Construction method of generative heterogeneous graph model, agent distribution method and related device

Abstract

The present disclosure provides a construction method of a generative heterogeneous graph model, an agent distribution method and related devices, which are related to the fields of artificial intelligence technologies such as generative model, generative search, graph model, agent, object distribution, etc. The method comprises: extracting a first interaction action chain containing an agent distribution behavior exceeding a preset satisfaction degree from historical search behaviors and search behavior feedbacks of a first sample user; extracting a second interaction action chain containing an agent distribution behavior exceeding a preset satisfaction degree from historical recommended behaviors and recommended behavior feedbacks of a second sample user; and constructing a generative heterogeneous graph model based on the first interaction action chain and the second interaction action chain, taking the user, the behavior object and the agent as different nodes, and taking the interaction action as an edge connecting the corresponding nodes. And comprising: obtaining user-related information of a target user; inputting the user-related information into a pre-trained generative heterogeneous graph model to obtain a target agent output by the generative heterogeneous graph model; and distributing the target agent to the target user.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of data processing, in particular to the fields of artificial intelligence such as generative model, graph model, agent, object distribution, and more particularly to a method for constructing a generative heterogeneous graph model, an agent distribution method, and a device, an electronic device, a computer readable storage medium, and a computer program product corresponding to the method. BACKGROUND

[0002] The mainstream content distribution mode is to use a search system or a recommendation system to achieve, that is, to use a multi-layer system funnel to recall, sort, and mix the results of related webpages, and each strategy funnel sorts the webpages based on basic relevance, user feedback behavior, authority, and other information and truncates the output to the next layer.

[0003] That is, in the search scenario, the search system is used to understand the content of the search term or search statement input by the user, and the search results are matched according to the content relevance; in the recommendation scenario, the recommendation system is used to construct a user portrait based on the user's historical behavior or user preference, and the content that the user is interested in is pushed to the user according to the user portrait. SUMMARY

[0004] The present disclosure provides a method for constructing a generative heterogeneous graph model and an agent distribution method, and a device, an electronic device, a computer readable storage medium, and a computer program product corresponding to the method.

[0005] In a first aspect, the present disclosure provides a method for constructing a generative heterogeneous graph model, comprising: extracting a first interactive action chain containing an agent distribution behavior exceeding a preset satisfaction degree from the historical search behavior and search behavior feedback of a first sample user; extracting a second interactive action chain containing an agent distribution behavior exceeding a preset satisfaction degree from the historical recommended behavior and recommended behavior feedback of a second sample user; and constructing a generative heterogeneous graph model based on the first interactive action chain and the second interactive action chain, taking the user, the behavior object, and the agent as different nodes, and taking the interactive action as an edge connecting the corresponding nodes.

[0006] In a second aspect, the embodiments of the present disclosure provide a construction device of a generative heterogeneous graph model, comprising: a first interaction action chain extraction unit configured to extract a first interaction action chain containing an agent distribution behavior exceeding a preset satisfaction degree from historical search behaviors and search behavior feedbacks of a first sample user; a second interaction action chain extraction unit configured to extract a second interaction action chain containing an agent distribution behavior exceeding a preset satisfaction degree from historical recommended behaviors and recommended behavior feedbacks of a second sample user; and a generative heterogeneous graph model construction unit configured to construct a generative heterogeneous graph model in a form of taking a user, a behavior object and an agent as different nodes and taking an interaction action as an edge connecting the corresponding nodes based on the first interaction action chain and the second interaction action chain.

[0007] In a third aspect, the embodiments of the present disclosure provide an agent distribution method, comprising: obtaining user related information of a target user; inputting the user related information into a pre-trained generative heterogeneous graph model to obtain a target agent output by the generative heterogeneous graph model, wherein the generative heterogeneous graph model is obtained according to the construction method of the generative heterogeneous graph model described in the first aspect; and distributing the target agent to the target user.

[0008] In a fourth aspect, the embodiments of the present disclosure provide an agent distribution device, comprising: a user related information obtaining unit configured to obtain user related information of a target user; a model using unit configured to input the user related information into a pre-trained generative heterogeneous graph model to obtain a target agent output by the generative heterogeneous graph model, wherein the generative heterogeneous graph model is obtained according to the construction device of the generative heterogeneous graph model described in the second aspect; and an agent distribution unit configured to distribute the target agent to the target user.

[0009] In a fifth aspect, the embodiments of the present disclosure provide an electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to implement the construction method of the generative heterogeneous graph model described in the first aspect or the agent distribution method described in the third aspect when executed.

[0010] In a sixth aspect, the embodiments of the present disclosure provide a non-transitory computer readable storage medium storing computer instructions for enabling a computer to implement the construction method of the generative heterogeneous graph model described in the first aspect or the agent distribution method described in the third aspect when executed.

[0011] In a seventh aspect, the embodiments of the present disclosure provide a computer program product comprising a computer program which, when executed by a processor, is capable of implementing the method for constructing a generative heterogeneous graph model as described in the first aspect or the method for distributing an agent as described in the third aspect.

[0012] The method for constructing a generative heterogeneous graph model provided by the embodiments of the present disclosure extracts a first interactive action chain containing an agent distribution behavior considered satisfactory by a user from historical search behaviors and search behavior feedbacks in a search scenario, extracts a second interactive action chain containing an agent distribution behavior considered satisfactory by a user from historical recommended behaviors and recommended behavior feedbacks in a recommendation scenario, and then extracts different nodes played by users, behavior objects and agents and edges played by interactive actions from each interactive action chain covering the search scenario and the recommendation scenario, so as to finally construct a generative heterogeneous graph model according to the nodes and edges connecting different nodes, so that the constructed generative heterogeneous graph model fuses the association relationships between different users and different agents in the search scenario and the recommendation scenario in the form of a graph structure.

[0013] The method for distributing an agent provided by the embodiments of the present disclosure can quickly obtain a target agent matched to a target user without distinguishing the types of scenarios by directly inputting user-related information of the target user into the generative heterogeneous graph model constructed in the above manner, so that the agent matched to the target user can be distributed to the user in all scenarios by using the generative heterogeneous graph model.

[0014] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0015] Other features, objects and advantages of the present disclosure will become more apparent through reading the following detailed description of non-limiting embodiments made with reference to the accompanying drawings:

[0016] Figure 1 is an exemplary system architecture to which the present disclosure can be applied;

[0017] Figure 2 is a flowchart of a method for constructing a generative heterogeneous graph model provided by the embodiments of the present disclosure;

[0018] Figure 3 is a flowchart of a method for determining a first interactive action chain corresponding to a search scenario provided by the embodiments of the present disclosure;

[0019] Figure 4A flowchart of a method for determining and recommending a second interaction action chain corresponding to a scene according to an embodiment of the present disclosure is provided.

[0020] Figure 5 A flowchart of a method for training a generative heterogeneous graph model according to an embodiment of the present disclosure is provided.

[0021] Figure 6 A flowchart of an agent distribution method according to an embodiment of the present disclosure is provided.

[0022] Figure 7 A flowchart of a method for distributing a target agent matching a user according to the user's search demand according to an embodiment of the present disclosure is provided.

[0023] Figure 8 A flowchart of a method for distributing a target agent matching a user according to the user's preference information according to an embodiment of the present disclosure is provided.

[0024] Figure 9-1 and Figure 9-2 A system architecture diagram and an example diagram according to an embodiment of the present disclosure are provided.

