A technical resource recommendation method based on knowledge graph preference propagation

Abstract

A method for recommending technical resources based on knowledge graph preference propagation includes: step S1 based on a technical resource library; step S2 preprocessing user historical browsing data; step S3 based on the interaction matrix between the technical resource knowledge graph and the user's technical resource; and step S4 ranking technical resources by interest value according to a user interest model. The method extracts topics based on the LDA model, considering the closeness of relationships between different nodes and representing them with weights, thus improving the practicality of the knowledge graph. During data preprocessing, the method considers the change in user interests over time and introduces a time forgetting curve, improving recommendation accuracy. In implementing the recommendation algorithm, it combines an embedding-based method, which utilizes the high-dimensional semantic information of the knowledge graph and is easy to implement, with a path-based method, introducing path information from the knowledge graph through preference propagation, thus improving the interpretability of the recommendation results.

Description

Technical Field

[0001] This invention relates to a knowledge graph-based technology resource recommendation algorithm, belonging to the field of recommendation algorithms. Background Technology

[0002] With the rapid development of internet technology, we have entered the information society, and data and information are exploding. We enjoy the convenience brought by information technology, but we also struggle with how to extract the information we need from this vast amount of data. Recommendation systems, as information filtering systems, can effectively solve the problem of information overload, promote productivity, and improve the quality of life. Current recommendation systems have been extensively researched and are applied in various fields, demonstrating significant practical value.

[0003] Personalized recommendation algorithms are the core of recommender systems, and they can be mainly divided into three categories: content-based recommendation algorithms, collaborative filtering-based recommendation algorithms, and hybrid recommendation algorithms. However, traditional recommendation algorithms still suffer from several problems, primarily the sparsity of behavioral relationship data between users and items, and the cold start problem when recommending new users or items. To address these issues, researchers have attempted to combine collaborative filtering recommendation algorithms with other auxiliary information (e.g., user and item attribute features, user social network information, etc.) to build hybrid recommender systems, thereby improving recommendation performance.

[0004] Knowledge graphs, as an effective auxiliary information in hybrid recommender systems, have attracted much attention from researchers in recent years. A knowledge graph is a directed heterogeneous network containing nodes – “entities” and directed edges – “relationships”, thus containing a wealth of background information about items and the relationships between them. Knowledge graphs can improve recommender systems in three ways: (1) they introduce semantic relevance between items, helping to discover their potential connections and improving the accuracy of the recommender system; (2) they consist of various types of relationships, which helps to reasonably expand user interests and increase the diversity of the recommender system; and (3) they connect user history records with recommendation history records, thereby improving the interpretability of the recommender system.

[0005] Currently, knowledge graph-based recommendation algorithms are mainly divided into embedding-based methods and path-based methods. Both methods still have certain problems. The former loses the utilization of multi-hop relationships in the graph during recommendation, resulting in a lack of interpretability in some recommendation results and difficulty in identifying users' potential interests; the latter heavily relies on manually designed meta-paths, which are difficult to optimize in practice. To address this issue, this invention proposes a knowledge graph-based recommendation algorithm that incorporates preference propagation technology. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a technical resource recommendation algorithm based on the preference propagation method of knowledge graph, which solves the problems of lack of interpretability of embedding-based methods and difficulty in designing and optimizing meta-paths of path-based methods, thereby improving the quality of recommendation results.

[0007] Technical resources refer to papers, blogs, manuals, technical documents, etc., that can help users learn, understand, and work. In this article, technical resources refer to various text-based documents.

[0008] A method for recommending technical resources based on knowledge graph preference propagation includes the following steps:

[0009] Step S1: Based on the technology resource library, construct a technology resource knowledge graph in the "node-edge-weight" pattern using the LDA model;

[0010] Step S2: Preprocess the user's historical browsing data, introduce the time forgetting curve, and obtain the user's technical resource interaction matrix;

[0011] Step S3: Based on the technical resource knowledge graph and the user technical resource interaction matrix, the user's interest in each technical resource is calculated using preference propagation in the recommendation algorithm module, and a user interest model is constructed.

[0012] Step S4: Based on the user interest model, rank the technical resources by interest value, visualize the top-ranked technical resources and provide reasons for the recommendations.

