User group expansion method and device, equipment, storage medium and program product

Abstract

The invention relates to the technical field of computers, and provides a user group expansion method and device, equipment, a storage medium and a program product. The method comprises the following steps: acquiring internet behavior data of candidate users; modeling user behaviors based on the user information, the external link information, the time information and the base station information, and constructing a knowledge graph; inputting the knowledge graph into a deep learning model, carrying out low-dimensional representation learning, and obtaining an embedded vector of a candidate user output by the deep learning model; the embedded vectors are used for representing user behavior rules and spatio-temporal dynamic characteristics of the candidate users; and determining whether the candidate user is a target group expansion user of the seed user set based on the embedded vector. According to the mode, consideration of time information and space information is introduced in the user group expansion process, the spatial and temporal change of user behaviors can be accurately captured, judgment and classification of candidate users are more accurate, and then the group expansion accuracy can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular, to a method, apparatus, device, storage medium, and program product for expanding a user group. Background Art

[0002] In the fields of advertising placement and user portrait construction, user group expansion technology has been widely applied. User group expansion refers to, based on an existing seed user group, screening out users from the entire user population who have similar or identical characteristics to the seed user group, and adding the screened users to the seed user group to achieve the expansion or sampling of the seed user group.

[0003] Currently, user group expansion technology mainly includes a user group expansion method based on interest groups and a user group expansion method through frequent item set mining. However, the former ignores the spatio-temporal dynamic characteristics and individual differences of user behavior and cannot meet the needs of precise group expansion; the latter overly relies on frequent item set mining and rule learning, lacks the mining and modeling of the spatio-temporal dynamic characteristics of users, and is likely to affect the accuracy of group expansion.

[0004] Therefore, the existing user group expansion methods do not consider the influence of time information and space information, and it is difficult to accurately capture changes in user behavior, thereby resulting in low accuracy of group expansion. Summary of the Invention

[0005] Embodiments of this application provide a method, apparatus, device, storage medium, and program product for expanding a user group, so as to solve the technical problem that the existing user group expansion methods do not consider the influence of time information and space information, and it is difficult to accurately capture changes in user behavior, thereby resulting in low accuracy of group expansion.

[0006] In a first aspect, embodiments of this application provide a method for expanding a user group, including: obtaining the Internet access behavior data of candidate users; the Internet access behavior data includes user information, external link information, time information, and base station information; based on the user information, external link information, time information, and base station information, modeling user behavior to construct a knowledge graph; inputting the knowledge graph into a deep learning model for low-dimensional representation learning to obtain the embedding vector of the candidate user output by the deep learning model; the embedding vector is used to characterize the user behavior pattern and spatio-temporal dynamic characteristics of the candidate user; based on the embedding vector, determining whether the candidate user is a target user for expanding the seed user set.

[0007] In one embodiment, the knowledge graph is represented in the form of quadruples; before inputting the knowledge graph into a deep learning model for low-dimensional representation learning to obtain the embedding vectors of candidate users output by the deep learning model, it further includes: obtaining a set of positive quadruple samples and a set of negative quadruple samples; the set of positive quadruple samples includes multiple positive quadruple samples, and the set of negative quadruple samples includes multiple negative quadruple samples. The positive quadruple samples are real user samples, and the negative quadruple samples are non-real user samples; based on the set of positive quadruple samples and the set of negative quadruple samples, pre-train the initial model constructed based on the TransE algorithm to obtain a deep learning model.

[0008] In one embodiment, the deep learning model is pre-trained using a preset loss function, and the preset loss function is determined based on the distance between positive quadruple samples and negative quadruple samples; the training objective of the deep learning model is to minimize the distance between the embedding vectors of all positive quadruple samples and maximize the distance between the embedding vectors of all negative quadruple samples by optimizing the preset loss function.