[0025] Figure 10 A structural block diagram of a generative heterogeneous graph model construction device according to an embodiment of the present disclosure is provided.

[0026] Figure 11 A structural block diagram of an agent distribution device according to an embodiment of the present disclosure is provided.

[0027] Figure 12 A structural diagram of an electronic device suitable for executing a generative heterogeneous graph model construction method and / or an agent distribution method according to an embodiment of the present disclosure is provided. DETAILED DESCRIPTION

[0028] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to assist in understanding, and should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Also, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description. It should be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0029] In the technical solutions of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in the technical solutions comply with relevant laws and regulations and do not violate public order and good customs.

[0030] Figure 1 An exemplary system architecture 100 is shown to which the embodiments of the construction method of the generative heterogeneous graph model, the agent distribution method, and the corresponding apparatus, electronic device, and computer-readable storage medium of the present application can be applied.

[0031] As shown in Figure 1 The system architecture 100 can include terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 serves as a medium to provide a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 can include various connection types, such as wired, wireless communication links, or optical fiber cables, and the like.

[0032] A user can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages, and the like. The terminal devices 101, 102, 103 and the server 105 can be installed with various applications for realizing information communication between the two, such as information search applications, information recommendation applications, search and recommendation integrated applications, model training or model construction applications, agent distribution applications, and the like.

[0033] The terminal devices 101, 102, 103 and the server 105 can be hardware or software. When the terminal devices 101, 102, 103 are hardware, they can be various electronic devices with display screens, including but not limited to smartphones, tablet computers, laptop computers, desktop computers, and the like. When the terminal devices 101, 102, 103 are software, they can be installed in the above-mentioned electronic devices, and can be implemented as multiple software or software modules, or as a single software or software module, which is not specifically limited here. When the server 105 is hardware, it can be implemented as a distributed server cluster composed of multiple servers, or as a single server. When the server 105 is software, it can be implemented as multiple software or software modules, or as a single software or software module, which is not specifically limited here.

[0034] The server 105 can provide various services through various built-in applications. Taking an agent distribution application that can provide an agent distribution service for a user as an example, the server 105 can achieve the following effects when running the agent distribution application: first, receiving user-related information transmitted by a target user through a terminal device 101, 102, 103 used by the target user through the network 104; then, inputting the user-related information into a pre-trained generative heterogeneous graph model to obtain a target agent output by the generative heterogeneous graph model; and finally, distributing the target agent to an application used by the target user.

[0035] The generative heterogeneous graph model can be trained by a model construction class application built in the server 105 according to the following steps: first, a first interaction action chain containing an agent distribution behavior exceeding a preset satisfaction degree is extracted from the historical search behavior and search behavior feedback of a first sample user; then, a second interaction action chain containing an agent distribution behavior exceeding a preset satisfaction degree is extracted from the historical recommended behavior and recommended behavior feedback of a second sample user; and then, based on the first interaction action chain and the second interaction action chain, a generative heterogeneous graph model is constructed by taking the user, the behavior object and the agent as different nodes and taking the interaction action as an edge connecting the corresponding nodes.

[0036] Since a lot of computing resources and strong computing power are required to train the generative heterogeneous graph model, the construction method of the generative heterogeneous graph model provided in the subsequent embodiments of the present application is generally executed by the server 105 with strong computing power and more computing resources, and accordingly, the construction device of the generative heterogeneous graph model is generally arranged in the server 105. However, it should be noted that when the terminal devices 101, 102 and 103 also have computing power and computing resources that meet the requirements, the terminal devices 101, 102 and 103 can also complete the above-mentioned operations by the server 105 through the construction class application of the generative heterogeneous graph model installed thereon, and then output the same result as the server 105. Accordingly, the construction device of the generative heterogeneous graph model can also be arranged in the terminal devices 101, 102 and 103. In this case, the example system architecture 100 can also not include the server 105 and the network 104.

[0037] Of course, the server for training the generative heterogeneous graph model can be different from the server for using the trained generative heterogeneous graph model. In particular, the generative heterogeneous graph model trained by the server 105 can also be distilled into a lightweight generative heterogeneous graph model suitable for being placed in the terminal devices 101, 102 and 103, that is, the lightweight generative heterogeneous graph model in the terminal devices 101, 102 and 103 can be flexibly selected according to the actual requirement of recognition accuracy, or the more complex generative heterogeneous graph model in the server 105 can be selected.

[0038] It should be understood that Figure 1 The number of terminal devices, networks and servers in the system architecture 100 is only illustrative. According to the implementation needs, there can be any number of terminal devices, networks and servers.

[0039] For reference Figure 2 , Figure 2 A flowchart of a construction method of a generative heterogeneous graph model provided by an embodiment of the present disclosure is shown in FIG. 2, wherein the flowchart 200 includes the following steps:

[0040] Step 201: Extract the first interaction action chain containing agent distribution behavior exceeding the preset satisfaction level from the historical search behavior and search behavior feedback of the first sample users;

[0041] This step aims to be implemented by the entity executing the generative heterogeneous graph model construction method (e.g., Figure 1 The server 105 shown extracts a search interaction action chain containing agent distribution behaviors exceeding a preset satisfaction level from the historical search behaviors and search behavior feedback of the first sample users, and uses it as the first interaction action chain for subsequent use. The preset satisfaction level indicates that the user is satisfied with and approves of the actual agent distributed by the agent distribution behavior, and that the satisfaction or approval level exceeds a predetermined threshold. That is, the distributed agent should be an agent that effectively meets the user's expectations or intentions through the search behavior. This threshold can be preset and is usually evaluated based on feedback mechanisms (such as user feedback ratings, click-through rates, dwell time, etc.) to determine whether the performance of the distributed agent truly meets the user's actual needs.

[0042] Historical search behavior refers to all the search records a user has made during previous interactions with the search system. These records include information such as the keywords the user queried, the pages they viewed, and the links they clicked. Search behavior feedback refers to the further actions or feedback that a user may take in the search results (such as clicking on a result, the duration of viewing a page, or the evaluation of certain results). These behaviors can help the search system understand the user's satisfaction or preference for specific search results.

[0043] In this disclosure, "intelligent agent" refers to an entity encapsulated using generative large language models (such as GPT series models, BERT models, etc.) that can perform specific functions according to user needs. It is typically based on a generative large language model, and encapsulated by pre-defined role settings, function settings, or scenario settings. Specifically, these intelligent agents can provide users with services such as information retrieval, question answering, and task execution through natural language understanding, generation, and reasoning capabilities. Examples include a travel planning intelligent agent for providing users with travel planning services, a poetry generation intelligent agent for providing users with poetry generation services, and a digital human intelligent agent for providing users with immersive chat dialogue services, etc.

[0044] The search interaction action chain refers to a plurality of interaction steps or operation processes between the user and the search system, which includes the user initiating a query, the search system responding, and possibly further input by the user, feedback by the system, and final result display, etc. The first interaction action chain refers to a part of the search interaction action chain that contains the intelligent agent distribution behavior that is satisfactory or recognized by the user.

[0045] Step 202: Extracting a second interaction action chain containing an intelligent agent distribution behavior exceeding a preset satisfaction degree from the historical recommended behavior and the recommended behavior feedback of the second sample user.

