, and r is the relationship between them. Through the knowledge graph, not only the semantic similarity between entities can be expressed, but also the isolation between various resources can be broken, and the manufacturing capability relationship can be reflected.

[0031] As shown in Figure 3 , each item of data obtained by the data preprocessing module will be input into the collaborative filtering module, the technical content calculation module and the knowledge graph module. The three modules are independently calculated to generate manufacturing resource collaborative filtering similarity, manufacturing resource technical content index and resource manufacturing capability similarity, and are provided to the similarity fusion module. The recommendation module generates a manufacturing resource recommendation list according to the finally generated similarity.

[0032] In the collaborative filtering module, the content and manufacturing capability of the recommended resource are not considered, which is a problem that needs to be solved by the knowledge graph module, and only the user's rating data for the resource and the price data of the resource are considered, and the main idea is that users tend to like similar items. U = {U1, U2, …, U m} represents m users, V = {V1, V2, …, V n} represents n resources, and the user-resource score matrix is R m×n , where R uiRui represents the rating of resource i by user u, Rui is the preference degree of user u to resource i. The higher the score, the higher the preference degree. The rating similarity is cosine similarity, which is calculated according to the user-resource rating table. Since the key to obtaining the best quality prediction using cosine similarity is to standardize the rating, the average rating of the manufacturing resource is subtracted from the rating table before calculating the similarity. Better results can be obtained. The rating similarity sim score The calculation formula is as follows:

[0033]

[0034] sim score (i,j) represents the rating similarity between resource i and resource j;

[0035] R ui Rui represents the rating of resource i by user u, Rui is the preference degree of user u to resource i. The higher the score, the higher the preference degree. The rating similarity is cosine similarity, which is calculated according to the user-resource rating table. Since the key to obtaining the best quality prediction using cosine similarity is to standardize the rating, the average rating of the manufacturing resource is subtracted from the rating table before calculating the similarity. Better results can be obtained. The rating similarity sim uj Ruj represents the rating of resource j by user u;

[0036] Rui represents the rating of resource i by user u, Rui is the preference degree of user u to resource i. The higher the score, the higher the preference degree. The rating similarity is cosine similarity, which is calculated according to the user-resource rating table. Since the key to obtaining the best quality prediction using cosine similarity is to standardize the rating, the average rating of the manufacturing resource is subtracted from the rating table before calculating the similarity. Better results can be obtained. The rating similarity sim

[0037] Next, the price similarity sim price of the resource is calculated according to the resource price list. The resource price similarity no longer relies on the interaction between users and resources. The smaller the price gap between the two goods, the greater the price similarity, and vice versa. p i pi represents the price of resource i, and the price gap between the two resources is divided by the lower-priced commodity to represent the gap. In manufacturing resources, when the price difference is greater than one times the price of the lower-priced resource, the price similarity of the two is low. In order to normalize the value to 0-1, when the price difference is greater than or equal to one times the price of the lower-priced resource, the similarity factor is set to ε, which can be 0.001 or other small numbers. The price similarity calculation formula is as follows:

[0038]

[0039] sim price (i,j) represents the price similarity between resource i and resource j;

[0040] p i pi represents the price of resource i, and the price gap between the two resources is divided by the lower-priced commodity to represent the gap. In manufacturing resources, when the price difference is greater than one times the price of the lower-priced resource, the price similarity of the two is low. In order to normalize the value to 0-1, when the price difference is greater than or equal to one times the price of the lower-priced resource, the similarity factor is set to ε, which can be 0.001 or other small numbers. The price similarity calculation formula is as follows: j pi represents the price of resource i, and the price gap between the two resources is divided by the lower-priced commodity to represent the gap. In manufacturing resources, when the price difference is greater than one times the price of the lower-priced resource, the price similarity of the two is low. In order to normalize the value to 0-1, when the price difference is greater than or equal to one times the price of the lower-priced resource, the similarity factor is set to ε, which can be 0.001 or other small numbers. The price similarity calculation formula is as follows:

[0041] ε is the set similarity factor.

