An information mining method and device and an information recommendation method and device
By optimizing the edge activation vector and random walk of heterogeneous graphs, the problems of excessive scale of heterogeneous graphs and unreasonable manual weighting of edges are solved, the information processing efficiency and user experience are improved, and more accurate information recommendation is achieved.
Patent Information
- Application Number
- CN201910417882.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2019-05-20
- Publication Date
- 2024-12-24
- Estimated Expiration
- 2039-05-20
AI Technical Summary
When handling heterogeneous patterns, the manual weight of edges is unreasonable, resulting in low noise introduction and operation efficiency. It is difficult to effectively carry out graph embedding learning on large-scale heterogeneous patterns, which affects the recommendation effect.
By obtaining the edge activation vector in the heterogeneous graph, using the fitness function to optimize the edge activation vector with task characteristics, and randomly walk on the optimized heterogeneous graph to generate the node sequence required for the graph embedding algorithm, directly Obtain a graph embedded representation of specific task characteristics from heterogeneous graphs.
It significantly reduces the running time, compresses the scale of heterogeneous graphs, avoids the irrationality of manually specified edge weights and the introduction of invalid edges, and improves the accuracy of user experience and recommendations.
Smart Images
Figure CN111966813B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to but is not limited to data mining technology, and in particular to an information mining method and device and an information recommendation method and device. Background Art
[0002] Personalized recommendation algorithms are used to build user profiles by mining user behavior data within business scenarios, and then recommend relevant content to users based on the user profiles.
[0003] Only when users have relatively rich behavioral data in the corresponding scenarios can the constructed user portrait be relatively comprehensive and accurate. When the user's behavior in this scenario is relatively sparse, the effect of the personalized recommendation algorithm will be greatly reduced. Take the application scenario where the user has never read anything through an APP as an example. In this case, the general strategy is that the APP will recommend popular content that is irrelevant to the user, and these recommended contents are not necessarily of interest to the user, which will greatly reduce the user's user experience in this scenario.
[0004] In addition to the behavior data in specific scenarios, there are also various behavior data in the system of large products, such as product behavior data, etc. These behavior data represent the user's interests from different angles. If the behavior data from multiple sources in the product can be integrated to supplement the single behavior data in a specific scenario, then the user's preferences can be captured more deeply.
[0005] When integrating and mining behavioral data from multiple sources, since the relationships between these behavioral data are complex and diverse, it is very suitable to characterize these behavioral data by constructing a heterogeneous graph; subsequently, the relevant graph embedding algorithm is used to mine on the constructed heterogeneous graph. In this case, the user representation obtained from the heterogeneous graph integrates multiple behavioral data, and also lays a good foundation for more accurate recommendation information to users in the future. Among them, a heterogeneous graph is a type of graph, which is a graph with more than one type of node or edge; correspondingly, a homogeneous graph is a type of graph, specifically a graph with only one type of node and edge. Graph embedding is a graph algorithm that represents the nodes of a graph as a multi-dimensional vector representation.
[0006] In related technologies, methods for graph embedding learning on heterogeneous graphs include:
[0007] One is to first convert the heterogeneous graph into a homogeneous graph, that is, by assigning different weights to different types of edges on the heterogeneous graph, and then using the weighted summation method to normalize the different types of edges between nodes into a single weight, thereby converting the heterogeneous graph into a homogeneous graph; then, use a homogeneous graph embedding algorithm, such as deep walk, to learn. Deep walk learns the social representation of graph's vertices. Deep walk improves random walk and is a shortened stream random walk method. In this method of graph embedding learning on heterogeneous graphs, the weights corresponding to different edges are manually specified, which is not only unreasonable, but also some edges are invalid for recommendation tasks in specific scenarios, but are still retained, thereby introducing noise and directly affecting the application effect.
[0008] The other is the traditional heterogeneous graph embedding algorithm, which either requires artificial prior knowledge to formulate path clusters to guide random walks, or faces the challenge of operating efficiency in walking on large-scale heterogeneous graphs. Summary of the invention
[0009] The present application provides an information mining method and device and an information recommendation method and device, which can improve the efficiency of information processing and enhance user experience.
[0010] An embodiment of the present invention provides an information mining method, including:
[0011] Get edge activation vectors in heterogeneous graphs;
[0012] The edge activation vector obtained is optimized according to at least the degree of association between two nodes in the heterogeneous graph to obtain an edge activation vector with task characteristics;
[0013] Random walks are performed on the heterogeneous graph after edge activation vector optimization to generate the node sequence required for graph embedding algorithm learning; based on the node sequence generated by random walks, a graph embedding representation with specific task characteristics is learned.