[0046] On the basis of step 201, this step aims to extract a recommended interaction action chain containing an intelligent agent distribution behavior exceeding a preset satisfaction degree from the historical recommended behavior and the recommended behavior feedback of the second sample user, and use it as a second interaction action chain for subsequent use. The exceeding of the preset satisfaction degree indicates that the user is satisfied with and recognizes the actual intelligent agent distributed by the intelligent agent distribution behavior, and the satisfaction or recognition degree exceeds a certain predetermined threshold. That is, the distributed actual intelligent agent should be a better intelligent agent that meets the user's potential needs. The threshold can be pre-set, and the performance of the distributed actual intelligent agent can be evaluated based on a feedback mechanism (such as user satisfaction score, click, acceptance or not, subsequent behavior such as purchase, continued browsing, or explicit feedback such as score, comment, and like) to determine whether the actual intelligent agent meets the user's potential needs.

[0047] The historical recommended behavior refers to the recommended information received by the user in the recommendation system before. Specifically, it can be content recommended based on the user's interest, search record, click, purchase, etc. These behaviors can help the recommendation system understand the user's past interest trends and needs. The recommended behavior feedback refers to the feedback information given by the user after receiving the recommended content, which can be the user's satisfaction score, click, acceptance or not, subsequent behavior (such as purchase, continued browsing, etc.), or explicit feedback (such as score, comment, and like). These feedback information provides important data for the system on whether the recommendation is successful and whether it meets the user's expectations.

[0048] wherein the recommended interaction action chain generally means a feedback chain or behavior chain formed in the process of interaction between the user and the recommendation system, the chain is composed of multiple steps, each step may involve a certain action of the user (for example, clicking, querying, scoring, confirming, etc.) and the response of the recommendation system (such as agent recommendation, feedback adjustment, etc.). In simple terms, the recommended interaction action chain contains the behavior sequence of the recommendation system adjusting and optimizing according to the feedback after the user gives feedback on the recommendation behavior. For example, the user clicks on a certain recommended agent, indicating interest, satisfaction or dissatisfaction with the recommended result, and the system further optimizes the recommendation algorithm according to this feedback to form a new recommendation or improved agent distribution behavior. The second interaction action chain specifically refers to the part of the recommended interaction action chain that contains the agent distribution behavior that is satisfactory or recognized by the user.

[0049] Step 203: Based on the first interaction action chain and the second interaction action chain, a generative heterogeneous graph model is constructed by taking the user, the behavior object and the agent as different nodes, and taking the interaction action as the edge connecting the corresponding nodes.

[0050] On the basis of step 202, this step aims to extract the user, the behavior object and the agent contained in the first interaction action chain corresponding to the search scenario and the second interaction action chain corresponding to the recommendation scenario as different nodes from the above-mentioned execution subject, and then extract the interaction action as the edge connecting the corresponding nodes, and use these nodes and edges to construct a generative heterogeneous graph model.

[0051] wherein the generative heterogeneous graph model is a technology combining graph neural network (GNN) and generative model (such as generative adversarial networks, GAN), aiming to learn the node representation of the heterogeneous graph, and at the same time can generate new nodes or edges. The following is an introduction to some key points of the generative heterogeneous graph model:

[0052] Heterogeneous graph (Heterogeneous Graph), contains multiple types of nodes and edges, which can represent different entities and relationships. For example, in a movie recommendation system, nodes can be users, movies, directors, and edges can be user ratings of movies, similarities between movies, etc.

[0053] Generative model (Generative Model), by learning the distribution of data to generate new data samples. In the context of graphs, this means that the model can not only learn the representation of nodes, but also generate new nodes or edges to extend the structure of the graph

[0054] GNN is a kind of neural network specially processing graph structure data, which can capture the neighborhood information of nodes and update the representation of nodes. In heterogeneous graphs, GNN needs to handle different types of nodes and edges, which is usually realized by technologies such as Heterogeneous Graph Attention Networks.

[0055] The generative heterogeneous graph model constructed by the present disclosure is expected to be able to generate and represent the relationship between different types of nodes (such as users, behavior objects, agents) and edges (interaction actions), and capture complex interaction patterns by means of the graph-related characteristics of the model, so as to provide matching agent distribution in the whole scene for users.

[0056] Among them, the behavior object can include at least one of the following:

[0057] User input search term / search sentence: refers to the keyword or sentence input by the user in the search engine, used for information retrieval;

[0058] Selected target search result: the specific link or item selected by the user in the search result, which reflects the user's interest and demand;

[0059] Recommended search term / search sentence: the keyword or sentence recommended by the recommendation system based on the user's historical behavior or current context;

[0060] Information or object recommended to present in the information recommendation position: information, functional entities or goods specially recommended to the user in the user interface, such as advertisements, popular news, agents or goods, etc.

[0061] In addition, the behavior object can also include the following several other cases:

[0062] User-generated content: content published by users on social platforms, forums or comment areas, which can be part of the behavior object; User's historical interaction data: including user's click history, browsing history, etc., which can be used to build user portrait and then affect the recommendation system; User's location information: user's location information can affect search results and recommended content, such as localized search and recommendation; User device information: device type (such as mobile phone, tablet, computer) and device settings used by the user, which can be used to optimize user experience and content presentation; Interaction in social networks: likes, shares, comments and other interactive behaviors of users in social networks, which can be used as social signals to affect the spread and recommendation of information.

[0063] The generative heterogeneous graph model construction method provided in this disclosure extracts a first interaction action chain containing agent distribution behaviors deemed satisfactory by the user from historical search behaviors and search behavior feedback in a search scenario, and extracts a second interaction action chain containing agent distribution behaviors deemed satisfactory by the user from historical recommended behaviors and recommendation behavior feedback in a recommendation scenario. This allows for the extraction of different nodes (users, behavior objects, and agents) and edges (interaction actions) from the interaction action chains covering both search and recommendation scenarios. Finally, a generative heterogeneous graph model is constructed based on these nodes and the edges connecting different nodes, enabling the constructed model to integrate the relationships between different users and different agents in both search and recommendation scenarios in a graph structure.

[0064] Please refer to Figure 3 , Figure 3 A flowchart illustrating a method for determining a first interaction action chain corresponding to a search scenario, provided in this embodiment of the disclosure, is shown below. Figure 2 Step 201 in process 200 provided a specific implementation. Other steps in process 200 are not adjusted; a new complete embodiment is obtained by replacing step 201 with the specific implementation provided in this embodiment. Process 300 includes the following steps:

[0065] Step 301: Obtain the historical search behavior and search behavior feedback of the first sample user;

[0066] In obtaining the historical search behavior and search behavior feedback of the first sample users, it can be done on a per-student historical search session basis, that is, the corresponding historical search behavior and search behavior feedback can be extracted from each historical search session.

[0067] Step 302: From historical search behavior and search behavior feedback, filter search interaction action chains that contain the distribution behavior of any intelligent agent to obtain the target search interaction action chain;

[0068] Building upon step 301, this step aims to have the aforementioned executing entity filter search interaction action chains containing the distribution behavior of any agent from historical search behaviors and search behavior feedback, thereby obtaining the target search interaction action chain. In other words, the search interaction action chains containing agent distribution behavior are first filtered out and used as the target search interaction action chain for subsequent processing.

[0069] Step 303: Determine the actual satisfaction of the first sample user with the distributed agent based on the subsequent interaction actions of the first sample user with the distributed agent in the target search interaction action chain.

[0070] On the basis of step 302, the present step aims to determine the actual satisfaction degree of the first sample user to the distributed intelligent agent according to the subsequent interaction action of the first sample user to the distributed intelligent agent in the target search interaction action chain by the above-mentioned execution subject.