[0042] Finally, the similarity sim cf of the collaborative filtering module is sim score sim price is added with a certain weight. That is:

[0043] sim cf (i,j)=αsim price (i,j)+(1-α)sim price (i,j)

[0044] sim cf (i,j) represents the final collaborative filtering similarity between resource i and resource j;

[0045] sim score (i,j) represents the rating similarity between resource i and resource j;

[0046] sim price (i,j) represents the price similarity between resource i and resource j;

[0047] α is a weight coefficient.

[0048] In the technical content index calculation module, the resource content is divided into 10 levels 1-10 according to the resource release time, and is converted into a decimal interval of 0-1, which is represented as a technical content index. First, the release time of all resources is counted, and all release times are divided into 10 time ranges from far to near, in units of days, for non-uniform division. The size of each time range is mainly determined according to expert experience, and according to the iteration period law of manufacturing resources, generally, the time range of low level is longer, and the time range of high level is shorter. Thus, the technical content index T=(t1,,t2,…,t n ) is obtained.

[0049] In the knowledge graph module, the knowledge graph is composed of triples represented by (h, r, t), and the entity representation I of industrial robots, stamping equipment, etc. is learned by the TransHR model. According to the network structure of the knowledge graph, the manufacturing capability semantic similarity between resources is calculated. The invention uses the Euclidean distance to calculate the similarity between resource i and resource j. Since the corresponding relationship vector between two entity vectors is not unique, for example, the experimental base and the industrial robot not only have the "own" relationship, but also may have the "design" relationship. These relationships can be divided into direct relationships and indirect relationships, and the weight λ of the direct relationship should be greater than that of the indirect relationship, and λ ranges from (0, 1). The manufacturing capability similarity, i.e. semantic similarity, is calculated as follows:

[0050]

[0051] sim sem (i,j) represents the semantic similarity between resource i and resource j;

[0052] I i represents the feature vector of resource i, and I jsim(i,j) represents the similarity between resource i and resource j.

[0053] N(I i ,I j ) and M(I i ,I j ) represent the direct relationship number and indirect relationship number between resource i and j, respectively.

[0054] λ represents the direct relationship weight.

[0055] In the similarity fusion module, the collaborative filtering similarity, the technology index content, and the manufacturing capability semantic similarity calculated by the above three modules are fused according to the fusion proportion to obtain the final similarity. The technology index content will improve the similarity between the two resources, so that the possibility of the resource being recommended to the user is greater. The final fusion similarity is:

[0056] sim(i,j) = αsim cf (i,j) + βsim sem (i,j) + γt i ;

[0057] sim(i,j) represents the similarity between resource i and resource j.

[0058] sim cf (i,j) represents the rating similarity between resource i and resource j.

[0059] sim sem (i,j) represents the semantic similarity between resource i and resource j.

[0060] t i represents the technology content index of resource i.

[0061] α, β and γ represent the fusion weight coefficients.

[0062] In the recommendation list generation module, the input is the similarity matrix between resources. The formula for calculating the predicted preference score of user u to manufacturing resource i is as follows:

[0063]

[0064] F ui represents the predicted preference score of user u to manufacturing resource i.

[0065] A(u) represents the resource set that user u has rated.

[0066] R uj represents the rating of user u to resource j.

[0067] sim(i,j) represents the similarity between resource i and resource j.

[0068] Finally, the F ui resources are sorted and the top k resources are selected to form a recommendation list.

