Information processing method and device, electronic equipment and storage medium

Abstract

The invention is applied to the technical field of data processing, particularly can be applied to various scenes such as cloud technology, artificial intelligence, intelligent traffic and auxiliary driving, and relates to an information processing method and device, electronic equipment and a storage medium, and the method comprises the steps: building graph data based on the historical operation conditions of each target object for a business object in a specified business scene; the graph data comprises a node information set constructed based on nodes corresponding to each target object and each business object, a time information set constructed by operation time corresponding to a connection edge, node features and edge interaction features; and based on the graph data, performing multi-round iterative training on the initial graph neural network model to obtain a trained target model and target embedding vectors constructed by the target model for each node. Therefore, the business object range which can be considered by the target object is improved, and a favorable analysis basis is provided for analyzing the adaptability of the business object to the target object.

Description

Background Art

[0002] In related technologies, in order to better perform business interactions with related objects, information processing is usually performed on the related objects, and then content that may be of interest to the related objects is determined based on the information processing results.

[0003] Currently, when information processing is performed on related objects, historical operation data associated with the related objects is usually obtained first, and then the historical operation data is processed to obtain content that meets the needs.

[0004] For example, in a travel navigation business scenario, historical operation data of related objects can be obtained, and by analyzing the historical operation data, the places where the related objects travel most frequently can be determined.

[0005] However, existing processing methods rely heavily on the historical operation data of the object itself. Therefore, when determining the content that meets the needs, on the one hand, it has great content limitations. On the other hand, it may be affected by accidental factors of operation and misdetermine the content that meets the needs, thereby reducing the efficiency of business interaction with related objects.

[0006] In view of this, a more efficient and intelligent information processing method is needed to assist related objects in better business processing. Summary of the invention

[0007] The embodiments of the present application provide an information processing method, device, electronic device and storage medium to reduce the adverse effects caused by the content limitations of processable information and obtain a processing result that can effectively analyze the adaptation relationship between related objects and business content.

[0008] First, an information processing method is proposed for training a graph neural network model, including:

[0009] Based on the historical operation of each target object on the business object in the specified business scenario, graph data is constructed; the graph data includes: a node information set constructed based on the nodes corresponding to each target object and each business object, a time information set constructed by the operation time corresponding to the connection edge between the nodes, the node features of each node and the edge interaction features of each connection edge; the edge interaction features represent: the time sequence operation of the target object on the business object;

[0010] Based on the graph data, the initial graph neural network model is iteratively trained for multiple rounds to obtain a trained target model and a target embedding vector constructed by the target model for each node, wherein in one round of iterative training, the following operations are performed:

[0011] The graph neural network model is used to extract the corresponding operation time features for each operation time, and the following operations are performed for each node: the node features, operation time features and edge interaction features corresponding to each associated neighboring node are aggregated to obtain the corresponding prediction embedding vector;

[0012] Determine the node pairs with neighborhood relationships connected by each connection edge, determine the reference nodes in each node pair, and adjust the model parameters based on the similarity of the predicted embedding vectors between the reference node and the domain node corresponding to each node pair, and between the reference node and other non-neighborhood nodes.

[0013] In the second aspect, an information processing device is proposed for training a graph neural network model, comprising:

[0014] A construction unit is used to construct graph data based on the historical operation of each target object on the business object in a specified business scenario; the graph data includes: a node information set constructed based on the nodes corresponding to each target object and each business object, a time information set constructed by the operation time corresponding to the connection edge between the nodes, a node feature of each node and an edge interaction feature of each connection edge; the edge interaction feature represents: the time sequence operation of the target object on the business object;

[0015] A training unit is used to perform multiple rounds of iterative training on the initial graph neural network model based on the graph data to obtain a trained target model and a target embedding vector constructed by the target model for each node, wherein in one round of iterative training, the following operations are performed:

[0016] The graph neural network model is used to extract the corresponding operation time features for each operation time, and the following operations are performed for each node: the node features, operation time features and edge interaction features corresponding to each associated neighboring node are aggregated to obtain the corresponding prediction embedding vector;

[0017] Determine the node pairs with neighborhood relationships connected by each connection edge, determine the reference nodes in each node pair, and adjust the model parameters based on the similarity of the predicted embedding vectors between the reference node and the domain node corresponding to each node pair, and between the reference node and other non-neighborhood nodes.