[0071] The subsequent interaction action for determining the actual satisfaction degree can include at least one of the following:

[0072] The first use duration of the distributed intelligent agent, the total use duration, the total use times or the average use duration of the distributed intelligent agent within a preset time period, whether the distributed intelligent agent is collected or forwarded, the number of forwarding operations or the forwarding channel of the distributed intelligent agent, whether the distributed intelligent agent is evaluated and the emotional tendency of the evaluation content, and other actions that can reflect the satisfaction degree of the user to the actual intelligent agent.

[0073] When the above-mentioned multiple specific interaction actions exist at the same time, the behaviors can also be processed into quantified scores or ratings through a quantification algorithm, and the actual satisfaction degree is calculated by a weighted calculation method combined with the weighted values converted by the influence degree of various interaction actions on the satisfaction degree.

[0074] Step 304: determining the target search interaction action chain with the actual satisfaction degree exceeding the preset satisfaction degree as the first interaction action chain.

[0075] The present embodiment aims to provide a specific implementation scheme for determining the actual satisfaction degree of the first interaction action chain by two-stage screening and evaluating the actual satisfaction degree of the distributed intelligent agent according to the subsequent interaction action, so as to make the provided implementation manner more down-to-earth and specific and feasible, and to make the determined first interaction action chain improve the accuracy as much as possible by combining multiple evaluation items.

[0076] Please refer to Figure 4 , Figure 4 The flowchart of a method for determining and recommending a second interaction action chain corresponding to a scene provided by the present embodiment, i.e., a specific implementation manner provided for step 202 in flowchart 200 shown in Figure 2 The other steps in flowchart 200 are not adjusted, and the specific implementation manner provided by the present embodiment is obtained in the form of replacing step 202 to obtain a new complete embodiment. Wherein flowchart 400 includes the following steps:

[0077] Step 401: obtaining the historical recommended behavior and recommended behavior feedback of the second sample user;

[0078] Wherein, in the process of obtaining the historical recommended behavior and the recommended behavior feedback of the second sample user, the historical recommended behavior and the recommended behavior feedback can be obtained in units of single historical recommended session, that is, the corresponding historical recommended behavior and the recommended behavior feedback can be extracted from each historical recommended session.

[0079] Step 402: screening the recommended interaction action chain containing the distribution behavior of any agent from the historical recommended behavior and the recommended behavior feedback to obtain the target recommended interaction action chain.

[0080] On the basis of step 401, the present step aims to screen the recommended interaction action chain containing the distribution behavior of any agent from the historical recommended behavior and the recommended behavior feedback by the above-mentioned execution subject to obtain the target recommended interaction action chain. That is, the recommended interaction action chain containing the distribution behavior of the agent is screened out as the target recommended interaction action chain for subsequent processing process.

[0081] Step 403: determining the actual satisfaction of the second sample user to the distributed agent according to the subsequent interaction action of the second sample user to the distributed agent in the target recommended interaction action chain.

[0082] On the basis of step 402, the present step aims to determine the actual satisfaction of the second sample user to the distributed agent according to the subsequent interaction action of the second sample user to the distributed agent in the target recommended interaction action chain by the above-mentioned execution subject.

[0083] Wherein, the subsequent interaction action for determining the actual satisfaction can include at least one of the following:

[0084] The viewing time length of the information recommendation position where the distributed agent is located, whether the distributed agent is run, the first running time length of the distributed agent and the time interval between the second running and the first running, whether the search operation related to the function of the distributed agent is actively performed after the distributed agent is run, and other actions that can reflect the satisfaction degree of the user to the distributed agent.

[0085] When the above-mentioned various specific interaction actions exist at the same time, the behaviors can also be processed into quantified scores or ratings through quantification algorithm, and the weighted weights converted from the influence degree of various interaction actions on the satisfaction are combined to calculate the comprehensive actual satisfaction through weighted calculation method.

[0086] Step 404: determining the target recommended interaction action chain with the actual satisfaction exceeding the preset satisfaction as the second interaction action chain.

[0087] The embodiment is aimed at providing a specific implementation scheme of the second interaction action chain by two-stage screening and evaluating the actual satisfaction of the actual agent distributed according to the subsequent interaction action, so as to make the provided implementation manner more down-to-earth and specific and feasible, and to make the determined second interaction action chain improve the accuracy as much as possible by combining various evaluation items.

[0088] Please refer to Figure 5 , Figure 5 A flowchart of a method for training a to-be-trained generative graph model to obtain a generative heterogeneous graph model is provided for the embodiment of the present disclosure, that is, a specific implementation manner is provided for step 203 in flow 200 shown in Figure 2 The other steps in flow 200 are not adjusted, and the specific implementation manner provided by the embodiment of the present disclosure is obtained in the form of replacing step 203 to obtain a new complete embodiment. Wherein flow 500 includes the following steps:

[0089] Step 501: obtaining a to-be-trained generative graph model;

[0090] The to-be-trained generative graph model can be a completely new model without training, or a basic model obtained by preliminary training on the same training samples in a similar scenario.

[0091] Step 502: extracting the interaction action between the user entity, the behavior object entity, the agent entity as different nodes and the different entities as the edges connecting different nodes in the first interaction action chain and the second interaction action chain;

[0092] Step 503: training the to-be-trained generative graph model by using the user entity, the behavior object entity, the agent entity and the interaction action between different entities to obtain a generative heterogeneous graph model.

[0093] The embodiment provides an implementation of a specific training of a generative heterogeneous graph model through steps 501-503: first, a generative graph model to be trained is needed as a basic graph model, which is used to learn the interaction mode between the user, the behavior object and the agent, and finally generate new interactions; then the user entity, the behavior object entity and the agent entity are extracted from the first and second interaction action chains, which will be used as nodes in the graph model, and the interaction actions between the different entities also need to be extracted, which will be used as edges connecting the nodes; next, the extracted entities and interaction actions are used to train the generative graph model, in this process, the model will learn how to represent the entities and the relationship between them, and the purpose of training is to enable the model to capture the complex interaction mode between the user, the behavior object and the agent, and to generate new interaction actions or predict future interactions based on the mode, and finally the trained generative heterogeneous graph model can not only understand the existing interaction data, but also generate new interaction data, thereby expanding and enriching the structure of the graph.

[0094] The above embodiments illustrate how to train a generative heterogeneous graph model from various aspects. In order to highlight the effect of the trained generative heterogeneous graph model as much as possible from the actual use scene, please also refer to Figure 6 , Figure 6 A scheme for implementing agent distribution using the trained generative heterogeneous graph model is shown, and the flow 600 includes the following steps:

[0095] Step 601: obtaining user-related information of a target user;

[0096] The target user can be a user who logs in to the service provided by the above execution subject in the service providing range, or a user who is determined by the application installed by the above execution subject to enter an active state or a use state. The user-related information of the user can be in different forms according to whether the target user initiates a search behavior.

[0097] Step 602: inputting the user-related information into the pre-trained generative heterogeneous graph model to obtain a target agent output by the generative heterogeneous graph model;

[0098] Step 603: distributing the target agent to the target user.

[0099] On the basis of step 602, this step aims to distribute the target agent to the target user in a suitable form at a suitable time by the above execution subject.