[0013] As a further improvement of the present invention, in step S1, based on the technical resource library, the LDA model is used to extract the topics of the titles, abstracts and keywords of the technical resources, and the technical resources are used as a set of nodes. A set of relationship edges is established according to the frequency of occurrence and the topic association. The influence of the closeness of the relationship between technical resources is also taken into account, and the importance of the relationship is represented by weights. A technical resource knowledge graph in the "node-edge-weight" mode is constructed.

[0014] Step S1 specifically includes:

[0015] Step S1-1: Preprocess the technical resources to obtain a technical resource library;

[0016] Step S1-2: Use the LDA model to extract the set of technical resource nodes based on the topic keywords; in this process, the threshold is set to 0.25, and nodes greater than the threshold are added to the node set;

[0017] Steps S1-3: Calculate the relationship weights between nodes to obtain the relationship edge set and weight set; in this process, the threshold is set to 0.5. When the weight is greater than the threshold, the relationship is added to the edge set and the weight value is added to the corresponding weight set.

[0018] Steps S1-4: Construct a technical resource knowledge graph G in the "node-edge-weight" pattern based on the node set, relation edge set, and weight set.

[0019] As a further improvement of the present invention, in step S2, considering the issue of user interest changing over time, the longer the time since the user last read a certain technical resource, i.e., the longer the reading interval, the lower the user's level of interest in the technical resource. Based on the information of the time when the user last read the technical resource, a time forgetting curve function is introduced to calculate the user's interest decay, and different weight values ​​are assigned to technical resources with different reading intervals.

[0020] Step S2 specifically involves: using the data preprocessing module to process existing user browsing technology resource records and convert them into a user technology resource interaction matrix y. uv And input it into the recommendation algorithm module; this matrix is ​​initially set to 1 if the user has read the corresponding technical resource, and 0 otherwise; then the forgetting weight is calculated based on the forgetting curve over time, and the specific calculation formula is as follows:

[0021]

[0022] Where α is the forgetting coefficient and T is the reading time interval; different weight values ​​w are assigned to the viewed technical resources according to the time interval T of the user's browsing of the technical resources. uv Multiply the initial matrix by the forgetting weight w uv The user technical resource interaction matrix y was then obtained. uv .

[0023] As a further improvement of the present invention, in step S3, the recommendation algorithm module is characterized by combining the two methods, considering the lack of interpretability of embedding-based methods and the difficulty in designing and optimizing meta-paths in path-based methods. On the one hand, the embedding layer is used to calculate the embedding vectors of all technical resources, and their high-dimensional semantic information is used for similarity calculation and matching. On the other hand, a preference propagation method is introduced, allowing nodes to perform diffusion matching along neighboring nodes in the knowledge graph. Furthermore, the use of path information in the graph eliminates the need for manual design of meta-paths, thus improving the interpretability of the recommendation results.

[0024] Step S3 specifically includes:

[0025] Step S3-1: For a given user u, input the user's technical resource interaction matrix, and establish an initial set v for preference propagation in the knowledge graph. u This refers to the technical resources that the user has browsed;

[0026] Step S3-2: Establish a prediction objective function for user u and a certain technical resource v. That is, to calculate the degree of interest of user u in technical resource v, the embedding vector of technical resource v is directly calculated based on the embedding layer;

[0027] Step S3-3: Based on the initial set v on the knowledge graph u Preference propagation is performed by combining the weights of the relation edges to obtain the one-hop propagation set. Calculate the correlation between the embedding vectors of each technical resource and technical resource v in the propagation set, and sum them with weights to obtain the result of user interests after the first round of preference propagation. Repeat the preference propagation several times and sum the weighted values ​​to obtain the embedding vector of user u;

[0028] Step S3-4: Calculate the interest value of user u for a specific technical resource v based on the embedding vector of user u;

[0029] Step S3-5: Change v, traverse the technology resource library, and repeat steps 3-2, 3-3, and 3-4 to obtain the user u's interest model in the technology resource library.

[0030] As a further improvement of the present invention, in step S3, the preference propagation method searches for neighboring nodes of the current set in the knowledge graph. Combining the weights of the relational edges, different weights are assigned to neighboring nodes to represent the user's degree of interest preference. The similarity between the current technical resource and the technical resources of these neighboring nodes is calculated to reflect the user's potential interest in the current technical resource. To prevent the relevant set from becoming too large and generating redundant information after multiple preference propagations, firstly, the weight is reduced as the number of preference propagations increases; secondly, the number of propagation rounds is reduced, thereby limiting the size of the propagation set.