[0014] In an exemplary embodiment, obtaining an edge activation vector in a heterogeneous graph includes:
[0015] According to different types of edges of the heterogeneous graph, k binary heterogeneous edge activation vectors are initialized; wherein the value of k is related to the complexity of the heterogeneous graph, and the more complex the heterogeneous graph is, the more types of edges there are, and the larger the k value is.
[0016] In an exemplary embodiment, the degree of association between two nodes in the heterogeneous graph is obtained by a fitness function for evaluating the suitability of different edge activation vectors in the heterogeneous graph.
[0017] In an exemplary embodiment, optimizing the obtained edge activation vector according to the fitness function includes:
[0018] Using the fitness function, respectively calculating fitness function values corresponding to p different edge activation vectors on a specific task; wherein each edge activation vector includes k values;
[0019] Select p / 2 edge activation vector pairs proportionally to the fitness function value;
[0020] Perform feature crossover processing on the selected p / 2 edge activation vector pairs; then perform gene mutation processing to generate p new edge activation vectors;
[0021] The step of returning to calculating the fitness function value is continued until convergence, and the edge activation vector that produces the best output is selected as the edge activation vector having the task characteristics.
[0022] In an exemplary embodiment, a graph algorithm is used to obtain the degree of association between two nodes in the heterogeneous graph.
[0023] The present application also provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the above-mentioned information mining methods.
[0024] The present application further provides a device for implementing information mining, including a memory and a processor, wherein the memory stores the following instructions executable by the processor: used to execute the steps of any of the above-mentioned information mining methods.
[0025] The present application also provides an information recommendation method, comprising:
[0026] Obtaining edge activation vectors from the heterogeneous graph, optimizing the obtained edge activation vectors at least according to the degree of association between two nodes in the heterogeneous graph, and obtaining edge activation vectors with task characteristics;
[0027] Perform random walks on the heterogeneous graph after edge activation vector optimization to generate the node sequence required for graph embedding algorithm learning; learn the graph embedding representation of specific task characteristics based on the node sequence generated by random walks;
[0028] A user profile is constructed based on the learned graph embedding representation of specific task characteristics, and information is recommended to the user based on the constructed user profile.
[0029] In an exemplary embodiment, the degree of association between two nodes in the heterogeneous graph is obtained by a fitness function for evaluating the suitability of different edge activation vectors in the heterogeneous graph.
[0030] In an exemplary embodiment, optimizing the obtained edge activation vector at least according to a fitness function includes:
[0031] Using the fitness function, respectively calculating the fitness function values corresponding to p different edge activation vectors on the specific task; wherein each edge activation vector includes k values; the value of k is related to the complexity of the heterogeneous graph, and the more complex the heterogeneous graph is, the more types of edges there are, and the larger the k value is;
[0032] Select p / 2 edge activation vector pairs proportionally to the fitness function value;
[0033] Perform feature crossover processing on the selected p / 2 edge activation vector pairs; then perform gene mutation processing to generate k new edge activation vectors;
[0034] The step of returning to calculating the fitness function value is continued until convergence, and the edge activation vector that produces the best output is selected as the edge activation vector having the task characteristics.
[0035] The present application further provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the above-mentioned information recommendation methods.
[0036] The present application further provides a device for implementing information recommendation, including a memory and a processor, wherein the memory stores the following instructions executable by the processor: used to execute the steps of any of the above-mentioned information recommendation methods.
[0037] The application simplifies large-scale heterogeneous graphs based on edge activation vectors, and the scale of heterogeneous graphs is greatly compressed. This application overcomes the problems of low efficiency of running directly on heterogeneous graphs due to the large scale of heterogeneous graphs, the irrationality of manually specifying edge weights during the conversion of heterogeneous graphs to homogeneous graphs, and the introduction of invalid edges. In other words, the information mining method of this application significantly reduces the running time and achieves better results than related technical methods, thereby improving the user experience.
[0038] The information recommendation method of the present application ensures more accurate information recommendation, thereby improving the user experience, because the user representation obtained directly from the heterogeneous graph integrates a variety of behavioral data information.
