Method and apparatus for recommending candidate seed users

Abstract

The application relates to a method and device for recommending candidate seed users. A method for recommending candidate seed users comprises: obtaining personal information of users, sharing content of the users and data about interaction behaviors between the users, the users including at least one seed user and at least one candidate user; inputting the data into a neural network model to obtain a production side vector for representing production side features of the users and a consumption side vector for representing consumption side features of the users; constructing a vector retrieval service for the production side vector of the candidate user; obtaining, by using the production side vector of the seed user and through the constructed vector retrieval service, the candidate user satisfying a first predetermined condition as a first candidate seed user; obtaining, by using the consumption side vector of the seed user and through the constructed vector retrieval service, the candidate user satisfying a second predetermined condition as a second candidate seed user; and recommending the first and second candidate seed users as result candidate seed users.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to methods and apparatus for recommending candidate seed users. Background Technology

[0002] Currently, there are conventional recommendation methods in recommendation systems, such as those based on user-grouping algorithms or lookalike algorithms. These methods typically use target users, such as those who produce high-quality content, as seed users, and then select users similar to those seed users from among the candidate users. However, due to the limitations of the prior conditions of the recommendation method, the number of users recommended may be less than expected.

[0003] Therefore, there is a need for recommendation methods that can recommend more candidate users. Summary of the Invention

[0004] This disclosure provides methods and apparatus for recommending candidate seed users.

[0005] According to a first aspect of this disclosure, a method for recommending candidate seed users is provided, comprising: acquiring data including personal information of users, shared content of users, and data about interactive behaviors between users, wherein the users include at least one seed user and at least one candidate user; inputting the data into a neural network model to obtain a production-side vector characterizing the production-side characteristics of the users and a consumption-side vector characterizing the consumption-side characteristics of the users; constructing a vector retrieval service for the production-side vector of each of the at least one candidate user; using the production-side vector of the at least one seed user, and through the constructed vector retrieval service, obtaining candidate users who meet a first predetermined condition as first candidate seed users; using the consumption-side vector of the at least one seed user, and through the constructed vector retrieval service, obtaining candidate users who meet a second predetermined condition as second candidate seed users; and recommending the first candidate seed user and the second candidate seed user as resultant candidate seed users.

[0006] According to a second aspect of this disclosure, an apparatus for recommending candidate seed users is provided, comprising: a memory having instructions stored thereon; and a processor configured to perform the method according to this disclosure based on the instructions stored in the memory.

[0007] According to a third aspect of this disclosure, a computer-readable storage medium is provided that stores computer-executable instructions thereon, which, when executed by a processor, cause the processor to perform the method according to this disclosure.

[0008] The at least one candidate user can be a user who posts and shares content more than four times per month; and the at least one seed user can be a target user selected manually.

[0009] The method according to this disclosure can also utilize the acquired data to construct a graph network as data for the neural network model, including: based on the data, obtaining the user as a first node, obtaining shared content as a second node, and obtaining interactive behavior as an edge; using the acquired first node, second node, and edges to construct a graph network; and inputting the constructed graph network as data into the neural network model to obtain the user's producer side vector and consumer side vector, wherein the neural network model includes a GCN (Graph Neural Network) model.

[0010] User interaction behaviors include direct behaviors between users and indirect behaviors based on shared content; the direct behaviors correspond to unidirectional edges, while the indirect behaviors correspond to bidirectional edges.

[0011] The GCN model may include the GraphSAGE (Graph Sample and aggregate) model.

[0012] The method according to this disclosure also includes improvements to the neural network model, in which the production side vector is updated based on the behavior of publishing and sharing content, and the consumption side vector is updated based on other behaviors besides publishing and sharing content.

[0013] According to the method of this disclosure, the data can be preprocessed, and the preprocessed data can be used in the neural network model. The data preprocessing includes: for the user, deleting users whose number of interaction behaviors is less than a predetermined minimum and greater than a predetermined maximum; for the shared content, deleting shared content whose ranking result does not meet a predetermined ranking condition based on the ranking of the number of behaviors other than posting shared content; and for the interaction behavior, if the number of interaction behaviors is greater than a predetermined number, randomly deleting the interaction behaviors so that the number of interaction behaviors after random deletion reaches a predetermined number.

[0014] The method according to this disclosure may further include: storing the user's production side vector and consumption side vector; and updating the user's production side vector and consumption side vector on a daily basis.