[0018] Optionally, when constructing edge interaction features in graph data, the construction unit is used to:

[0019] Based on the historical operation data representing the historical operation conditions associated with each target object in the specified business scenario, the operation information of the business objects operated by each target object at different historical times is aggregated;

[0020] Corresponding to each target object and each business object covered by each historical operation data, each node is constructed, and according to the operation relationship between the target object and the business object, a connection edge is constructed between the corresponding nodes;

[0021] Based on the aggregated operation information groups corresponding to each connection edge, the corresponding edge interaction features are generated respectively.

[0022] Optionally, when aggregating the operation information of the target objects operating the business objects at different historical times, the construction unit is configured to include:

[0023] For each historical operation data associated with each target object, data content of each historical operation data is split according to a preset segmentation identifier to determine the operation information of the target object on the business object at different historical times;

[0024] For the target object and each business object of the operation, the operation information at different historical times is aggregated according to the information type to obtain an aggregated operation information group.

[0025] Optionally, when generating the corresponding edge interaction features based on the aggregated operation information groups corresponding to the respective connection edges, the construction unit is used to:

[0026] A preset encoding technology is used to encode different types of information content in each aggregated operation information group to obtain corresponding edge interaction features.

[0027] Optionally, when constructing a time information set in graph data, the construction unit is used to:

[0028] For each historical operation data associated with each target object, based on each historical time recorded in different historical operation data when the target object operates the same business object, the corresponding operation time is obtained;

[0029] Based on the operation time of each target object on the operated business object, a time information set is obtained.

[0030] Optionally, when obtaining a corresponding prediction embedding vector from the node features, operation time features, and edge interaction features corresponding to each of the neighborhood nodes associated with the aggregation, the training unit is used to:

[0031] Determine each neighboring node associated with the node according to a preset model depth and a neighborhood sampling function;

[0032] A temporal edge aggregator corresponding to the depth of the model is used to aggregate the comprehensive features corresponding to each of the associated neighboring nodes for the node to obtain a corresponding prediction embedding vector, wherein the comprehensive features are obtained by concatenating the node features, operation time features, and edge interaction features associated with the neighboring nodes.

[0033] Optionally, after obtaining the trained target model and the target embedding vectors respectively constructed by the target model for each node, the device further includes a recommendation unit, and the recommendation unit is further used to:

[0034] In response to an access indication of an access object, searching for a target embedding vector matching the access object;

[0035] When a target embedding vector matching the access object is found, a target business object adapted to the access object is determined based on the target embedding vector corresponding to the access object and the vector similarity between the target embedding vectors of each business object, and the target business object is recommended to the access object.

[0036] Optionally, when no target embedding vector matching the access object is found, the recommendation unit is configured to:

[0037] Based on the historical operation of the access object in the specified business scenario, determine the historical business object operated by the access object, and determine the target embedding vector matching the historical business object in each target embedding vector constructed by the target model for each node;

[0038] Based on the target embedding vector of the historical business object, a target embedding vector corresponding to the access object is generated, and based on the vector similarity between the target embedding vector of the access object and the target embedding vectors of each business object, a target business object recommended to the access object is determined.

[0039] Optionally, after obtaining the trained target model and the target embedding vectors respectively constructed by the target model for each node, the device further includes a recommendation unit, and the recommendation unit is further used to:

[0040] In response to an access instruction of accessing the object, obtaining an access time carried in the access instruction;

[0041] When it is determined that the access object is a target object in the graph data, a time constraint determined according to the access time is added to a neighborhood sampling function in the target model, and the target embedding vectors corresponding to the access object and each business object are updated using the adjusted target model;

[0042] Based on the updated target embedding vector and the vector similarity between the target embedding vectors of each business object, a target business object recommended to the access object is determined.

[0043] In a third aspect, an electronic device is proposed, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above method when executing the program.