[0009] In one embodiment, determining whether a candidate user is a target expanding user of a seed user set based on the embedding vectors includes: determining the average similarity between the candidate user and the seed user set based on the embedding vectors; the average similarity is the average of the similarities between the candidate user and each seed user in the seed user set; determining a similarity score between the candidate user and the seed user set based on the average similarity; and determining whether the candidate user is a target expanding user of the seed user set based on the similarity score.

[0010] In one embodiment, the number of candidate users is at least two; determining whether a candidate user is a target expanding user of a seed user set based on the similarity score includes: sorting all candidate users in descending order according to the similarity score of each candidate user to determine the similarity ranking of each candidate user; and selecting a preset number of candidate users as the target expanding users of the seed user set based on the similarity ranking of each candidate user.

[0011] In one embodiment, the knowledge graph includes user nodes, external link nodes, time nodes, and base station nodes; modeling user behavior based on user information, external link information, time information, and base station information to construct a knowledge graph includes: modeling user information, external link information, time information, and base station information as user nodes, external link nodes, time nodes, and base station nodes respectively; modeling the network access behavior of candidate users as edges, and connecting user nodes, external link nodes, time nodes, and base station nodes based on the edges to generate a knowledge graph.

[0012] Second aspect, an embodiment of the present application provides a user group expansion device, including: an acquisition module, configured to acquire the Internet behavior data of candidate users; the Internet behavior data includes user information, external link information, time information, and base station information; a knowledge graph construction module, configured to model user behaviors based on the user information, external link information, time information, and base station information, and construct a knowledge graph; a vector generation module, configured to input the knowledge graph into a deep learning model for low-dimensional representation learning, and obtain the embedding vector of the candidate user output by the deep learning model; the embedding vector is used to characterize the user behavior pattern and spatio-temporal dynamic characteristics of the candidate user; a user group expansion module, configured to determine whether the candidate user is a target user to be expanded in the seed user set based on the embedding vector.

[0013] Third aspect, an embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the computer program, it implements any one of the above user group expansion methods.

[0014] Fourth aspect, an embodiment of the present application provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements any one of the above user group expansion methods.

[0015] Fifth aspect, an embodiment of the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements any one of the above user group expansion methods.

[0016] In the user group expansion method, device, equipment, storage medium, and program product provided by the embodiments of the present application, time information and base station information are added during the user group expansion process, thereby introducing considerations of time information and space information. First, the user information, external link information, time information, and base station information of the candidate user are used to model user behaviors and construct a knowledge graph. Then, the knowledge graph is input into a deep learning model for low-dimensional representation learning to obtain an embedding vector that characterizes the user behavior pattern and spatio-temporal dynamic characteristics of the candidate user. Finally, based on the embedding vector, it is determined whether the candidate user is a target user to be expanded in the seed user set. Through the above method, considerations of time information and space information are introduced during the user group expansion process, which can accurately capture the spatio-temporal changes of user behaviors, make the judgment and classification of candidate users more accurate, and thus effectively improve the accuracy of group expansion. Description of the Drawings

[0017] To more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 It is one of the schematic flowcharts of the user group expansion method provided by the embodiments of the present application.

[0019] Figure 2 It is the second schematic flowchart of the user group expansion method provided by the embodiments of the present application.

[0020] Figure 3 It is the schematic structural diagram of the user group expansion device provided by the embodiments of the present application.

[0021] Figure 4 It is the schematic structural diagram of the electronic device provided by the embodiments of the present application. Specific embodiments

[0022] To make the objectives, technical solutions, and advantages of the present application clearer, the following will clearly and completely describe the technical solutions in the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts fall within the scope of protection of the present application.

[0023] Please refer to Figure 1 and Figure 2 , Figure 1 is one of the schematic flowcharts of the user group expansion method provided by the embodiments of the present application, Figure 2 is the second schematic flowchart of the user group expansion method provided by the embodiments of the present application. As Figure 1 shown, in the embodiments of the present application, the user group expansion method includes steps S110 to S140, and the specific steps are as follows: S110: Obtain the Internet behavior data of candidate users.