[0015] In the aforementioned steps for obtaining recommended candidate seed users, the first predetermined condition is that the candidate user is the one whose search results are in the top 5% after the vector retrieval service is performed using the production side vector of the at least one seed user; and the second predetermined condition is that the candidate user is the one whose search results are in the top 5% after the vector retrieval service is performed using the consumption side vector of the at least one seed user.

[0016] According to the method disclosed herein, data processing can also be performed on the first candidate seed user and the second candidate seed user, and the processed user can be used as the result candidate seed user. The data processing includes deleting users from the first candidate seed user and the second candidate seed user who meet the following specific conditions: the user is reported by other users at a predetermined frequency; the user's abnormal sharing content accounts for a predetermined proportion; and the user is identified as a fake user.

[0017] Other features and advantages of the present technology will become clearer from the following detailed description of exemplary embodiments of the present disclosure with reference to the accompanying drawings. Attached Figure Description

[0018] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the specification, serve to explain the principles of the technology.

[0019] Referring to the accompanying drawings and the following detailed description, the present technology can be understood more clearly, wherein:

[0020] Figure 1 A flowchart of a candidate seed user recommendation method according to an exemplary embodiment of the present disclosure is shown.

[0021] Figure 2 A block diagram of a candidate seed user recommendation apparatus according to an exemplary embodiment of the present disclosure is shown.

[0022] Figure 3 A block diagram of a model training unit according to an exemplary embodiment of the present disclosure is shown.

[0023] Figure 4 A block diagram of a recommendation processing unit according to an exemplary embodiment of the present disclosure is shown.

[0024] Figure 5 A schematic diagram of the aggregation process of the GraphSAGE model is shown.

[0025] Figure 6 An exemplary configuration of a computing device that can implement embodiments of the present disclosure is shown. Detailed Implementation

[0026] Preferred embodiments of the present technology will now be described in detail with reference to the accompanying drawings. Details and functions not essential to the present technology have been omitted so as not to obscure the understanding of the present technology.

[0027] The following description of at least one exemplary embodiment is merely illustrative and is not intended to limit the scope of this disclosure or its application or use.

[0028] Please note that similar reference numbers and letters refer to similar items in the figure, and therefore once an item is defined in a figure, it does not need to be discussed in subsequent figures.

[0029] In this disclosure, the terms “first,” “second,” etc., are used only to distinguish between elements or steps and are not intended to indicate chronological order, priority, or importance.

[0030] Figure 1 A flowchart illustrating a candidate seed user recommendation method according to an exemplary embodiment of this disclosure is shown. Figure 1 As shown, the candidate seed user recommendation method according to this disclosure may include steps S101 to S106, and may be executed by the candidate seed user recommendation device.

[0031] In step S101, data including users' personal information, users' shared content, and data about interactions between users are obtained, wherein users can include at least one seed user and at least one candidate user.

[0032] Seed users can be manually selected target users, while candidate users can be users who post or share content more than four times per month. Note that the empirical parameter 4 is beneficial for effectively filtering candidate users and can significantly save computational resources.

[0033] Users can share content that includes at least one of the following: text, images, audio, and video.

[0034] User interaction behaviors can include direct behaviors between users and indirect behaviors based on shared content. Direct behaviors can include at least one of the following: private messaging, following, and clicking on profiles. Indirect behaviors include at least one of the following: posting shared content, liking, commenting, sharing, saving, and clicking on shared images.

[0035] Furthermore, in a further embodiment, the candidate seed user recommendation method according to this disclosure can further preprocess the acquired data, and use the preprocessed data in an improved neural network model according to this disclosure (hereinafter, in conjunction with...). Figure 5(Detailed description), wherein data preprocessing may include: for users, deleting users whose number of interaction behaviors is less than a predetermined minimum (e.g., 5 interactions in a month) and greater than a predetermined maximum (e.g., 5000 interactions in a month); for shared content, deleting shared content whose ranking does not meet predetermined ranking criteria (e.g., the ranking of the number of behaviors other than posting does not reach 80% of the ranking results) based on the ranking of the number of behaviors other than posting; and for interaction behaviors, randomly deleting interaction behaviors when the number of interaction behaviors exceeds a predetermined number (e.g., 5000), so that the number of interaction behaviors after random deletion reaches the predetermined number. Note that empirical parameters in preprocessing are beneficial for further effectively screening candidate users and can significantly save computational performance.