Claims

1. A manufacturing resource recommendation system based on knowledge graphs and collaborative filtering, characterized in that, It includes six modules: data preprocessing module, collaborative filtering module, technical content calculation module, knowledge graph module, similarity fusion module, and recommendation module. The input to the collaborative filtering module is the user-resource rating table and the resource price table, and the output is the collaborative filtering similarity, which is then provided to the similarity fusion module. The technical content calculation module takes a list of resource release times as input and outputs a technical content index of the resource, which is then provided to the similarity fusion module. The knowledge graph module takes as input a resource relationship knowledge graph and outputs a manufacturing capability semantic similarity, which is provided to the similarity fusion module. Based on the knowledge graph, the many-to-many model TransHR is used to obtain the representation of each resource entity, and Euclidean distance is used to calculate the manufacturing capability similarity between each resource. The similarity fusion module takes as input the results of three calculation modules: collaborative filtering module, technical content calculation module, and knowledge graph module, and outputs the final similarity obtained by fusion according to the fusion ratio, which will be output to the user resource recommendation list module. In the collaborative filtering module, users tend to like similar items. Represents m users, Represented as n resources, the user-resource rating matrix is ​​as follows: , among them This represents the rating of user u for resource i. A higher rating indicates a greater degree of preference for that resource. The rating similarity is calculated using cosine similarity, which is based on the user-resource rating table. Rating Similarity The calculation formula is as follows: This indicates the score similarity between resource i and resource j; This represents user u's rating of resource i. This represents user u's rating of resource j; This represents the average rating of the resource from all users. Finally, the similarity of the collaborative filtering module. for and They are added together with a certain weight; that is: This represents the final collaborative filtering similarity between resource i and resource j; This indicates the score similarity between resource i and resource j; This indicates the price similarity between resource i and resource j; These are the weighting coefficients; In the similarity fusion module, the collaborative filtering similarity, technical index content, and manufacturing capability semantic similarity calculated by the collaborative filtering module, technical content calculation module, and knowledge graph module are fused according to a fusion ratio to obtain the final similarity. The technical index content increases the similarity between two resources, making it more likely that the resource will be recommended to the user. The final fused similarity is: ； This represents the fusion similarity between resource i and resource j; This indicates the score similarity between resource i and resource j; This indicates the semantic similarity between resource i and resource j; This represents the technological content index of resource i; , and This represents the fusion weighting coefficient.

2. The recommendation system as described in claim 1, characterized in that, The data preprocessing module obtains data including: user ratings of resources, resource price lists, resource release time lists, and resource relationship knowledge graphs. Among them, the user rating matrix and resource price lists will be input into the collaborative filtering module, the resource release time lists will be input into the technology content calculation module, and the resource knowledge graphs will be input into the knowledge graph module.

3. The recommendation system as described in claim 1, characterized in that, The data preprocessing module is divided into three sub-modules: resource demand characteristics, technology content characteristics, and manufacturing capability characteristics. The data processing required by the resource demand characteristics module and the technology content characteristics module is basic processing; in the resource demand characteristics submodule, the input is user-resource rating and resource price data, and the output is a user-resource co-occurrence matrix. List of resource prices ; The technical content feature submodule outputs a list of resource release times. ; In the manufacturing capability characteristics submodule, the input is information about manufacturing resources, including titles, descriptions, and information about the manufacturing resource providers. First, a resource knowledge graph ontology is constructed, including product, component, project, and enterprise ontology. Then, relationship types are defined: a "ownership" relationship between the enterprise ontology and the equipment, and a "application" relationship between raw materials and the equipment. Next, by populating the instance layers of the knowledge graph, 3D printing equipment, robots, and stamping equipment manufacturing resources are mapped to the ontology, and the relationships between the equipment and other instances are described, including a "ownership" relationship between the experimental base and the 3D printer, and an "application" relationship between printing raw materials and the 3D printer. Finally, the manufacturing resource relationship knowledge graph is completed and output as a triple, i.e., (h, r, t); where h and t represent the 3D printing equipment, robot, and stamping equipment manufacturing resources, as well as the enterprise and project entities, and r represents the relationships between them.

4. The recommendation system as described in claim 1, characterized in that, The formula for calculating manufacturing capability similarity, or semantic similarity, is as follows: This indicates the semantic similarity between resource i and resource j; This represents the feature vector of resource i. This represents the feature vector of resource j; and Let these represent the direct and indirect relationship numbers between resources i and j, respectively. λ represents the weight of direct relationships.

5. The recommendation system as described in claim 1, characterized in that, The recommendation list generation module receives the final fusion similarity and calculates the user's demand score for each manufacturing resource based on the nearest neighbor principle. The higher the score, the more likely the user is to purchase the resource. Therefore, the top k resources are selected from high to low scores to form the recommendation list. In the recommendation list generation module, the input is the similarity matrix between resources; the formula for calculating the predicted preference score of user u for manufacturing resource i is as follows: This represents the predicted preference score of user u for manufacturing resource i; This represents the set of resources that user u has already rated; This represents user u's rating of resource j; This represents the fusion similarity between resource i and resource j; Finally, for Sort the resources and select the top k to form a recommendation list.

Images