[0024] The Internet behavior data includes user information, external link information, time information, and base station information.

[0025] Specifically, as Figure 2As shown, using the operator's user access records and base station data, the Internet access behavior data of candidate users is comprehensively collected. These Internet access behavior data should cover core information such as the web pages visited by candidate users, the specific time of visiting web pages, the frequency of visiting web pages, and the duration of staying on web pages. At the same time, in order to further improve the accuracy of group expansion, it is also necessary to combine the web page content information of external links (i.e., external link information), such as web page titles, article abstracts, or key content fragments. For links with inaccessible content, only rely on their user behavior characteristics for modeling. Through the comprehensive integration of multi-source heterogeneous data, a basic data framework covering the interests and behavior habits of candidate users is constructed.

[0026] Optionally, after the Internet access behavior data is collected, to ensure the accuracy and efficiency of the subsequent modeling process, all Internet access behavior data needs to be cleaned and preprocessed.

[0027] Optionally, the steps of data cleaning mainly include operations such as removing duplicate records, eliminating noise data, and repairing data anomalies to ensure the integrity and consistency of the data.

[0028] S120: Based on user information, external link information, time information, and base station information, model user behavior and construct a knowledge graph.

[0029] After completing the data preprocessing, use the knowledge graph method to deeply model the Internet access behavior data of candidate users.

[0030] Specifically, model user information, external link information, time information, and base station information as user nodes, external link nodes, time nodes, and base station nodes respectively, and model the network access behavior of candidate users as edges. Use the edges as relationships to connect user nodes, external link nodes, time nodes, and base station nodes to generate a knowledge graph.

[0031] It should be noted that the knowledge graph constructed in the embodiments of this application is a multi-dimensional and dynamic network representation structure. The design of the knowledge graph not only carries the direct interaction information between candidate users and external links, but also further depicts the temporal changes and geographical distribution characteristics of users' Internet access behaviors through the dynamic features of time information and space information, thereby enhancing the data expression ability in terms of breadth and depth. The nodes of the knowledge graph are mainly divided into two categories: user nodes and external link nodes. User nodes represent each user in the operator, and the node features include the basic information of the user (such as user ID) and the behavioral features in the access records, such as statistical information such as access frequency and stay duration. External link nodes correspond to the web pages or links accessed by candidate users, and the features include web page content (such as title, abstract, keywords, etc.) or meta-information related to the web page (such as category, URL, etc.). For external links for which specific content cannot be obtained, feature representation is only based on the associated user behavior. The knowledge graph also includes time nodes and base station nodes: The time nodes use timestamps as units to quantitatively represent the time information of user access behaviors, such as the granularity of hours, days, and months; the base station information is further processed through position encoding technology so that it can effectively participate in vectorization modeling. Each base station node can reflect the geographical location information of users' Internet access behaviors, and the features of the base station nodes include the unique identifier of the base station and the geographical location encoding to which it belongs (such as longitude and latitude).

[0032] Optionally, the knowledge graph is represented in the form of quadruples. For example, the knowledge graph can be represented as (User A, time point, location, external link B). The time information and space information are embedded in the representation of the knowledge graph relationships, making each relationship have dynamic features. This design ensures that the knowledge graph can reflect the complex Internet access behavior patterns of users, such as the preference of users to access specific links at a certain specific time.

[0033] S130: Input the knowledge graph into a deep learning model for low-dimensional representation learning to obtain the embedding vector of the candidate user output by the deep learning model.

[0034] The embedding vector is used to characterize the user behavior patterns and spatio-temporal dynamic features of the candidate user.

[0035] In the embodiments of this application, the method of low-dimensional representation learning is based on graph embedding technology, which mainly learns the low-dimensional representations of each node and relationship in the knowledge graph based on the TransE model.