[0036] In a further embodiment, the candidate seed user recommendation method according to this disclosure can also utilize the acquired data to construct a network graph, and use the constructed network graph as data for an improved neural network model. Specifically, based on the acquired data, users are obtained as first nodes, shared content is obtained as second nodes, and interactive behaviors are obtained as edges, wherein direct behaviors correspond to unidirectional edges, and indirect behaviors correspond to bidirectional edges; a graph network is constructed using the acquired first nodes, second nodes, and edges; and the constructed graph network is used as data for an improved neural network model. Furthermore, the first node is also called a user node, and the user node uses user basic information features, user posting features, and user interactive behavior features as node features; the second node is also called a post node, and the post node uses post category information as node features.

[0037] Alternatively, the acquired data can be preprocessed as described above, and then the preprocessed data can be used to construct a network graph as data for improving the neural network model.

[0038] In step 102, the data is input into an improved neural network model to obtain a production-side vector characterizing the user's production-side characteristics and a consumption-side vector characterizing the user's consumption-side characteristics. In the improved neural network model, the production-side vector can be updated based on the behavior of publishing and sharing content, and the consumption-side vector can be updated based on behaviors other than publishing and sharing content. Furthermore, the improved neural network model may include an improved GCN model. Additionally, the improved GCN model includes an improved GraphSAGE model. In the following sections, this will be combined with... Figure 5 The improved neural network model according to this disclosure is further described in detail using the GraphSAGE model as an example.

[0039] In a further embodiment, the candidate seed user recommendation method according to this disclosure may further include storing the production side vector and consumption side vector of the user obtained by the above method in a recording medium, preferably in a Redis storage system.

[0040] In a further embodiment, the candidate seed user recommendation method according to this disclosure may also include updating (stored) the user's production side vector and consumption side vector on a daily basis.

[0041] In step 103, a vector retrieval service is constructed for the production side vectors of each of the at least one candidate user. Preferably, constructing the vector retrieval service may include, for example, constructing a weighted vector for the production side vectors of the at least one candidate user.

[0042] In step 104, using the production side vectors of at least one seed user, a vector retrieval service is constructed to obtain candidate users who meet a first predetermined condition as first candidate seed users. The first predetermined condition may be that the candidate users are those whose vectors rank in the top 5% of the search results after performing the vector retrieval service using the production side vectors of at least one seed user. Note that a higher threshold here results in higher retrieval accuracy; a lower threshold results in a wider coverage. A threshold of 5% is beneficial for obtaining high-precision and wide-ranging search results. Step 104 is beneficial for using the production side vectors of seed users to obtain candidate users whose production characteristics are similar to those of the seed users.

[0043] In step 105, using the consumption side vectors of the at least one seed user, a vector retrieval service is constructed to obtain candidate users who meet a second predetermined condition as second candidate seed users. The second predetermined condition may be that the candidate users corresponding to the vectors whose retrieval results are in the top 5% after executing the vector retrieval service using the consumption side vectors of the at least one seed user. Note that a higher threshold here results in higher retrieval accuracy; a lower threshold results in a wider coverage. A threshold of 5% is beneficial for obtaining high-precision and wide-ranging retrieval results. Step 105 is beneficial for using the consumption side vectors of seed users to obtain candidate users from among the candidate users who are favored by the seed users in terms of consumption.

[0044] In step 106, the first candidate seed user obtained in step 104 and the second candidate seed user obtained in step 105 are recommended as the result candidate seed users.

[0045] Furthermore, in a further embodiment, the candidate seed user recommendation method according to this disclosure may further include data processing of the first and second candidate seed users obtained through the above steps, and using the processed users as the result candidate seed users. The data processing may include deleting users from the first and second candidate seed users obtained through the above steps who meet the following specific conditions: the user has been reported by other users at a predetermined frequency (e.g., more than 30 reports within one month); the user's abnormal sharing content accounts for a predetermined proportion (e.g., the proportion of users failing manual review exceeds 30%); and the user is identified as a fake user (e.g., a user who steals images or has a fake persona).

[0046] Note that in the current recommendation platform, high-quality creators are the most important foundation for building a content ecosystem. However, the number of creators recommended by existing recommendation methods is relatively small, thus lacking an effective way to recommend creators. To address this issue, the technology disclosed herein can further consider the differences in users' production, consumption, and sharing of content, thereby effectively identifying more high-quality creators. This allows for the provision of traffic to these creators, which can help improve their retention, enhance the browsing experience of other users, and ultimately promote the development of the platform's content ecosystem.