[0031] As a further improvement of the present invention, in step S4, the preference propagation paths of the top three technical resources in the knowledge graph are saved, and the reasons for recommending the technical resources are given, making it easier for users to accept the recommendation results.

[0032] As a further improvement to this invention, a solution is established to address the cold start problem. For new technology resources, they are added to a technology resource repository and included when building the technology resource knowledge graph. For new users, user registration information is introduced. The system automatically initializes the user's historical browsing information using the registration information, thus treating new users as "old users." Recommendations are made using the historical browsing information provided by the system, and then corrected and improved by the user during their usage, continuously improving the accuracy of the recommendations. On the other hand, considering the lack of historical browsing information for new users, the recommendation algorithm will place greater emphasis on utilizing the information in the technology resource knowledge graph, increasing the number of rounds of preference propagation while limiting the size of the neighbor set for preference propagation. Through multiple iterative learning solutions, the optimal recommendation result can be obtained. Furthermore, due to the increased number of propagation rounds, the user's potential interests can be discovered, and the user's interests and preferences can be determined more quickly, facilitating the establishment of a user interest model. This invention combines path-based and embedding-based algorithms, overcoming the limitations of embedding-based methods in terms of interpretability and path-based methods in terms of designing and optimizing meta-paths. By introducing preference propagation technology into the knowledge graph, it fully utilizes the high-dimensional semantic and path propagation information within the knowledge graph, improving not only the quality and accuracy of recommendation results but also making them more acceptable to users. Furthermore, this invention addresses the cold start problem and provides a solution.

[0033] In summary, this invention extracts topics based on the LDA model, considers the closeness of relationships between different nodes and represents them with weights, thus improving the practicality of the knowledge graph. During data preprocessing, it considers the changing interests of users over time by introducing a time forgetting curve, improving recommendation accuracy. In implementing the recommendation algorithm, it combines an embedding-based method, enabling the method to utilize the high-dimensional semantic information of the knowledge graph and facilitating implementation, with a path-based method, introducing path information from the knowledge graph through preference propagation, improving the interpretability of the recommendation results. Finally, this invention also provides a solution to the cold start problem, improving the robustness of the recommendation system. Attached Figure Description

[0034] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0035] Figure 1 This is an interaction diagram of the recommendation system modules of the present invention;

[0036] Figure 2 This is a flowchart of the recommended method steps of the present invention;

[0037] Figure 3 For the present invention Figure 2 A flowchart of the steps involved in constructing a knowledge graph in China;

[0038] Figure 4 For the present invention Figure 2 A flowchart of the steps involved in building a user interest model;

[0039] Figure 5 For the present invention Figure 2 A diagram illustrating the data interaction and calculation steps for displaying the recommendation results. Detailed Implementation

[0040] This section will further explain the invention with reference to the accompanying drawings.

[0041] Figure 1 This is an interaction diagram of the recommendation system modules of the present invention;

[0042] Figure 2 The flowchart of the recommendation method of the present invention is shown in the figure. The technical resource recommendation algorithm based on the knowledge graph-based preference propagation method of the present invention includes the following steps:

[0043] Step S1: Based on the technology resource library, construct a technology resource knowledge graph in the "node-edge-weight" pattern using the LDA model;

[0044] Step S2: Preprocess the user's historical browsing data, introduce the time forgetting curve, and obtain the user's technical resource interaction matrix;

[0045] Step S3: Based on the technical resource knowledge graph and the user technical resource interaction matrix, the user's interest in each technical resource is calculated using preference propagation in the recommendation algorithm module, and a user interest model is constructed.

[0046] Step S4: Based on the user interest model, rank the technical resources by interest value, visualize the top-ranked technical resources and provide reasons for the recommendations.

[0047] like Figure 3 As shown, in this invention, step S1 specifically includes: step S1-1, preprocessing the technical resources, using the titles, abstracts and keywords of technical resources from CNKI between 2010 and 2020 as a technical resource set, and processing them through jieba word segmentation, stop word removal and word form restoration to obtain a technical resource library;

[0048] Step S1-2: Use the LDA model to extract a set of technical resource nodes based on topic keywords. In this process, a threshold of 0.25 is set; nodes exceeding this threshold are added to the node set.