[0100] The agent distribution method provided by the embodiments of the present disclosure can quickly obtain a target agent matched with the target user without distinguishing the scene type by directly inputting the user-related information of the target user into the generative heterogeneous graph model constructed in the above manner, that is, the generative heterogeneous graph model can be used to distribute the matched agent to the user in the whole scene

[0101] To deepen the understanding of the scheme provided by the above embodiments, the following will provide two different specific implementation manners respectively through Figure 7 and Figure 8 two different cases or scenarios, wherein Figure 7 The flow of a method for distributing a matched target agent to a user according to obtained user search demand provided by the embodiments of the present disclosure includes the following steps:

[0102] Step 701: In response to the target user actively initiating a search behavior in the current session, obtaining the search demand from the search behavior actively initiated by the target user;

[0103] Step 702: Inputting the search demand into the pre-trained generative heterogeneous graph model to obtain the target agent output by the generative heterogeneous graph model;

[0104] Step 703: Distributing the target agent as the first candidate search result to the target user.

[0105] The embodiments are aimed at the case that the target user has an actively initiated search behavior in the current session, and the search demand is obtained from the actively initiated search behavior and input into the generative heterogeneous graph model as the specific user-related information in this case, so that the generative heterogeneous graph model matches the appropriate target agent for the target user according to the input search demand.

[0106] Figure 8 The flowchart of a method for distributing a matched target agent to a user according to determined user preference information provided by the embodiments of the present disclosure includes the following steps:

[0107] Step 801: In response to the target user not actively initiating a search behavior in the current session, obtaining the user preference information from the historical search session and / or the historical recommendation session of the target user;

[0108] Step 802: Inputting the user preference information into the pre-trained generative heterogeneous graph model to obtain the target agent output by the generative heterogeneous graph model;

[0109] Step 803: Determining a target information recommendation position matched with the target agent;

[0110] The display position of the target information recommendation position in the information stream presentation page is determined according to at least one of the following:

[0111] The total browsing duration of all information by the target user in the current information browsing behavior (that is, inserting an interaction request at a suitable information browsing position in the information browsing interface when the total browsing duration reaches a certain set duration threshold), the browsing duration of any information by the user in the current information browsing behavior (that is, inserting an interaction request at a suitable information browsing position in the information at the moment when the browsing duration of a single information reaches a certain set duration threshold), the total number of information browsed by the user in the current information browsing behavior (that is, inserting an interaction request at a suitable information browsing position in the information browsing interface when the total number of browsed information reaches a certain set number threshold), a preset position of the presented waterfall information stream (that is, inserting an interaction request at a certain preset dynamic insertion position of the waterfall information stream), a position of the end of the presented information body, and the like. Here, only a limited number of possible ways can be listed, but it is not limited to the above listed several cases.

[0112] Step 804: issuing the target agent to the target user in a manner of displaying at the target information recommendation position.

[0113] The embodiment is aimed at the case where the target user does not actively initiate a search behavior in the current session. Since there is no user-initiated demand information, the user preference information obtained from the historical search session and / or the historical recommendation session of the target user can only be used as the specific user-related information in this case to input the generative heterogeneous graph model, so as to match the appropriate target agent for the user according to the user preference information by the generative heterogeneous graph model.

[0114] On the basis of any of the above embodiments, feedback information of the target user on the distributed target agent can also be obtained, and then a new interaction action chain is generated according to the target user and the feedback information of the target agent, so as to update the generative heterogeneous graph model by using the new interaction action chain.

[0115] To deepen the understanding, the embodiment starts from the following difficulties in the task completion problem in the open field, and tries to provide a solution based on the generative heterogeneous graph model search and push integration:

[0116] 1) How to efficiently match the agent and the user demand in the capability space;

[0117] 2) How to efficiently model using the same algorithm for different scenarios;

[0118] 3) How to solve the distribution problem end-to-end and reduce the multi-layer funnel loss.

[0119] Specifically, the proposed agent search recommendation integrated distribution algorithm based on generative graph model can be applied to various scenarios that can perform agent distribution, such as search, recommendation, content hanging, etc. The scheme realizes accurate distribution of agents and user needs based on a generative heterogeneous graph model. The scheme is designed to have the following functions:

[0120] 1) Constructing a generative graph model based on large-scale user feedback and agent interaction data;

[0121] 2) Using user input and user state as input to perform walk recommendation on the graph model to output distributed agent candidates;

[0122] 3) In different distribution scenarios, generative ranking modeling is performed based on the unified representation of agent and user state in the graph model, and the model is estimated from the perspective of agent and demand capability matching to output agents that meet the needs for distribution.

[0123] And it is expected to be applied to agent distribution systems in search, information flow and other scenarios.

[0124] The implementation framework of the scheme is as follows Figure 9-1 As shown in the figure, the input of the system is: multiple scene agent distribution user behavior feedback logs (including in search scenarios, recommendation scenarios, various content hanging scenarios, etc.), and the output is the agent distribution result in a given scenario.

[0125] The entire distribution system is based on a unified heterogeneous graph model, which mainly includes several parts:

[0126] 1) Offline trained agent unified heterogeneous graph model: overall understanding, decomposition and concatenation of tasks, and final result;

[0127] 2) Online graph model generative distribution (online generative ranking): recall and ranking distribution based on graph model according to different scene inputs.

[0128] Figure 9-2 Then a specific example of how to construct a generative heterogeneous graph model from multiple user interaction action chains related to poetry and poets (such as Li Bai) is shown under the implementation framework provided in Figure 9-1 The integrated graph structure is also shown.

[0129] The technical effects that can be achieved by the above-mentioned scheme include:

[0130] 1. Unified graph model modeling of multi-scene data fusion

[0131] For the current search engine can do the relevance matching information retrieval of the request, the recommendation system can make content recommendation in the no Query scene, the scheme provided by the embodiment unifies the model from the search, recommendation mounting and other scenes, forms a set of unified agent graph model of multi-scene and multi-heterogeneous nodes, and serves the online algorithm distribution process. The unified graph model can greatly integrate the data of different scenes, obtain more connection information, and further deeply mine the relationship between different nodes to achieve more efficient description of nodes and edges.

[0132] 2. Online unified system distribution of different scenes based on graph model + generative model

[0133] According to the learning result of the multi-scene unified graph model, the graph model node is used as the input to construct an end-to-end generative model for different use scenes, different scenes can be distributed through the unified system, thereby greatly improving the efficiency and method effect of the system. The integration of search and recommendation is realized on the online system.

[0134] 3. Satisfactory degree matching in capability space through a generative model

[0135] The current mainstream search and recommendation algorithm system uses a semantic level matching distribution model for traditional content. This scheme is relatively suitable for deterministic content such as videos and texts, but the description and setting of the agent cannot completely represent the ability of the agent. The agent distribution needs to model the demand matching degree from the ability matching. The embodiment realizes the in-depth understanding and matching modeling of user demand and agent ability by constructing a capability matching space satisfaction estimation large model through a generative model, and continuously improves the distribution effect of each scene through online user feedback reinforcement learning alignment training.

[0136] Further reference Figure 10 and Figure 11 As an implementation of the method shown in each of the above figures, the disclosure respectively provides a generative heterogeneous graph model construction device embodiment and an agent distribution device embodiment. The generative heterogeneous graph model construction device embodiment corresponds to the generative heterogeneous graph model construction method embodiment shown in Figure 2 , and the agent distribution device embodiment corresponds to the agent distribution method embodiment shown in Figure 6 . The above device can be applied to various electronic devices.