[0047] Figure 2 A block diagram of a candidate seed user recommendation apparatus according to an exemplary embodiment of the present disclosure is shown.

[0048] like Figure 2 As shown, the candidate seed user recommendation device 20 according to this disclosure may include a model training unit 201 and a recommendation processing unit 202, which can perform the above-described combination. Figure 1 The method for recommending candidate seed users is described below. The following will combine... Figure 3 The model training unit 201 is described in detail, and combined with... Figure 4 The recommendation processing unit 202 is described in detail.

[0049] Additionally, the candidate seed user recommendation device 20 according to this disclosure may include a memory and a processor for performing combined... Figure 1 The method for recommending candidate seed users is described.

[0050] Figure 3 A block diagram of a model training unit 201 according to an exemplary embodiment of the present disclosure is shown.

[0051] like Figure 3 As shown, the model training unit 201 may include a training data acquisition module 301, a model training module 302, and a result output module 303.

[0052] The training data acquisition module 301 may include a data acquisition module 3011, and may optionally include at least one of a data preprocessing module 3012 and a data mapping module 3013.

[0053] Through the data acquisition module 3011, the training data acquisition module 301 can acquire data including users' personal information, users' shared content, and data about user interaction behavior, and use the acquired data, along with the pre-obtained production side vector and consumption side vector, as training data.

[0054] Users can include at least one seed user and at least one candidate user. Seed users can be target users selected manually, while candidate users can be users who post or share content more than four times per month. Note that the empirical parameter 4 is beneficial for effectively filtering candidate users and can significantly save computational resources.

[0055] Users can share content that includes at least one of the following: text, images, audio, and video.

[0056] User interaction behaviors can include direct behaviors between users and indirect behaviors based on shared content. Direct behaviors can include at least one of the following: private messaging, following, and clicking on profiles. Indirect behaviors include at least one of the following: posting shared content, liking, commenting, sharing, saving, and clicking on shared images.

[0057] In a further embodiment, the training data acquisition module 301 can further preprocess the acquired data through the data preprocessing module 3012, and use the preprocessed data as training data for the improved neural network model according to this disclosure. The data preprocessing may include: for users, deleting users whose interaction counts are less than a predetermined minimum (e.g., 5 interactions per month) and greater than a predetermined maximum (e.g., 5000 interactions per month); for shared content, deleting shared content whose ranking does not meet predetermined ranking conditions (e.g., the ranking of interactions other than posting does not reach 80% of the ranking results) based on the ranking of the number of interactions other than posting; and for interactive behaviors, randomly deleting interactive behaviors when the number of interactions exceeds a predetermined number (e.g., 5000), such that the number of interactions after random deletion reaches the predetermined number. Note that the empirical parameters in the preprocessing are beneficial for further effectively screening candidate users and can significantly save computational performance.

[0058] In a further embodiment, the training data acquisition module 301 can also construct a network graph using the acquired data through the data graph construction module 3013, and use the constructed network graph as training data for an improved neural network model. Specifically, based on the acquired data, users are obtained as first nodes, shared content as second nodes, and interactive behaviors as edges, where direct behaviors correspond to unidirectional edges, and indirect behaviors correspond to bidirectional edges; the acquired first nodes, second nodes, and edges are used to construct a graph network; and the constructed graph network is used as data for an improved neural network model. Furthermore, the first node is also called a user node, and the user node uses user basic information features, user posting features, and user interactive behavior features as node features; the second node is also called a post node, and the post node uses post category information as node features.

[0059] Alternatively, the acquired data can be preprocessed as described above, and then the preprocessed data can be used to construct a network graph as data for an improved neural network model. The improved neural network model may include an improved GCN model (e.g., a GraphSAGE model).

[0060] The improved neural network can be trained using training data through the model training module 302, preferably offline. Specifically, the model training unit 201 can input the training data obtained from the training data acquisition module 301 into the improved neural network model for training, to obtain a production side vector representing the user's production-side characteristics and a consumption side vector representing the user's consumption-side characteristics. In the improved neural network model, the production side vector can be updated based on the behavior of publishing and sharing content, and the consumption side vector can be updated based on behaviors other than publishing and sharing content. Furthermore, the improved neural network model may include an improved GCN model. Additionally, the improved GCN model includes an improved GraphSAGE model. In the following sections, this will be combined with... Figure 5 The improved neural network model according to this disclosure is further described in detail using the GraphSAGE model as an example.