[0036] Specifically, the knowledge graph is input into a deep learning model constructed based on the TransE algorithm for low-dimensional representation learning. In the deep learning model, each node (such as user nodes, external link nodes, time nodes, and base station nodes) and relationship (such as network access behavior, spatio-temporal relationship) in the knowledge graph are mapped to vectors. Therefore, the deep learning model can finally output the embedding vectors of candidate users.

[0037] It should be noted that in the deep learning model, the representation of relationships needs to capture the mutual associations between nodes through vector addition. Since the knowledge graph input into the deep learning model is represented in the form of quadruples, the core idea of the deep learning model is to make the embedding vectors of the quadruple ) satisfy the following conditions: ; Among them, is the embedding vector of the user node, is the embedding vector of the external link node, is the embedding vector of the relationship (i.e., network access behavior), is the embedding vector of the time series (i.e., time node).

[0038] Optionally, in order to express time information more accurately, the embedding vector of the time node can draw on the position encoding method in the Transformer framework. Therefore, for the embedding vector of the time node, its encoding adopts the following formula: ; ; Among them, represents the even-bit encoding of the embedding vector of the time node, represents the odd-bit encoding of the embedding vector of the time node, represents the embedding dimension.

[0039] This encoding method can provide a periodic and global expression for time information in the low-dimensional space, and is used to capture the time dynamics of users' Internet access behaviors.

[0040] S140: Based on the embedding vectors, determine whether the candidate user is the target expansion user of the seed user set.

[0041] An embodiment of the present application proposes a user group expansion method based on operator user Internet behavior data, which is particularly applicable to the advertising placement scenario. By comprehensively analyzing user behavior data, external link content, and spatio-temporal information, accurate expansion of seed users is achieved, and dynamic expansion of the user group is completed by constructing and optimizing a knowledge graph. This method makes full use of the operator's massive data resources, including user access records, base station information, etc. Through joint modeling of multi-dimensional data, characteristic patterns of potential user groups can be effectively mined.

[0042] In the user group expansion method provided by the embodiment of the present application, time information and base station information are added during the user group expansion process, thereby introducing considerations of time information and space information. First, the user information, external link information, time information, and base station information of candidate users are used to model user behavior and construct a knowledge graph. Then, the knowledge graph is input into a deep learning model for low-dimensional representation learning to obtain an embedding vector representing the user behavior law and spatio-temporal dynamic characteristics of the candidate user. Finally, based on the embedding vector, it is determined whether the candidate user is a target expanded user of the seed user set. Through the above method, considerations of time information and space information are introduced during the user group expansion process, which can accurately capture the spatio-temporal changes of user behavior, make the judgment and classification of candidate users more accurate, and thus effectively improve the accuracy of group expansion.

[0043] In some embodiments, the knowledge graph is represented in the form of quadruples; before inputting the knowledge graph into a deep learning model for low-dimensional representation learning to obtain the embedding vector of the candidate user output by the deep learning model, it further includes: obtaining a positive quadruple sample set and a negative quadruple sample set; the positive quadruple sample set includes multiple positive quadruple samples, and the negative quadruple sample set includes multiple negative quadruple samples. The positive quadruple samples are real user samples, and the negative quadruple samples are non-real user samples; based on the positive quadruple sample set and the negative quadruple sample set, the initial model constructed based on the TransE algorithm is pre-trained to obtain a deep learning model.

[0044] It can be understood that before performing low-dimensional representation learning using a deep learning model, it is necessary to first train the deep learning model.

[0045] Specifically, obtain a positive quadruple sample set and a negative quadruple sample set.

[0046] The positive quadruple sample set includes multiple positive quadruple samples, and the negative quadruple sample set includes multiple negative quadruple samples.

[0047] Among them, the positive quadruple sample ) is a real user sample, representing real user access behavior; the negative quadruple sample ) is a non-genuine user sample, which is a meaningless combination generated by random sampling.