[0137] As Figure 10As shown, the generation heterogeneous graph model construction device 1000 in this embodiment can include a first interaction action chain extraction unit 1001, a second interaction action chain extraction unit 1002, and a generation heterogeneous graph model construction unit 1003. The first interaction action chain extraction unit 1001 is configured to extract, from the historical search behavior and search behavior feedback of the first sample user, a first interaction action chain containing an agent distribution behavior exceeding a preset satisfaction degree. The second interaction action chain extraction unit 1002 is configured to extract, from the historical recommended behavior and recommended behavior feedback of the second sample user, a second interaction action chain containing an agent distribution behavior exceeding a preset satisfaction degree. The generation heterogeneous graph model construction unit 1003 is configured to construct a generation heterogeneous graph model based on the first interaction action chain and the second interaction action chain, taking the user, the behavior object, and the agent as different nodes, and taking the interaction action as an edge connecting the corresponding nodes.

[0138] In this embodiment, the specific processes of the first interaction action chain extraction unit 1001, the second interaction action chain extraction unit 1002, and the generation heterogeneous graph model construction unit 1003 in the generation heterogeneous graph model construction device 1000 and the technical effects brought by the specific processes can be respectively referred to Figure 2 The related descriptions of steps 201-203 in the corresponding embodiment will not be repeated here.

[0139] In some optional implementation manners of this embodiment, the first interaction action chain extraction unit 1001 is further configured to:

[0140] obtain the historical search behavior and the search behavior feedback of the first sample user;

[0141] filter, from the historical search behavior and the search behavior feedback, a search interaction action chain containing a distribution behavior of any agent, to obtain a target search interaction action chain;

[0142] determine an actual satisfaction degree of the first sample user to the distributed agent according to a subsequent interaction action of the first sample user to the distributed agent in the target search interaction action chain;

[0143] determine, as the first interaction action chain, the target search interaction action chain having an actual satisfaction degree exceeding a preset satisfaction degree.

[0144] In some optional implementation manners of this embodiment, the subsequent interaction action for determining the actual satisfaction degree includes at least one of the following:

[0145] a first use duration of the distributed agent;

[0146] a total use duration, a total use frequency, or an average use duration of the distributed agent within a preset time period;

[0147] whether a collection operation or a forwarding operation is performed on the distributed intelligent agent;

[0148] the number of forwarding operations or the forwarding channel of the distributed intelligent agent;

[0149] whether an evaluation is performed on the distributed intelligent agent and the sentiment tendency of the evaluation content.

[0150] In some optional implementations of the embodiment, the second interaction action chain extraction unit 1002 is further configured to:

[0151] obtain historical recommended behaviors and recommended behavior feedbacks of the second sample user;

[0152] from the historical recommended behaviors and the recommended behavior feedbacks, filter a recommended interaction action chain containing a distribution behavior of any intelligent agent to obtain a target recommended interaction action chain;

[0153] determine an actual satisfaction degree of the second sample user to the distributed intelligent agent according to a subsequent interaction action of the second sample user to the distributed intelligent agent in the target recommended interaction action chain;

[0154] determine the target recommended interaction action chain with an actual satisfaction degree exceeding a preset satisfaction degree as the second interaction action chain.

[0155] In some optional implementations of the embodiment, the subsequent interaction action for determining the actual satisfaction degree includes at least one of the following:

[0156] a viewing duration of an information recommendation position where the distributed intelligent agent is located;

[0157] whether the distributed intelligent agent is run;

[0158] a first running duration of the distributed intelligent agent and a time interval between a second running and the first running;

[0159] whether a search operation related to a function of the distributed intelligent agent is actively performed after the distributed intelligent agent is run.

[0160] In some optional implementations of the embodiment, the behavior object serving as the node can include at least one of the following:

[0161] a search word / search sentence input by the user, a selected target search result, a recommended search word / search sentence, or an information or object recommended to be presented in the information recommendation position.

[0162] In some optional implementations of the embodiment, the generative heterogeneous graph model construction unit 1003 is further configured to:

[0163] obtain a generative graph model to be trained;

[0164] extract, from the first interaction action chain and the second interaction action chain, an interaction action between different entities as different nodes of a user entity, an action object entity, an agent entity, and an interaction action between different entities as edges connecting different nodes;

[0165] train the generative graph model to be trained using the interaction action between different entities as different nodes of a user entity, an action object entity, an agent entity, and an interaction action between different entities as edges connecting different nodes, to obtain a generative heterogeneous graph model.

[0166] As shown in Figure 11 The agent distribution apparatus 1100 of the embodiment can include a user-related information obtaining unit 1101, a model using unit 1102, and an agent distribution unit 1103. The user-related information obtaining unit 1101 is configured to obtain user-related information of a target user. The model using unit 1102 is configured to input the user-related information into a pre-trained generative heterogeneous graph model to obtain a target agent output by the generative heterogeneous graph model. The generative heterogeneous graph model is obtained according to the construction apparatus 1000 of the generative heterogeneous graph model. The agent distribution unit 1103 is configured to distribute the target agent to the target user.

[0167] In the embodiment, the user-related information obtaining unit 1101, the model using unit 1102, and the agent distribution unit 1103 in the agent distribution apparatus 1100 can perform specific processing and bring about technical effects, which can be respectively referred to Figure 6 Corresponding to the related descriptions of steps 601-603 in the embodiment, no further description is given here.

[0168] In some optional implementation manners of the embodiment, the user-related information obtaining unit 1101 is further configured to:

[0169] In response to the target user actively initiating a search behavior in a current session, obtain a search demand from the search behavior actively initiated by the target user; and the user-related information includes the search demand.

[0170] In some optional implementation manners of the embodiment, the agent distribution unit 1103 is further configured to:

[0171] Distribute the target agent to the target user as a first candidate search result.

[0172] In some optional implementation manners of the embodiment, the user-related information obtaining unit 1101 is further configured to:

[0173] In response to the target user not actively initiating a search behavior in the current session, user preference information is obtained from a historical search session and / or a historical recommendation session of the target user; wherein the user-related information comprises the user preference information.

[0174] In some optional implementations of the embodiment, the intelligent agent distribution unit 1103 is further configured to:

[0175] determine a target information recommendation position matched with the target intelligent agent;

[0176] distribute the target intelligent agent to the target user in a manner of displaying the target intelligent agent in the target information recommendation position.

[0177] In some optional implementations of the embodiment, a display position of the target information recommendation position in an information stream presentation page is determined according to at least one of the following:

[0178] a total browsing duration of all information within the current information browsing behavior of the target user, a browsing duration of any information within the current information browsing behavior of the target user, a total number of information browsed within the current information browsing behavior of the target user, a preset position of presenting a waterfall information stream, and an end position of presenting an information body of the browsed information.

[0179] In some optional implementations of the embodiment, the intelligent agent distribution apparatus 1100 further comprises:

[0180] a feedback information obtaining unit configured to obtain feedback information of the target user on the distributed target intelligent agent;

[0181] a new interaction action chain generating unit configured to generate a new interaction action chain according to the target user and the feedback information on the distributed target intelligent agent, so as to update the generative heterogeneous graph model by using the new interaction action chain.