[0061] The result output module 303 can, for example, be combined with the following... Figure 3 The recommendation processing unit 202 described outputs the results obtained from the model training module 302, including the user's production side vector 3031 and consumption side vector 3032.

[0062] In a further embodiment, the model training unit 201 may also include a data storage module (not shown), through which the production side vector and consumption side vector generated by the above method can be stored in a recording medium, preferably in a Redis storage system.

[0063] In a further embodiment, the model training unit 201 may also include a data update module (not shown), through which the (stored) user production side vectors and consumption side vectors can be updated on a daily basis.

[0064] Figure 4 A block diagram of a recommendation processing unit 202 according to an exemplary embodiment of the present disclosure is shown.

[0065] like Figure 4 As shown, the recommendation processing unit 202 may include a result acquisition module 401, a vector retrieval module 402, and a result recommendation module 403.

[0066] The result acquisition module 401 can acquire, for example, the production side vector and consumption side vector of at least one seed user and the production side vector and consumption side vector of at least one candidate user from the model training unit 201 for the recommendation processing unit 202.

[0067] Seed users can be manually selected target users, while candidate users can be users who post or share content more than four times per month. Note that the empirical parameter 4 is beneficial for effectively filtering candidate users and can significantly save computational resources.

[0068] The vector retrieval module 402 can construct a vector retrieval service for the production side vector of each of the at least one candidate users, and then use the production side vector and consumption side vector of the at least one seed user to obtain candidate users that meet the predetermined conditions through the constructed retrieval service.

[0069] Preferably, constructing a vector retrieval service may include, for example, constructing a weighted vector for the production side vector of each of at least one candidate user.

[0070] Specifically, the vector retrieval module 402 can utilize the production side vectors of at least one seed user to obtain candidate users who meet a first predetermined condition as first candidate seed users through a constructed vector retrieval service. The first predetermined condition can be that the candidate users are those whose vectors rank in the top 5% of the search results after executing the vector retrieval service using the production side vectors of at least one seed user. Note that a higher threshold here results in higher retrieval accuracy; a lower threshold results in a wider coverage. A threshold of 5% is beneficial for obtaining high-precision and wide-ranging search results. This is advantageous for using the production side vectors of seed users to obtain candidate users whose production characteristics are similar to those of the seed users.

[0071] Furthermore, the vector retrieval module 402 can utilize the consumption side vectors of the at least one seed user to obtain candidate users who meet a second predetermined condition as second candidate seed users through the constructed vector retrieval service. The second predetermined condition can be that the candidate users corresponding to the vectors whose retrieval results rank in the top 5% after executing the vector retrieval service using the consumption side vectors of the at least one seed user. Note that the higher the threshold set here, the higher the retrieval accuracy; the lower the threshold set, the wider the coverage. A threshold of 5% is beneficial for obtaining high-precision and wide-ranging retrieval results. This is advantageous for using the consumption side vectors of seed users to obtain candidate users from among the candidate users who are favored by the seed user for making purchases.

[0072] The result recommendation module 403 can recommend the first candidate seed user and the second candidate seed user (retrieval result) obtained from the vector retrieval module 402 as result candidate seed users.

[0073] In a further embodiment, the result recommendation module 403 can also process the search results and use the processed users as candidate seed users. The data processing may include deleting users from the first and second candidate seed users obtained by the vector retrieval module 402 who meet the following specific conditions: the user has been reported by other users at a predetermined frequency (e.g., more than 30 reports within one month); the user's abnormal sharing content accounts for a predetermined proportion (e.g., the proportion of users failing manual review exceeds 30%); and the user is identified as a fake user (e.g., a user who steals images or has a fake persona).

[0074] Figure 5 A schematic diagram of the aggregation process of the GraphSAGE model is shown.

[0075] First, refer to Figure 5 In step (a), neighbors are randomly sampled, with no more than Sk neighbors sampled in each hop. The first hop (k=1) samples 3 neighbors, and the second hop (k=2) samples 5 neighbors. For example, Figure 5 The neighbor nodes in the data include user nodes and shared content nodes.