[0039] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] The accompanying drawings are used to provide further understanding of the technical solution of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application and do not constitute a limitation on the technical solution of the present application.
[0041] Figure 1 A flowchart of the information mining method for this application;
[0042] Figure 2 Flowchart of the recommended approach for this application information. DETAILED DESCRIPTION
[0043] In order to make the purpose, technical solution and advantages of the present application more clear, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other arbitrarily without conflict.
[0044] In a typical configuration of the present application, a computing device includes one or more processors (CPU), an input / output interface, a network interface, and a memory.
[0045] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.
[0046] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include non-transitory media such as modulated data signals and carrier waves.
[0047] The steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions. Also, although a logical sequence is shown in the flowchart, in some cases, the steps shown or described can be performed in a sequence different from that shown here.
[0048] Figure 1 The flowchart of the information mining method for this application is as follows: Figure 1 As shown, including:
[0049] Step 100: Obtain edge activation vectors in the heterogeneous graph.
[0050] In an exemplary embodiment, this step may include:
[0051] According to the different types of edges in the heterogeneous graph, k binary heterogeneous edge activation vectors are initialized. For example, if there are 20 different types of edges on the heterogeneous graph, k 20-dimensional vectors can be constructed by initialization, where each dimension in the vector has a value of either 0 or 1, and the value 0 or 1 is randomly selected at the initialization.
[0052] The value of k is a variable parameter, which is related to the complexity of the heterogeneous graph. The more complex the heterogeneous graph is and the more types of edges it has, the larger the value of k is.
[0053] Step 101: Optimize the obtained edge activation vector at least according to the degree of association between two nodes in the heterogeneous graph to obtain an edge activation vector with task characteristics.
[0054] In an exemplary embodiment, the degree of association between two nodes in the heterogeneous graph may be obtained by a fitness function for evaluating the suitability of different edge activation vectors in the heterogeneous graph.
[0055] In an exemplary embodiment, the fitness function is also used to evaluate the suitability of different edge activation vectors in a heterogeneous graph, and may include the formula (1):
[0056]
[0057] In formula (1), i and j represent two nodes on the heterogeneous graph, z represents the common neighbors of node i and node j in the heterogeneous graph, n and m represent the specific types of edges; C is the candidate edge activation vector, e is an edge of a certain type on the heterogeneous graph, ω(e) is the weight of the edge type e; E C (ν) represents the set of edges currently valid for node ν under the given edge activation vector C; E C i,j represents the set of valid edges between node i and node j under the given edge activation vector C.
[0058] The fitness function shown in formula (1) well characterizes the degree of effective information retention between nodes. By calculating the fitness function values corresponding to p different edge activation vectors on a specific task, p / 2 edge activation vector pairs are selected proportionally according to the fitness function values. The higher the fitness, the greater the probability of being selected. Each edge activation vector includes k values.
[0059] In an exemplary embodiment, optimizing the obtained edge activation vector according to the fitness function in this step includes:
[0060] Calculate the fitness function values corresponding to p different edge activation vectors on the specific task respectively; each edge activation vector includes k values;
[0061] Select p / 2 edge activation vector pairs proportionally to the fitness function value;
[0062] The selected p / 2 edge activation vector pairs are subjected to feature crossover processing, that is, the values of the two activation vectors in a certain position are replaced with a certain probability; then, gene mutation processing (a technical term in genetic algorithms) is performed, that is, the value of a certain position of the activation vector is reversed with a certain probability to produce p new edge activation vectors.
[0063] The step of returning to calculating the fitness function value continues until convergence, such as the number of repetitions reaches a certain set value, and the edge activation vector that produces the best output (such as the edge activation vector with the largest fitness function value) is selected as the edge activation vector with task characteristics.
[0064] In an exemplary embodiment, in the above process of optimizing the edge activation vector obtained according to the fitness function, the specific task may be: a task in a certain scenario is to recommend relevant articles to users, and the user's operation data on the article (such as reading, liking, and forwarding) can be used as the input of the algorithm under this task.
[0065] In an exemplary instance, in the above-mentioned process of optimizing the obtained edge activation vector according to the fitness function, the proportional selection may include, for example: when selecting the edge activation vector, the selection may be made based on the fitness value calculated from the edge activation vector, for example, the higher the fitness, the greater the chance that the edge activation vector will be selected, and so on.