[0182] The above embodiment exists as a device embodiment corresponding to the above method embodiment. The device for constructing a generative heterogeneous graph model provided by the above embodiment extracts a first interaction action chain containing intelligent agent distribution behavior considered satisfactory by a user from historical search behavior and search behavior feedback in a search scenario, extracts a second interaction action chain containing intelligent agent distribution behavior considered satisfactory by a user from historical recommended behavior and recommended behavior feedback in a recommendation scenario, and then extracts different nodes played by users, behavior objects and intelligent agents and edges played by interaction actions from each interaction action chain covering the search scenario and the recommendation scenario. Finally, the generative heterogeneous graph model is constructed according to these nodes and edges connecting different nodes, so that the constructed generative heterogeneous graph model fuses the association relationship between different users and different intelligent agents in the search scenario and the recommendation scenario in the form of a graph structure.

[0183] The intelligent entity distribution apparatus provided by the above embodiments can obtain the target user related information, and can quickly obtain the target intelligent entity matched with the target user without distinguishing the scene type by directly inputting the user related information into the generative heterogeneous graph model constructed in the above manner, that is, the generative heterogeneous graph model can be used to distribute the matched intelligent entity to the user in all scenes.

[0184] According to an embodiment of the present disclosure, the present disclosure further provides an electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to implement the construction method of the generative heterogeneous graph model and / or the intelligent entity distribution method described in any of the above embodiments when executed.

[0185] According to an embodiment of the present disclosure, the present disclosure further provides a readable storage medium, which stores computer instructions for enabling a computer to implement the construction method of the generative heterogeneous graph model and / or the intelligent entity distribution method described in any of the above embodiments when executed.

[0186] The present disclosure provides a computer program product, which can implement the steps of the construction method of the generative heterogeneous graph model and / or the steps of the intelligent entity distribution method described in any of the above embodiments when executed by a processor.

[0187] Figure 12 A schematic block diagram of an example electronic device 1200 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present disclosure described and / or claimed in this document.

[0188] As Figure 12As shown, the device 1200 includes a computing unit 1201 that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) 1202 or a computer program loaded from the storage unit 1208 into a random access memory (RAM) 1203. In the RAM 1203, various programs and data required for the operation of the device 1200 can also be stored. The computing unit 1201, the ROM 1202, and the RAM 1203 are connected to each other through a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0189] A plurality of components in the device 1200 are connected to the I / O interface 1205, including: an input unit 1206, such as a keyboard, a mouse, and the like; an output unit 1207, such as various types of displays, speakers, and the like; a storage unit 1208, such as a magnetic disk, an optical disk, and the like; and a communication unit 1209, such as a network card, a modem, a wireless communication transceiver, and the like. The communication unit 1209 allows the device 1200 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0190] The computing unit 1201 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1201 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, and the like. The computing unit 1201 performs various methods and processes described above, such as the construction method of the generative heterogeneous graph model and / or the agent distribution method. For example, in some embodiments, the construction method of the generative heterogeneous graph model and / or the agent distribution method can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 1208. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 1200 via the ROM 1202 and / or the communication unit 1209. When the computer program is loaded into the RAM 1203 and executed by the computing unit 1201, one or more steps of the construction method of the generative heterogeneous graph model and / or the agent distribution method described above can be performed. Alternatively, in other embodiments, the computing unit 1201 can be configured to perform the construction method of the generative heterogeneous graph model and / or the agent distribution method by any other appropriate means, such as by means of firmware.

[0191] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0192] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0193] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0194] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0195] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0196] The computer system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. A server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of large management difficulty and weak business scalability in traditional physical host and virtual private server (VPS, Virtual Private Server) services.

[0197] The technical scheme of the embodiments of the present disclosure extracts the first interactive action chain containing the intelligent agent distribution behavior considered satisfactory by the user from the historical search behavior and search behavior feedback in the search scenario, extracts the second interactive action chain containing the intelligent agent distribution behavior considered satisfactory by the user from the historical recommended behavior and recommended behavior feedback in the recommendation scenario, and then can extract different nodes played by the user, the behavior object and the intelligent agent and the edges played by the interactive action from each interactive action chain covering the search scenario and the recommendation scenario, and finally obtains the generative heterogeneous graph model constructed according to these nodes and the edges connecting different nodes, so that the generative heterogeneous graph model constructed in the form of a graph structure fuses the association relationship between different users and different intelligent agents in the search scenario and the recommendation scenario. After obtaining the user-related information of the target user, the target intelligent agent matched with the target user can be quickly obtained without distinguishing the scene type by directly inputting the user-related information into the generative heterogeneous graph model constructed in the above manner, that is, the user can be distributed with the matched intelligent agent in the whole scene by using the generative heterogeneous graph model.

[0198] It should be understood that the steps shown above can be reordered, added, or deleted. For example, the steps described in the present disclosure can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical scheme of the present disclosure can be achieved, and the present disclosure is not limited herein.

[0199] The above detailed description does not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A construction method of a generative heterogeneous graph model, comprising: obtaining historical search behavior and search behavior feedback of a first sample user; from the historical search behavior and the search behavior feedback, screening a search interaction action chain containing distribution behavior of any agent to obtain a target search interaction action chain; determining a first actual satisfaction of the first sample user to the distributed agent according to a subsequent interaction action of the first sample user to the distributed agent in the target search interaction action chain; determining the target search interaction action chain with a first actual satisfaction exceeding a first preset satisfaction as a first interaction action chain; obtaining historical recommended behavior and recommended behavior feedback of a second sample user; from the historical recommended behavior and the recommended behavior feedback, screening a recommended interaction action chain containing distribution behavior of any agent to obtain a target recommended interaction action chain; determining a second actual satisfaction of the second sample user to the distributed agent according to a subsequent interaction action of the second sample user to the distributed agent in the target recommended interaction action chain; determining the target recommended interaction action chain with a second actual satisfaction exceeding a second preset satisfaction as a second interaction action chain; extracting users, behavior objects, and agents as different nodes and extracting interaction actions as edges connecting corresponding nodes from the first interaction action chain covering a search scenario and the second interaction action chain covering a recommended scenario to construct a generative heterogeneous graph model, the generative heterogeneous graph model being capable of generating new interaction actions.

2. The method of claim 1, wherein, The subsequent interaction action for determining the actual satisfaction of the first sample user to the distributed agent includes at least one of the following: first use duration of the distributed agent; total use duration, total use times, or average use duration of the distributed agent within a preset time period; whether a collection operation is performed on the distributed agent or whether a forwarding operation is performed on the distributed agent; times of forwarding operation or forwarding channels of the distributed agent; whether an evaluation is performed on the distributed agent and emotional tendency of the evaluation content.

3. The method of claim 1, wherein, The subsequent interaction action for determining the actual satisfaction of the second sample user to the distributed agent includes at least one of the following: viewing duration of an information recommendation position where the distributed agent is located; whether the distributed agent is run; first running duration of the distributed agent and time interval between second running and first running; whether a search operation related to a function of the distributed agent is actively performed after the distributed agent is run.

4. The method of claim 1, wherein, The behavior object taken as the node includes at least one of the following: a search word / search sentence input by a user, a selected target search result, a recommended search word / search sentence, information or an object recommended to be presented in an information recommendation position.

5. The method according to any one of claims 1 to 4, wherein, The construction of the generative heterogeneous graph model from the first interaction action chain covering the search scenario and the second interaction action chain covering the recommended scenario includes: obtaining a to-be-trained generative graph model; extracting, from the first interaction action chain covering the search scenario and the second interaction action chain covering the recommendation scenario, an interaction action between a user entity, a behavior object entity, an agent entity as different nodes and different entities as edges connecting the different nodes; training the to-be-trained generative graph model by using the interaction action between the user entity, the behavior object entity, the agent entity as different nodes and the different entities as edges connecting the different nodes, to obtain the generative heterogeneous graph model.