[0049] Steps S1-3: Calculate the relationship weights between nodes to obtain the relationship edge set and weight set. In this process, a threshold of 0.5 is set; when a weight is greater than the threshold, the relationship is added to the edge set and the weight value is added to the corresponding weight set.

[0050] Steps S1-4: Construct a technical resource knowledge graph G in the "node-edge-weight" pattern based on the node set, relation edge set, and weight set.

[0051] In this invention, step S2 involves using a data preprocessing module to process existing user browsing technology resource records and convert them into a user technology resource interaction matrix y. uv And input it into the recommendation algorithm module. This matrix is ​​initially set to 1 if the user has viewed the corresponding technical resource, and 0 otherwise. Then, the forgetting weight is calculated based on the forgetting curve over time; the specific calculation formula is as follows:

[0052]

[0053] Where α is the forgetting coefficient and T is the reading time interval. Different weight values ​​w are assigned to the viewed technical resources based on the time interval T between the user's browsing of the resources. uv Multiply the initial matrix by the forgetting weight w. uv The user technical resource interaction matrix y was then obtained. uv .

[0054] like Figure 4 As shown, in this invention, step S3 specifically includes:

[0055] Step S3-1: For a given user u, input the user's technical resource interaction matrix, and establish an initial set v for preference propagation in the knowledge graph. u This refers to the technical resources that the user has browsed;

[0056] Step S3-2: Establish a prediction objective function for user u and a certain technical resource v. That is, to calculate the degree of interest of user u in technical resource v, the embedding vector of technical resource v is directly calculated based on the embedding layer;

[0057] Step S3-3: Based on the initial set v on the knowledge graph u Preference propagation is performed by combining the weights of the relation edges to obtain the one-hop propagation set. Calculate the correlation between the embedding vectors of each technical resource and technical resource v in the propagation set, and sum them with weights to obtain the result of user interests after the first round of preference propagation. Repeat the preference propagation several times and sum the weighted values ​​to obtain the embedding vector of user u;

[0058] Step S3-4: Calculate the interest value of user u for a specific technical resource v based on the embedding vector of user u;

[0059] Step S3-5: Change v, traverse the technology resource library, and repeat steps S3-2, S3-3, and S3-4 to obtain the user u's interest model in the technology resource library.

[0060] In this invention, step S3-1 involves defining a set of related technical resource entities based on the user technical resource interaction matrix:

[0061]

[0062] Where G represents the technical resource knowledge graph, (h,r,t) represents the triple in the knowledge graph, h represents the head entity, r represents the relation, and t represents the tail entity. This is a set of technical resource entities. When k=0, it represents technical resources that the user has browsed; when k=1, it represents the tail entity of a triple with the technical resources browsed as the head entity, which is the technical resource entity obtained after one round of preference propagation; and so on. It refers to the technical resources obtained after k rounds of preference propagation.

[0063] Define the propagation set:

[0064]

[0065] set This means that the triplet is obtained after k preference propagations.

[0066] In this invention, step S3-2 involves establishing a prediction objective function. Θ represents the degree of interest of user u in technical resource v, and Θ represents the embedding vector of all technical resource entities and relationships calculated based on the technical resource knowledge graph.

[0067] In this invention, the main purpose of steps S3-3 is to calculate the embedding vector of user u based on the existing technical resource v's embedding. First, the one-hop propagation set after the first preference propagation is obtained. For the i-th triplet (h) in the set i ,r i ,t i ), calculate the head entity h. i and relation r i The correlation with a given technical resource v is calculated using the following formula:

[0068]

[0069] Where h i For a single jump propagation set The head entity of the i-th triple, r i This relates to the corresponding triplet. The head entity h... i With relation r iMultiply by the product, then multiply by the technical resource v to obtain the relevance score. Then, normalize the relevance score using the softmax function to obtain the relevance p. i .

[0070] The formula for calculating the result of user interests after the first round of preference propagation is as follows:

[0071] Where t i For a single jump propagation set The tail entity of the i-th triple. With correlation p i For all t, the weights are... i The weighted summation yields the results of one round of dissemination of user interests.