[0048] Furthermore, based on the positive quadruple sample set and the negative quadruple sample set, the initial model constructed based on the TransE algorithm is pre-trained to obtain a deep learning model.

[0049] In some embodiments, the deep learning model is pre-trained using a preset loss function, and the preset loss function is determined based on the distance between the positive quadruple samples and the negative quadruple samples; the training objective of the deep learning model is to minimize the distance between the embedding vectors of all positive quadruple samples and maximize the distance between the embedding vectors of all negative quadruple samples by optimizing the preset loss function.

[0050] In the embodiments of the present application, the deep learning model is pre-trained using a preset loss function, and the core objective of its training process is to minimize the distance between the embedding vectors of all positive quadruple samples and maximize the distance between the embedding vectors of all negative quadruple samples by optimizing the preset loss function. In this way, the deep learning model can effectively distinguish real user access behaviors from randomly generated negative quadruple samples, thereby learning the relationships between users and multi-dimensional features such as external links, time, and location. To achieve this goal, the design of the preset loss function adopts a typical contrastive learning method, and the formula of the preset loss function is as follows: ; where is the positive quadruple sample set, representing real user access behaviors, which includes the behavior patterns of sample users accessing a certain link; is the negative quadruple sample set, usually some pseudo data generated by random sampling, and these data represent untrue or meaningless user behaviors; is the distance between vectors, and usually the Euclidean distance is used to measure the distance between two vectors in the embedding space, which is defined as: ; where is a hyperparameter used to control the minimum distance boundary between the positive quadruple samples and the negative quadruple samples, that is, to control the interval between the positive quadruple samples and the negative quadruple samples. This hyperparameter is crucial for the convergence speed and effect of model training.

[0051] Through the preset loss function, the distance of the positive quadruple samples is gradually reduced during the training process, and the distance of the negative quadruple samples is increased as much as possible, and finally an embedding space that can effectively capture the relationship between user behaviors and external links is formed.

[0052] In addition, since the preset loss function uses operations, which ensures that the loss is penalized only when the distance of the negative quadruple samples is less than that of the positive quadruple samples, thus avoiding unnecessary gradient updates. In this way, the model can gradually optimize the embedding vectors, making the embedding vectors of the positive quadruple samples closer in space, while the embedding vectors of the negative quadruple samples are far from those of the positive quadruple samples, so as to effectively distinguish real user access behaviors from forged user behaviors.

[0053] Generally speaking, during the training process, each node and relationship in the sample knowledge graph need to be randomly initialized first, and at the same time, corresponding encoding vectors are generated for the time information; further, positive quadruple samples are sampled from the sample knowledge graph as real user behavior records, and negative quadruple samples are generated by randomly replacing one of the elements for training comparison. The model calculates the embedding representations of each positive quadruple sample and negative quadruple sample through forward propagation, and uses the preset loss function to evaluate the embedding compactness of the positive quadruple samples and the dispersion of the negative quadruple samples, and adopts the gradient descent method to optimize the embedding representations of the nodes and relationships, making the representations of the positive quadruple samples closer, while the representations of the negative quadruple samples gradually move away. The above process is continuously iterated until the preset loss function converges or reaches the maximum number of training rounds, so as to finally learn high-quality embedding vectors that can capture user behavior patterns and spatio-temporal dynamic characteristics.

[0054] In some embodiments, based on the embedding vectors, determining whether a candidate user is a target expanding group user of the seed user set includes: determining the average similarity between the candidate user and the seed user set based on the embedding vectors; the average similarity is the average of the similarities between the candidate user and each seed user in the seed user set; determining the similarity score between the candidate user and the seed user set based on the average similarity; and determining whether the candidate user is a target expanding group user of the seed user set based on the similarity score.