[0066] In an exemplary embodiment, in the process of optimizing the edge activation vector obtained according to the fitness function, the certain probability may be a probability value set in advance by the user; and the certain position may be a randomly selected position. In an exemplary embodiment, the value of a certain position of the activation vector is reversed according to a certain probability, that is, a position is randomly selected, and the values of the two activation vectors at this position are replaced with a fixed probability.
[0067] In an exemplary embodiment, this step may include:
[0068] The degree of association between two nodes in the heterogeneous graph can be obtained by using graph algorithms in related technologies such as SimRank.
[0069] In an exemplary embodiment, in this step, in optimizing the edge activation vector obtained according to the graph algorithm in the related art, the genetic algorithm is no longer used to learn the edge activation vector. The edge activation vector is used as a part of the training parameters in the graph algorithm and is directly learned using the graph algorithm.
[0070] This step uses edge activation vectors to simplify and compress large-scale heterogeneous graphs, deletes some edges and nodes that are invalid for subsequent tasks, and simplifies the scale of the graph to different degrees depending on the activation vectors. In this way, the simplified heterogeneous graph can be connected to the graph embedding learning method on heterogeneous graphs in related technologies, ensuring the application effect while improving the operation efficiency.
[0071] Step 102: Perform a random walk on the heterogeneous graph after the edge activation vector is optimized to generate a node sequence required for learning the graph embedding algorithm; and learn a graph embedding representation of specific task characteristics based on the node sequence generated by the random walk.
[0072] The specific implementation of this step can adopt the method provided in the related art, and the specific implementation is not used to limit the protection scope of this application. It is emphasized here that the heterogeneous graph at this time is a heterogeneous graph simplified by the edge activation vector of this application.
[0073] The large-scale heterogeneous graph is simplified based on the edge activation vector in this application, and the scale of the heterogeneous graph is greatly compressed. It can be seen from practical applications that the total number of edges in the heterogeneous graph can be simplified by about 76%, and the total number of nodes can be simplified by about 65%. At the same time, this application overcomes the problems of low efficiency of running directly on the heterogeneous graph due to the large scale of the heterogeneous graph, the irrationality of manually specifying edge weights during the conversion of heterogeneous graphs to homogeneous graphs, and the introduction of invalid edges. In other words, the information mining method of this application has achieved better results than related technical methods while significantly reducing the running time, thereby improving the user experience.
[0074] The information mining method provided in this application can be used to implement subsequent sorting tasks and information recommendations based on the mined information.
[0075] The present application also provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the above-mentioned information mining methods.
[0076] The present application further provides a device for implementing information mining, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor; when the computer program is executed by the processor, the steps of any of the above-mentioned information mining methods are implemented.
[0077] Figure 2 A flowchart of the recommended method for this application information, such as Figure 2 As shown, including:
[0078] Step 200: Obtain an edge activation vector in the heterogeneous graph, and optimize the obtained edge activation vector at least according to the degree of association between two nodes in the heterogeneous graph to obtain an edge activation vector with task characteristics.
[0079] The specific implementation of this step is as described in steps 100 to 101, and will not be repeated here.
[0080] Step 201: Perform a random walk on the heterogeneous graph after edge activation vector optimization to generate a node sequence required for graph embedding algorithm learning; learn a graph embedding representation of specific task characteristics based on the node sequence generated by the random walk.
[0081] Step 202: construct a user profile based on the learned graph embedding representation of specific task characteristics, and recommend information to the user based on the constructed user profile.
[0082] The information recommendation method of the present application ensures more accurate information recommendation, thereby improving the user experience, because the user representation obtained directly from the heterogeneous graph integrates a variety of behavioral data information.
[0083] The present application also provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the above-mentioned information recommendation methods.
[0084] The present application further provides a device for implementing information recommendation, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor; when the computer program is executed by the processor, the steps of any of the above-mentioned information recommendation methods are implemented.
[0085] Although the embodiments disclosed in this application are as above, the contents described are only embodiments adopted to facilitate understanding of this application and are not intended to limit this application. Any technician in the field to which this application belongs can make any modifications and changes in the form and details of implementation without departing from the spirit and scope disclosed in this application, but the scope of patent protection of this application shall still be based on the scope defined in the attached claims.