[0072] The formula for calculating the embedding vector of user u is defined in this invention as follows:

[0073]

[0074] in The result of a round of dissemination that represents user interests. The result of the second round of propagation, representing user interest, is calculated similarly. Here, to limit the size of the propagation set, only two rounds are propagated, and the results of the two rounds are added together to obtain u.

[0075] Within this invention, steps S3-4 involve calculating the user u's interest value in a specific technical resource v based on the user u's embedding vector. The formula for calculating the target function is as follows: The final prediction function value is calculated using the sigmoid function of the inner product of u and v.

[0076] like Figure 5As shown, in this invention, step S4, after obtaining the user interest model, proceeds to the recommendation result display module. The module ranks the technical resources based on the user's interest value and visualizes the top ten resources, providing corresponding titles, links, etc. The top three technical resources can be displayed with more detail and corresponding recommendation reasons. These reasons originate from paths in the knowledge graph, which can be technical resources the user has previously read. This invention also provides a solution to the cold start problem. For new technical resources, they are added to the technical resource library and included when building the technical resource knowledge graph. For new users, user registration information is introduced. The system automatically initializes the user's browsing history using the registration information, thus treating new users as "old users." Recommendations are made using the system's provided browsing history, and then corrected and improved by the user during usage, continuously improving the accuracy of the recommendations. On the other hand, considering the lack of historical browsing information for new users, recommendation algorithms will pay more attention to utilizing technical resource knowledge graph information, increase the number of rounds of preference propagation and limit the size of the neighbor set for preference propagation, and obtain the optimal recommendation result through multiple iterations of learning and solving. Also, due to the increase in the number of propagations, the potential interests of users can be discovered, and the user's interests and preferences can be determined more quickly, which is convenient for building a user interest model.

[0077] This invention combines path-based and embedding-based algorithms, overcoming the limitations of embedding-based methods in terms of interpretability and path-based methods in terms of designing and optimizing meta-paths. By introducing preference propagation technology into the knowledge graph, it fully utilizes the high-dimensional semantic and path propagation information within the knowledge graph, improving not only the quality and accuracy of recommendation results but also making them more acceptable to users. Furthermore, this invention addresses the cold start problem and provides a solution.

Claims

1. A method for recommending technical resources based on knowledge graph preference propagation, comprising the following steps: Step S1: Based on the technology resource library, construct a technology resource knowledge graph in the "node-edge-weight" pattern using the LDA model; Step S2: Preprocess the user's historical browsing data, introduce the time forgetting curve, and obtain the user's technical resource interaction matrix; Step S3: Based on the technical resource knowledge graph and the user technical resource interaction matrix, the user's interest in each technical resource is calculated using preference propagation in the recommendation algorithm module, and a user interest model is constructed. Step S4: Based on the user interest model, rank the technical resources according to their interest values, visualize the top-ranked technical resources, and provide reasons for the recommendations; In step S3, the recommendation algorithm module is characterized in that, Considering the lack of interpretability of embedding-based methods and the difficulty in designing and optimizing meta-paths in path-based methods, this paper combines the two approaches. On the one hand, the embedding layer is used to calculate the embedding vectors of all technical resources, and their high-dimensional semantic information is used for similarity calculation and matching. On the other hand, a preference propagation method is introduced, allowing nodes to perform diffusion matching along neighboring nodes in the knowledge graph. Furthermore, the use of path information in the graph eliminates the need for manual design of meta-paths, thus improving the interpretability of the recommendation results. Step S3 specifically includes: Step S3-1, for a specific user Input the user's technical resource interaction matrix to establish an initial set for preference propagation in the knowledge graph. This refers to the technical resources that the user has browsed; Step S3-2, for users and a certain technical resource Establish the prediction objective function That is, calculating users For technical resources The degree of interest is directly calculated based on the embedding layer to obtain the technical resources. The embedding vector; Step S3-3: Based on the initial set on the knowledge graph Preference propagation is performed by combining the weights of relation edges to obtain the one-hop propagation set. Computational propagation focuses on each technical resource and technical resources The embedded vectors are correlated and weighted to obtain the user interests after the first round of preference propagation. Repeat several rounds of preference propagation and weighted summation to obtain user preferences. The embedding vector; Step S3-4, according to the user Embedded vector computation for specific technical resources Interest value; Step S3-5, Change Traverse the technical resource library, repeating steps 3-2, 3-3, and 3-4 to obtain the user... Interest model for technology resource repositories; The purpose of step S3-3 is to utilize existing technical resources The embedding is calculated for the user The embedding vector; first, obtain the one-hop propagation set after the first preference propagation. For the set of the first Three pairs Calculate the head entity within it. and relationships The correlation with a given technical resource v is calculated using the following formula: in For a single jump propagation set The Middle The head entity of a triplet, For the corresponding triple relationship; the head entity With Relationship Multiply, and then combine with technological resources The correlation scores are obtained by multiplying the results, and then normalized using the softmax function to obtain the final correlation score. ; The formula for calculating the result of user interests after the first round of preference propagation is as follows: in For a single jump propagation set The Middle Tail entities of triples; with correlation For weights, for all The weighted summation yields the results of one round of dissemination of user interests; user The formula for calculating the embedding vector is defined as follows: in The result of a round of dissemination that represents user interests. The calculation method for the second-round propagation result representing user interests is similar; here, to limit the size of the propagation set, only two rounds of propagation are performed, and the results of the two rounds are added together. .