[0055] After obtaining the embedding vectors of the candidate users, it is necessary to realize the expansion of the group from the seed user set to the target users through the oversampling operation. Specifically, first, the embedding vectors of each seed user are extracted from the seed user set, and then by calculating the average of these embedding vectors of the seed users, a seed user embedding representation of the seed user set is obtained.

[0056] It can be understood that the number of candidate users is at least two, so the number of embedding vectors of the candidate users is at least two.

[0057] Furthermore, for each candidate user, calculate the average similarity between the candidate user and the seed user set.

[0058] Assume that the seed user set is denoted as , where the set of seed users includes seed users, represents the embedding vector of the -th seed user; assuming is the embedding vector of the -th candidate user, then the average similarity between the -th candidate user and the set of seed users can be calculated by the following formula : ; where, is the embedding vector of the -th seed user and the -th candidate user's embedding vector between the similarities.

[0059] It can be seen from the above formula that the average similarity is the average of the similarities between the candidate user and each seed user in the set of seed users.

[0060] Optionally, the similarity is the cosine similarity, and the calculation formula of the cosine similarity is: .

[0061] Further, for each candidate user, after obtaining its average similarity with the set of seed users, the similarity score of the candidate user with the set of seed users can be determined according to the average similarity, and based on the similarity score, it can be determined whether the candidate user is the target expansion user of the set of seed users.

[0062] In some embodiments, the number of candidate users is at least two; determining whether the candidate user is the target expansion user of the set of seed users based on the similarity score includes: sorting all candidate users in descending order according to the similarity score of each candidate user, and determining the similarity ranking of each candidate user; based on the similarity ranking of each candidate user, selecting a preset number of candidate users as the target expansion users of the set of seed users.

[0063] Specifically, the number of candidate users is at least two. After obtaining the similarity scores of all candidate users with the set of seed users, all candidate users are sorted in descending order according to the similarity score of each candidate user, and the similarity ranking of each candidate user is determined.

[0064] Among them, the higher the similarity score, the higher the similarity ranking, indicating that the candidate user is the most similar to the set of seed users.

[0065] Further, based on the similarity ranking of each candidate user, a preset number of candidate users are selected as the target expansion users of the seed user set.

[0066] Specifically, all candidate users after descending order can be used as potential expansion users of the seed user set. At this time, the preset number can be determined according to the advertising placement requirements, and a preset number of candidate users are screened out as the target expansion users of the seed user set.

[0067] For example, assume that there are 1 million candidate users after descending order, and there are 10,000 seed users in the seed user set. According to the advertising placement requirements, 100,000 seed users need to be advertised. Since the number of seed users is insufficient, the seed user set needs to be expanded, that is, 90,000 seed users need to be added. Then, the top 90,000 candidate users can be selected from 1 million candidate users according to the similarity ranking of each candidate user as the target expansion users of the seed user set, and these 90,000 candidate users are added to the seed user set to achieve the expansion of the seed user set and the accurate placement of advertisements.

[0068] This method can effectively avoid the deviation of single-user behavior by comprehensively considering the behavior characteristics of multiple seed users, so as to better expand the target user group that meets the advertising placement requirements.

[0069] The user expansion method provided by the embodiments of this application is different from the frequent itemset mining method based on rules. It uses the representation learning technology in deep learning to model the complex relationship between users and external links, enabling the model to capture more potential user behavior patterns, being able to perform more personalized modeling of users, avoiding the problem of "unified recommendation" in the interest group method, and thus improving the accuracy and adaptability of expansion; at the same time, by dynamically modeling the spatio-temporal changes of user behavior and dynamically capturing the changes of user behavior, especially under the influence of different times and different locations, more accurate user expansion can be achieved, avoiding the problem that the recommendation method based on static interest groups cannot adapt to the dynamic changes of users.