Claims
1. An information mining method, comprising: Get edge activation vectors in heterogeneous graphs; Optimizing the edge activation vector obtained at least according to the degree of association between two nodes in the heterogeneous graph to obtain an edge activation vector with task characteristics; Performing random walks on the heterogeneous graph after edge activation vector optimization to generate a node sequence required for graph embedding algorithm learning; learning a graph embedding representation of specific task characteristics based on the node sequence generated by the random walk; The edge activation vector is used to indicate a vector having the same dimension as the number of edge types of the heterogeneous graph; The step of optimizing the edge activation vector obtained at least according to the degree of association between two nodes in the heterogeneous graph includes: performing feature crossover processing and gene mutation processing on the edge activation vector according to the degree of association between two nodes in the heterogeneous graph; The specific task characteristics are used to indicate recommendation of article information to a user, and the user's operation data on the article information is used to generate the graph embedding representation.
2. The information mining method according to claim 1, wherein: The obtaining of the edge activation vector in the heterogeneous graph includes: According to different types of edges of the heterogeneous graph, k binary heterogeneous edge activation vectors are initialized; wherein the value of k is related to the complexity of the heterogeneous graph, and the more complex the heterogeneous graph is, the more types of edges there are, and the larger the k value is.
3. The information mining method according to claim 1, wherein: The degree of association between two nodes in the heterogeneous graph is obtained by using a fitness function for evaluating the suitability of different edge activation vectors in the heterogeneous graph.
4. The information mining method according to claim 3, wherein: Optimizing the obtained edge activation vector according to the fitness function includes: Using the fitness function, respectively calculating fitness function values corresponding to p different edge activation vectors on a specific task; wherein each edge activation vector includes k values; Selecting p / 2 edge activation vector pairs in proportion to the fitness function value; Performing the feature crossover process within the selected p / 2 edge activation vector pairs; and then performing the gene mutation process to generate p new edge activation vectors; The step of returning to calculating the fitness function value is continued until convergence, and the edge activation vector with the best output is selected as the edge activation vector with task characteristics.
5. The information mining method according to claim 2, wherein: A graph algorithm is used to obtain the degree of association between two nodes in the heterogeneous graph.
6. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the information mining method described in any one of claims 1 to claim 5.
7. A device for implementing information mining, comprising a memory and a processor, wherein: The memory stores the following instructions executable by the processor: used to execute the steps of the information mining method described in any one of claims 1 to claim 5.
8. An information recommendation method, comprising: Obtaining an edge activation vector in a heterogeneous graph, and optimizing the obtained edge activation vector at least according to a degree of association between two nodes in the heterogeneous graph to obtain an edge activation vector with task characteristics; Performing random walks on the heterogeneous graph after edge activation vector optimization to generate a node sequence required for graph embedding algorithm learning; learning a graph embedding representation of specific task characteristics based on the node sequence generated by the random walk; Constructing a user profile based on the learned graph embedding representation of specific task characteristics, and recommending information to the user based on the constructed user profile; The edge activation vector is used to indicate a vector having the same dimension as the number of edge types of the heterogeneous graph; The step of optimizing the edge activation vector obtained at least according to the degree of association between two nodes in the heterogeneous graph includes: performing feature crossover processing and gene mutation processing on the edge activation vector according to the degree of association between two nodes in the heterogeneous graph; The specific task characteristics are used to indicate recommendation of article information to a user, and the user's operation data on the article information is used to generate the graph embedding representation.
9. The information recommendation method according to claim 8, wherein: The degree of association between two nodes in the heterogeneous graph is obtained by using a fitness function for evaluating the suitability of different edge activation vectors in the heterogeneous graph.
10. The information recommendation method according to claim 9, wherein: Optimizing the obtained edge activation vector at least according to the fitness function, comprising: Using the fitness function, respectively calculating the fitness function values corresponding to p different edge activation vectors on the specific task; wherein each edge activation vector includes k values; the value of k is related to the complexity of the heterogeneous graph, and the more complex the heterogeneous graph is, the more types of edges there are, and the larger the k value is; Selecting p / 2 edge activation vector pairs in proportion to the fitness function value; Performing the feature crossover process within the selected p / 2 edge activation vector pairs; and then performing the gene mutation process to generate k new edge activation vectors; The step of returning to calculating the fitness function value is continued until convergence, and the edge activation vector with the best output is selected as the edge activation vector with task characteristics.
11. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the information recommendation method according to any one of claims 8 to 10.
12. A device for implementing information recommendation, comprising a memory and a processor, wherein: The memory stores the following instructions executable by the processor: used to execute the steps of the information recommendation method described in any one of claims 8 to claim 10.