2. The recommended method as described in claim 1, characterized in that, In step S1, based on the technology resource library, the LDA model is used to extract topics from the titles, abstracts, and keywords of the technology resources. The technology resources are treated as a set of nodes, and a set of relationship edges is established based on the frequency of occurrence and topic association. The influence of the closeness of relationships between technology resources is also considered, and weights are used to represent the importance of relationships, thus constructing a technology resource knowledge graph in a "node-edge-weight" pattern. Step S1 specifically includes: Step S1-1: Preprocess the technical resources to obtain a technical resource library; Step S1-2: Use the LDA model to extract the set of technical resource nodes based on the topic keywords; in this process, the threshold is set to 0.25, and nodes greater than the threshold are added to the node set; Steps S1-3: Calculate the relationship weights between nodes to obtain the relationship edge set and weight set; in this process, the threshold is set to 0.

5. When the weight is greater than the threshold, the relationship is added to the edge set and the weight value is added to the corresponding weight set. Steps S1-4: Construct a technical resource knowledge graph in a "node-edge-weight" pattern based on the node set, relation edge set, and weight set. .

3. The recommended method as described in claim 1, characterized in that, In step S2, considering the issue of user interests changing over time, the longer the time since the last time a user read a certain technical resource is away from the current time, i.e., the longer the reading interval, the lower the user's interest in the technical resource; based on the information of the time when the user last read the technical resource, a time forgetting curve function is introduced to calculate the user's interest decay, and different weight values ​​are assigned to technical resources with different reading intervals; Step S2 specifically involves: using the data preprocessing module to process existing user browsing technology resource records and convert them into a user technology resource interaction matrix. And input it into the recommendation algorithm module; this matrix is ​​initially set to 1 if the user has read the corresponding technical resource, and 0 otherwise; then the forgetting weight is calculated based on the forgetting curve over time, and the specific calculation formula is as follows: in The forgetting factor is T, where T is the reading time interval. Different weight values ​​are assigned to the viewed technical resources based on the time interval T during which the user browses them. ; Multiply the initial matrix by the forgetting weight. The user technical resource interaction matrix was subsequently obtained. .

4. The recommended method as described in claim 1, characterized in that, In step S3-1, based on the user technical resource interaction matrix, a set of related technical resource entities is defined: ； in Represents a knowledge graph of technical resources. This represents a triple in a knowledge graph. Indicates the head entity. Indicates a relationship. Indicates the tail entity; This is a collection of technical resource entities. When k=0, it represents technical resources that the user has browsed; when k=1, it represents the tail entity of a triple with the technical resources browsed as the head entity, which is the technical resource entity obtained after one round of preference propagation; and so on. These are the technical resources obtained after k rounds of preference propagation. Define the propagation set: ； set This means that the triplet is obtained after k preference propagations.

5. The recommended method as described in claim 1, characterized in that, In step S3-2, the prediction objective function is established. , Indicates user For technical resources The degree of interest This represents the embedding vector of all technical resource entities and relationships calculated based on the technical resource knowledge graph.

6. The recommended method as described in claim 1, characterized in that, Steps S3-4, according to the user Embedded vector computation for specific technical resources The interest value, and the formula for calculating the prediction objective function are: The final prediction function value is obtained through... and The inner product is calculated using the sigmoid function.

Images