[0044] In a fourth aspect, a computer-readable storage medium is proposed, on which a computer program is stored, and the computer program implements the above method when executed by a processor.

[0045] In a fifth aspect, a computer program product is proposed, comprising a computer program, wherein the computer program implements the above method when executed by a processor.

[0046] The beneficial effects of this application are as follows:

[0047] In an embodiment of the present application, an information processing method, device, electronic device and storage medium are proposed. By means of the historical operation of each target object on the business object in a specified business scenario, graph data for unsupervised model training is constructed, and the target embedding vector corresponding to each target object and each business object is obtained by training the graph neural network model. In the constructed graph data, there are node information sets constructed based on the nodes corresponding to each target object and each business object, time information sets constructed by the operation time corresponding to the connecting edges between nodes, node features of each node, and edge interaction features of each connecting edge, wherein the edge interaction feature characterizes: the time sequence operation of the target object on the business object. Moreover, in the specific model training process, the corresponding operation time features can be generated based on each operation time in the time information set, and then when generating the predicted embedding vector corresponding to the node, the node features, edge interaction features, and operation time features of the neighboring nodes are aggregated to obtain the embedding representation corresponding to the node.

[0048] In this way, when analyzing the historical operations of each target object on the business object in the specified business scenario and constructing the corresponding graph data, not only are nodes and connecting edges constructed based on the target object's operations on the business object, but also operation time features and edge interaction features are constructed based on the target object's operation time on the business object and the target object's sequential operations on the business object. This allows the introduction of considerations for operation time and operation conditions at different operation times into the constructed graph data, which not only helps to express the differences in the target object's operations on the same business object at different operation times in the graph data, but also helps to analyze the impact of operation time on the interaction between the target object and the business object, thereby improving the accuracy of the graph data's description of the interaction conditions in the specified business scenario.

[0049] Furthermore, in the process of training the graph neural network model for time feature extraction and embedding vector generation, for each node, the corresponding prediction embedding vector is obtained by aggregating the node features, operation time features and edge interaction features of the associated neighborhood nodes, so that the generated prediction embedding vector incorporates the influence of operation time, thereby enhancing the sensitivity of the node's embedding representation to the operation time; after completing the model training, by constructing the target embedding vector for each node respectively, the embedded representation of the target object and the business object can be realized from the global perspective of the specified business scenario, which provides a processing basis for analyzing the operation possibility between the target object and the business object from a global scope, increases the scope of business objects that can be considered for the target object, and provides a favorable basis for analyzing the adaptability of the business object to the target object. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of possible application scenarios in the embodiments of the present application;

[0051] Figure 2A This is a schematic diagram of the information processing process in the embodiment of the present application;

[0052] Figure 2B A schematic diagram of a process for determining a business object from various historical operation data in an embodiment of the present application;

[0053] Figure 2C A schematic diagram of a process for determining an operation time in an embodiment of the present application;

[0054] Figure 2D Schematic diagram of the process of aggregating operation information of a target object on a business object in an embodiment of the present application;

[0055] Figure 3A This is a schematic diagram of the process of performing a round of iterative training in an embodiment of the present application;

[0056] Figure 3B A schematic diagram of the directions of sampling neighborhood nodes and features associated with aggregation neighborhood nodes in an embodiment of the present application;

[0057] Figure 3C A schematic diagram of a process for aggregating comprehensive features of neighborhood nodes in an embodiment of the present application;

[0058] Figure 4 A schematic diagram of the process of constructing graph data for a travel navigation scenario in an embodiment of the present application;

[0059] Figure 5 A schematic diagram of presenting location recommendations for a visiting object in an embodiment of the present application;

[0060] Figure 6This is a schematic diagram of the logical structure of the information processing device in the embodiment of the present application;

[0061] Figure 7 A schematic diagram of a hardware structure of an electronic device according to an embodiment of the present application;

[0062] Figure 8 Schematic diagram of the structure of a computing device in an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the technical solution of the present application, rather than all of the embodiments. Based on the embodiments recorded in the application documents, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the technical solution of the present application.

[0064] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0065] Some of the terms used in the embodiments of the present application are explained below to facilitate understanding by those skilled in the art.