[0070] In some embodiments, the knowledge graph includes user nodes, external link nodes, time nodes, and base station nodes; based on user information, external link information, time information, and base station information, user behavior is modeled to construct a knowledge graph, including: modeling user information, external link information, time information, and base station information as user nodes, external link nodes, time nodes, and base station nodes respectively; modeling the network access behavior of candidate users as edges, and connecting user nodes, external link nodes, time nodes, and base station nodes based on the edges to generate a knowledge graph.

[0071] The embodiments of this application also provide a user expansion device. Please refer to Figure 3 ,Figure 3 This is a schematic structural diagram of the user group expansion device provided by an embodiment of the present application. In the embodiment of the present application, the user group expansion device includes an acquisition module 310, a knowledge graph construction module 320, a vector generation module 330, and a user group expansion module 340.

[0072] The acquisition module 310 is used to acquire the Internet behavior data of candidate users.

[0073] The Internet behavior data includes user information, external link information, time information, and base station information.

[0074] The knowledge graph construction module 320 is used to model user behavior based on user information, external link information, time information, and base station information, and construct a knowledge graph.

[0075] The vector generation module 330 is used to input the knowledge graph into a deep learning model for low-dimensional representation learning, and obtain the embedding vector of the candidate user output by the deep learning model.

[0076] The embedding vector is used to characterize the user behavior pattern and spatio-temporal dynamic characteristics of the candidate user.

[0077] The user group expansion module 340 is used to determine whether the candidate user is the target user for expanding the seed user set based on the embedding vector.

[0078] In some embodiments, the knowledge graph is represented in the form of quadruples.

[0079] The vector generation module 330 is used to obtain a set of positive quadruple samples and a set of negative quadruple samples; the set of positive quadruple samples includes multiple positive quadruple samples, the set of negative quadruple samples includes multiple negative quadruple samples, the positive quadruple samples are real user samples, and the negative quadruple samples are non-real user samples; based on the set of positive quadruple samples and the set of negative quadruple samples, pre-train the initial model constructed based on the TransE algorithm to obtain a deep learning model.

[0080] In some embodiments, the deep learning model is pre-trained using a preset loss function, and the preset loss function is determined based on the distance between the positive quadruple samples and the negative quadruple samples; the training objective of the deep learning model is to minimize the distance between the embedding vectors of all positive quadruple samples and maximize the distance between the embedding vectors of all negative quadruple samples by optimizing the preset loss function.

[0081] In some embodiments, the user group expansion module 340 is configured to determine the average similarity between a candidate user and a set of seed users based on the embedding vectors; the average similarity is the average of the similarities between the candidate user and each seed user in the set of seed users; determine a similarity score between the candidate user and the set of seed users based on the average similarity; and determine whether the candidate user is a target user for group expansion of the set of seed users based on the similarity score.

[0082] In some embodiments, the number of candidate users is at least two.

[0083] The user group expansion module 340 is configured to sort all candidate users in descending order according to the similarity score of each candidate user to determine the similarity ranking of each candidate user; and select a preset number of candidate users as the target users for group expansion of the set of seed users based on the similarity ranking of each candidate user.

[0084] In some embodiments, the knowledge graph includes user nodes, external link nodes, time nodes, and base station nodes.

[0085] Based on the user information, the knowledge graph construction module 320 is configured to model the user information, external link information, time information, and base station information as user nodes, external link nodes, time nodes, and base station nodes respectively; model the network access behavior of the candidate user as an edge, and connect the user nodes, external link nodes, time nodes, and base station nodes based on the edge to generate a knowledge graph.

[0086] An embodiment of the present application further provides an electronic device. Figure 4 It is a schematic structural diagram of the electronic device provided by the embodiment of the present application. As Figure 4 shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440. Among them, the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call the logical instructions in the memory 430 to execute the user group expansion method.

[0087] In addition, when the logical instructions in the above-mentioned memory 430 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0088] An embodiment of the present application also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the user group expansion method provided by the above-mentioned various methods is implemented.

[0089] An embodiment of the present application also provides a computer program product. The computer program product includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the user group expansion method provided by the above-mentioned various methods.