6. An agent distribution method, comprising: obtaining user-related information of a target user; inputting the user-related information into a pre-trained generative heterogeneous graph model to obtain a target agent output by the generative heterogeneous graph model; wherein the generative heterogeneous graph model is obtained according to the construction method of the generative heterogeneous graph model of any one of claims 1-5; distributing the target agent to the target user.

7. The method of claim 6, wherein, The obtaining of the user-related information of the target user comprises: in response to the target user actively initiating a search behavior in a current session, obtaining a search demand from the search behavior actively initiated by the target user; wherein the user-related information comprises the search demand.

8. The method of claim 7, wherein, The distribution of the target agent to the target user comprises: distributing the target agent to the target user as a first candidate search result.

9. The method of claim 6, wherein, The obtaining of the user-related information of the target user comprises: in response to the target user not actively initiating a search behavior in a current session, obtaining user preference information from a historical search session and / or a historical recommendation session of the target user; wherein the user-related information comprises the user preference information.

10. The method of claim 9, wherein, The distribution of the target agent to the target user comprises: determining a target information recommendation position matched with the target agent; distributing the target agent to the target user in a manner of displaying on the target information recommendation position.

11. The method of claim 10, wherein, The display position of the target information recommendation position in an information stream presentation page is determined according to at least one of the following: a total browsing time length of all information within this information browsing behavior of the target user, a browsing duration time of any information within this information browsing behavior of the user, a total number of browsed information within this information browsing behavior of the user, a preset position of presenting a waterfall information stream, and an end position of presenting an information body of the browsed information.

12. The method of any one of claims 6-11, further comprising: obtaining feedback information of the target user on the distributed target agent; generating a new interaction action chain according to the target user and the feedback information on the distributed target agent, so as to update the generative heterogeneous graph model by using the new interaction action chain.

13. A construction device of a generative heterogeneous graph model, comprising: a first interaction action chain extraction unit configured to obtain a historical search behavior and a search behavior feedback of a first sample user; The search interaction action chain containing the distribution behavior of any agent is filtered from the historical search behavior and the search behavior feedback to obtain a target search interaction action chain; and the first actual satisfaction of the first sample user to the distributed agent is determined according to a subsequent interaction action of the first sample user to the distributed agent in the target search interaction action chain. The target search interaction action chain with the first actual satisfaction exceeding the first preset satisfaction is determined as the first interaction action chain. The second interaction action chain extraction unit is configured to obtain historical recommended behavior and recommended behavior feedback of a second sample user. The recommended interaction action chain containing the distribution behavior of any agent is filtered from the historical recommended behavior and the recommended behavior feedback to obtain a target recommended interaction action chain; and the second actual satisfaction of the second sample user to the distributed agent is determined according to a subsequent interaction action of the second sample user to the distributed agent in the target recommended interaction action chain; and the target recommended interaction action chain with the second actual satisfaction exceeding the second preset satisfaction is determined as the second interaction action chain. The generative heterogeneous graph model construction unit is configured to extract a user, a behavior object and an agent as different nodes and extract an interaction action as an edge connecting the corresponding nodes from the first interaction action chain covering a search scenario and the second interaction action chain covering a recommendation scenario, to construct a generative heterogeneous graph model, which can generate a new interaction action.

14. The apparatus of claim 13, wherein, The subsequent interaction action used to determine the actual satisfaction includes at least one of the following: a first use duration of the distributed agent; a total use duration, a total use frequency or an average use duration of the distributed agent within a preset time period; whether a collection operation is performed on the distributed agent or whether a forwarding operation is performed on the distributed agent; a forwarding frequency or a forwarding channel of the distributed agent; whether an evaluation is performed on the distributed agent and a sentiment tendency of an evaluation content.

15. The apparatus of claim 13, wherein, The subsequent interaction action used to determine the actual satisfaction includes at least one of the following: a viewing duration of an information recommendation position where the distributed agent is located; whether the distributed agent is run; a first run duration of the distributed agent and a time interval between a second run and the first run; whether a search operation related to a function of the distributed agent is actively performed after the distributed agent is run.

16. The apparatus of claim 13, wherein, The behavior object taken as the node includes at least one of the following: a search word / search sentence input by a user, a selected target search result, a recommended search word / search sentence, recommended information or an object presented in an information recommendation position.

17. The apparatus of any of claims 13-16, wherein, The generative heterogeneous graph model construction unit is further configured to: obtain a generative graph model to be trained; extract, from the first interaction action chain covering a search scenario and the second interaction action chain covering a recommendation scenario, a user entity, a behavior object entity and an agent entity as different nodes and an interaction action between different entities as an edge connecting different nodes; and train the generative graph model to be trained based on the extracted user entity, behavior object entity, agent entity and interaction action. The generative heterogeneous graph model is trained by using the interaction actions between the user entities, the behavior object entities, the agent entities as different nodes and different entities as edges connecting the different nodes. 18.An agent distribution apparatus, comprising: a user-related information obtaining unit configured to obtain user-related information of a target user; a model using unit configured to input the user-related information into a pre-trained generative heterogeneous graph model to obtain a target agent output by the generative heterogeneous graph model; wherein the generative heterogeneous graph model is obtained by the construction apparatus of the generative heterogeneous graph model according to any one of claims 13-17; an agent distribution unit configured to distribute the target agent to the target user.

19. The apparatus of claim 18, wherein, The user-related information obtaining unit is further configured to: obtain a search demand from a search behavior initiated by the target user in response to the target user actively initiating the search behavior in a current session; wherein the user-related information comprises the search demand.

20. The apparatus of claim 19, wherein, The agent distribution unit is further configured to: distribute the target agent as a first candidate search result to the target user.

21. The apparatus of claim 18, wherein, The user-related information obtaining unit is further configured to: obtain user preference information from a historical search session and / or a historical recommendation session of the target user in response to the target user not actively initiating a search behavior in a current session; wherein the user-related information comprises the user preference information.

22. The apparatus of claim 21, wherein, The agent distribution unit is further configured to: determine a target information recommendation position matching the target agent; distribute the target agent to the target user in a manner of displaying the target agent in the target information recommendation position.

23. The apparatus of claim 22, wherein, The display position of the target information recommendation position in an information stream presentation page is determined according to at least one of the following: a total browsing time length of all information within the current information browsing behavior of the target user, a browsing duration time length of any information within the current information browsing behavior of the target user, a total number of browsed information within the current information browsing behavior of the target user, a preset position of presenting a waterfall information stream, and an end position of presenting a browsed information body. 24.The apparatus of any one of claims 18-23, further comprising: a feedback information obtaining unit configured to obtain feedback information of the target agent distributed to the target user; a new interaction action chain generating unit configured to generate a new interaction action chain according to the feedback information of the target user and the target agent distributed to the target user, so as to update the generative heterogeneous graph model by using the new interaction action chain. 25.An electronic device, comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the construction method of the generative heterogeneous graph model according to any one of claims 1-5 and / or the agent distribution method according to any one of claims 6-12. 26.A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the construction method of the generative heterogeneous graph model according to any one of claims 1-5 and / or the agent distribution method according to any one of claims 6-12. 27.A computer program product comprising a computer program which, when executed by a processor, implements the steps of the construction method of the generative heterogeneous graph model according to any one of claims 1-5 and / or the steps of the agent distribution method according to any one of claims 6-12.

Images