[0076] Then, refer to Figure 3In step (b), the target node's embedding is generated by first aggregating the features of two-hop neighbors to generate the vectors of one-hop neighbors, and then aggregating the vectors of one-hop neighbors to generate the target node's vector. In the improved neural network model according to this disclosure (e.g., the GraphSAGE model), two vectors (i.e., a producer vector and a consumer vector) are maintained for each user node. When sampling and aggregating user node vectors, the producer vector of the user node is updated from the vectors of neighboring nodes corresponding to the publishing behavior (e.g., content sharing nodes), and the consumer vector of the user node is updated from the vectors of neighboring nodes corresponding to behaviors other than publishing (e.g., content sharing nodes and user nodes).

[0077] Finally, refer to Figure 5 In step (c), the target node's vector is input into the fully connected network to obtain the predicted value of the target node. Note that, compared to existing neural network models, the improved neural network model according to this disclosure can provide both production side vectors and consumption side vectors to better identify target users.

[0078] Figure 6 An exemplary configuration of a computing device that can implement embodiments of the present technology is shown. A computing device is an example of a hardware device to which the foregoing aspects of the present technology can be applied. A computing device can be any machine configured to perform processing and / or computation. A computing device can be, but is not limited to, a workstation, server, desktop computer, laptop computer, tablet computer, personal data assistant (PDA), smartphone, in-vehicle computer, or a combination thereof.

[0079] like Figure 6As shown, the computing device may include one or more components that may be connected to or communicate with bus 601 via one or more interfaces. Bus 601 may include, but is not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus. The computing device may include, for example, one or more processors 602, one or more input devices 603, and one or more output devices 604. The one or more processors 602 may be any kind of processor and may include, but is not limited to, one or more general-purpose processors or special-purpose processors (such as dedicated processing chips). Input devices 603 may be any type of input device capable of inputting information to the computing device and may include, but is not limited to, a mouse, keyboard, touchscreen, microphone, and / or remote controller. Output devices 604 may be any type of device capable of presenting information and may include, but is not limited to, a monitor, speaker, video / audio output terminal, vibrator, and / or printer.

[0080] The computing device may also include or be connected to a non-transitory storage device 607, which can be any non-transitory storage device capable of storing data, and may include, but is not limited to, disk drives, optical storage devices, solid-state storage, floppy disks, flexible disks, hard disks, magnetic tapes or any other magnetic media, compressed disks or any other optical media, cache memory and / or any other storage chip or module, and / or any other medium from which a computer can read data, instructions and / or code. The computing device may also include random access memory (RAM) 605 and read-only memory (ROM) 606. ROM 606 may store executable programs, utilities, or processes in a non-volatile manner. RAM 605 provides volatile data storage and stores instructions related to the operation of the computing device. The computing device may also include a network / bus interface 608 coupled to a data link 609. The network / bus interface 608 may be any kind of device or system capable of enabling communication with external devices and / or networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication devices and / or chipsets (such as Bluetooth). TM Equipment, 1302.11 equipment, WiFi equipment, WiMax equipment, cellular communication facilities, etc.

[0081] The various aspects, implementations, specific implementations, or features of the foregoing embodiments may be used individually or in any combination. The various aspects of the foregoing embodiments may be implemented by software, hardware, or a combination of hardware and software.

[0082] For example, the aforementioned embodiments may be embodied as computer-readable instructions recorded on a recording medium (e.g., a computer-readable storage medium). The recording medium is any data storage device capable of storing data that can subsequently be read by a computer system. Examples of recording media include read-only memory, random access memory, CD-ROM, DVD, magnetic tape, hard disk drive, solid-state drive, and optical data storage devices. The recording medium may also be distributed across network-coupled computer systems, allowing the computer-readable instructions to be stored and executed in a distributed manner.

[0083] For example, the aforementioned implementation scheme can take the form of hardware circuits. Hardware circuits can include any combination of combinational logic circuits, clock storage devices (such as floppy disks, flip-flops, latches, etc.), finite state machines, memories such as static random access memory or embedded dynamic random access memory, custom-designed circuits, programmable logic arrays, etc.