[0090] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative labor.

[0091] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the technical solution, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for expanding a user group, characterized in that, Including: Obtaining the Internet behavior data of candidate users; The Internet behavior data includes user information, external link information, time information, and base station information; Based on the user information, the external link information, the time information, and the base station information, modeling user behavior and constructing a knowledge graph; Inputting the knowledge graph into a deep learning model for low-dimensional representation learning to obtain the embedding vector of the candidate user output by the deep learning model; The embedding vector is used to characterize the user behavior pattern and spatio-temporal dynamic characteristics of the candidate user; Based on the embedding vector, determining whether the candidate user is a target expansion user of the seed user set.

2. The user group expansion method according to claim 1, wherein The knowledge graph is represented in the form of quadruples; Before inputting the knowledge graph into the deep learning model for low-dimensional representation learning to obtain the embedding vector of the candidate user output by the deep learning model, it further includes: Obtaining a positive quadruple sample set and a negative quadruple sample set; the positive quadruple sample set includes multiple positive quadruple samples, the negative quadruple sample set includes multiple negative quadruple samples, the positive quadruple samples are real user samples, and the negative quadruple samples are non-real user samples; Based on the positive quadruple sample set and the negative quadruple sample set, pre-training the initial model constructed based on the TransE algorithm to obtain the deep learning model.

3. The user group expansion method according to claim 2, wherein The deep learning model is pre-trained using a preset loss function, and the preset loss function is determined based on the distance between the positive quadruple samples and the negative quadruple samples; The training objective of the deep learning model is to minimize the distance between the embedding vectors of all the positive quadruple samples and maximize the distance between the embedding vectors of all the negative quadruple samples by optimizing the preset loss function.

4. The user group expansion method according to claim 1, characterized in that The determining whether the candidate user is a target expansion user of the seed user set based on the embedding vector includes: Based on the embedding vector, determining the average similarity between the candidate user and the seed user set; the average similarity is the average of the similarities between the candidate user and each seed user in the seed user set; Based on the average similarity, determining the similarity score between the candidate user and the seed user set; Based on the similarity score, determining whether the candidate user is a target expansion user of the seed user set.

5. The user group expansion method according to claim 4, wherein The number of candidate users is at least two; The determining whether the candidate user is a target expansion user of the seed user set based on the similarity score includes: Sorting all the candidate users in descending order according to the similarity score of each candidate user to determine the similarity ranking of each candidate user; Based on the similarity ranking of each candidate user, selecting a preset number of candidate users as the target expansion users of the seed user set.

6. The user group expansion method according to claim 1, wherein The knowledge graph includes user nodes, external link nodes, time nodes, and base station nodes; Modeling the user behavior based on the user information, the external link information, the time information, and the base station information to construct a knowledge graph, including: Modeling the user information, the external link information, the time information, and the base station information as the user node, the external link node, the time node, and the base station node respectively; Modeling the network access behavior of the candidate user as an edge, and connecting the user node, the external link node, the time node, and the base station node based on the edge to generate the knowledge graph.

7. A user group expansion device, characterized in that, Including: An acquisition module for acquiring the Internet access behavior data of a candidate user; The Internet access behavior data includes user information, external link information, time information, and base station information; A knowledge graph construction module for modeling the user behavior based on the user information, the external link information, the time information, and the base station information to construct a knowledge graph; A vector generation module for inputting the knowledge graph into a deep learning model for low-dimensional representation learning to obtain the embedding vector of the candidate user output by the deep learning model; the embedding vector is used to characterize the user behavior pattern and spatio-temporal dynamic characteristics of the candidate user; A user expansion module for determining whether the candidate user is a target expansion user of the seed user set based on the embedding vector.

8. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, the user expansion method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the user expansion method according to any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, the user expansion method according to any one of claims 1 to 6 is implemented.

Images