[0084] In one embodiment, the hardware circuit according to this disclosure can be implemented by encoding a circuit description using a hardware description language (HDL) such as Verilog or VHDL. An HDL description can be synthesized from a cell library designed for a given integrated circuit manufacturing technology and can be modified for timing, power, and other reasons to obtain a final design database, which can be transferred to a factory for integrated circuit production via a semiconductor manufacturing system. The semiconductor manufacturing system can produce integrated circuits by (e.g., on a wafer that may include a mask) depositing semiconductor material, removing material, changing the shape of the deposited material, modifying the material (e.g., by doping the material or modifying the dielectric constant using ultraviolet treatment), etc. Integrated circuits can include transistors and may also include other circuit elements (e.g., passive elements such as capacitors, resistors, inductors, etc.) and interconnections between transistors and circuit elements. Some embodiments may implement multiple integrated circuits coupled together to implement the hardware circuit, and / or discrete components may be used in some embodiments.

[0085] While specific embodiments of the present technology have been illustrated in detail by way of examples, those skilled in the art should understand that the examples are intended to be illustrative only and do not limit the scope of the present technology. It should be recognized that some steps in the foregoing methods are not necessarily performed in the order shown in the figures, but may be performed simultaneously, in different orders, or in an overlapping manner. Furthermore, those skilled in the art may add or omit steps as needed. Some components in the foregoing system are not necessarily arranged as shown in the figures; those skilled in the art may add or omit components as needed. Those skilled in the art should understand that the above embodiments can be modified without departing from the scope and spirit of this disclosure. The scope of this disclosure is defined by the appended claims.

Claims

1. A method for recommending candidate seed users, comprising: The data includes users’ personal information, users’ shared content and data about interactions between users, wherein the users include at least one seed user and at least one candidate user; Based on the data, the user is obtained as the first node, the shared content is obtained as the second node, and the interactive behavior is obtained as the edge. The interactive behavior between users includes direct behavior between users and indirect behavior based on shared content. The direct behavior corresponds to a one-way edge, and the indirect behavior corresponds to a two-way edge. Construct a graph network using the obtained first node, second node, and edges; The constructed graph network is input as data into the neural network model to obtain a production side vector representing the user's production-side characteristics and a consumption side vector representing the user's consumption-side characteristics. The neural network model includes a graph convolutional neural network (GCN) model, wherein the production side vector is updated based on the behavior of publishing and sharing content, and the consumption side vector is updated based on other behaviors besides publishing and sharing content. A vector retrieval service is constructed for the production side vector of each of the at least one candidate users; Using the production side vector of the at least one seed user, a vector retrieval service is constructed to obtain candidate users who meet the first predetermined conditions as first candidate seed users; Using the consumption side vectors of at least one seed user, a vector retrieval service is constructed to obtain candidate users who meet the second predetermined conditions as second candidate seed users; and The first and second candidate seed users are recommended as the final candidate seed users.

2. The method according to claim 1, wherein, The at least one candidate user is a user who posts or shares content more than four times per month; and The at least one seed user is a target user selected manually.

3. The method according to claim 1, wherein, The GCN model includes the graph sampling and aggregation GraphSAGE model.

4. The method according to claim 1, wherein, The data is preprocessed, and the preprocessed data is used in the neural network model. The data preprocessing includes: For the aforementioned users, delete those whose number of interaction behaviors is less than a predetermined minimum and greater than a predetermined maximum; For shared content, based on the number of times related to actions other than posting shared content, delete shared content whose ranking does not meet the predetermined ranking criteria; and For interactive behaviors, if the number of interactive behaviors exceeds the predetermined number, the interactive behaviors will be randomly deleted so that the number of interactive behaviors after the random deletion process reaches the predetermined number.

5. The method according to claim 1, further comprising: Store the user's production side vector and consumption side vector; and The user's production side vector and consumption side vector are updated on a daily basis.

6. The method according to claim 1, wherein, The first predetermined condition is that the candidate users whose search results are among the top 5% after performing the vector retrieval service using the production side vectors of the at least one seed user; and The second predetermined condition is that after executing the vector retrieval service using the consumption side vector of the at least one seed user, the candidate user whose retrieval result is in the top 5% of the vectors.

7. The method according to claim 1, further comprising: Data processing is performed on the first and second candidate seed users, and the processed users are used as the final candidate seed users. The data processing includes deleting users from the first and second candidate seed users who meet the following specific conditions: The user was reported by other users at a predetermined frequency; The percentage of abnormally shared content by the user reached a predetermined proportion; and The user was identified as a fake user.

8. An apparatus for recommending candidate seed users, comprising: A memory that stores instructions; and The processor is configured to execute the method according to any one of claims 1 to 7 based on instructions stored in the memory.

9. A computer-readable storage medium having stored thereon computer-executable instructions, which, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 7.

Images