[0066] Graph neural network model: The graph neural network model proposed in the embodiment of the present application includes: a time feature extraction subnetwork and a graph neural subnetwork constructed based on a graph sampling and aggregation (Graph Sample and aggregate, GraphSage) architecture, wherein, in the embodiment of the present application, the graph neural subnetwork corresponds to a graph-based inductive learning algorithm, which can utilize node feature information and structural information to sample neighborhood nodes for nodes, and obtain the target embedding vector corresponding to the node by aggregating the features of neighborhood nodes and nodes.

[0067] Node: refers to the point in the graph constructed for a specified business scenario that represents each target object and each business object.

[0068] Connection edge: used to connect nodes with operational relationships. In the embodiment of the present application, the operational relationship of the corresponding target object to the business object is represented by means of the connection edge between the nodes.

[0069] The following is a brief introduction to the design concept of the embodiment of the present application:

[0070] In related technologies, in order to better perform business interactions with related objects, information processing is usually performed on the related objects, and then content that may be of interest to the related objects is determined based on the information processing results.

[0071] Currently, when information processing is performed on related objects, statistics are usually performed based on the historical operation data of the related objects to determine the operation frequency of the business objects, and then the content that the related objects may be interested in is determined based on the operation frequency of the business objects.

[0072] However, existing processing methods are heavily dependent on the historical operation data of the related objects themselves. Therefore, when determining the content that meets the needs, on the one hand, they have great content limitations. On the other hand, they may be affected by accidental factors of operation and incorrectly determine the content that meets the needs, thereby reducing the efficiency of business interaction with related objects.

[0073] In view of this, a more efficient and intelligent information processing method is needed to assist related objects in better business processing.

[0074] Correspondingly, in the application embodiment, an information processing method, device, electronic device and storage medium are proposed, which constructs graph data for unsupervised model training by means of the historical operation of each target object on the business object in the specified business scenario, and obtains the target embedding vector corresponding to each target object and each business object by training the graph neural network model. In the constructed graph data, it includes a node information set constructed based on the nodes corresponding to each target object and each business object, a time information set constructed by the operation time corresponding to the connecting edge between the nodes, each node feature of each node, and each edge interaction feature of each connecting edge, wherein the edge interaction feature represents: the time sequence operation of the target object on the business object. Moreover, in the specific model training process, the corresponding operation time feature can be generated based on each operation time in the time information set, and then when generating the predicted embedding vector corresponding to the node, the node features, edge interaction features, and operation time features of the neighboring nodes are aggregated to obtain the embedding representation corresponding to the node.

[0075] In this way, when analyzing the historical operations of each target object on the business object in the specified business scenario and constructing the corresponding graph data, not only are nodes and connecting edges constructed based on the target object's operations on the business object, but also operation time features and edge interaction features are constructed based on the target object's operation time on the business object and the target object's sequential operations on the business object. This allows the introduction of considerations for operation time and operation conditions at different operation times into the constructed graph data, which not only helps to express the differences in the target object's operations on the same business object at different operation times in the graph data, but also helps to analyze the impact of operation time on the interaction between the target object and the business object, thereby improving the accuracy of the graph data's description of the interaction conditions in the specified business scenario.

[0076] Furthermore, in the process of training the graph neural network model for time feature extraction and embedding vector generation, for each node, the corresponding prediction embedding vector is obtained by aggregating the node features, operation time features and edge interaction features of the associated neighboring nodes, so that the generated prediction embedding vector incorporates the influence of the operation time, and enhances the sensitivity of the node's embedding representation to the operation time; after completing the model training, by constructing the target embedding vector for each node, the embedding representation of the target object and the business object can be realized from the global perspective of the specified business scenario, which provides a processing basis for analyzing the operation possibility between the target object and the business object from a global scope, increases the scope of business objects that can be considered for the target object, and provides a favorable basis for analyzing the adaptability of the business object to the target object. The preferred embodiments of the present application are described below in conjunction with the drawings of the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application, and in the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0077] See also Figure 1 , which is a schematic diagram of a possible application scenario in an embodiment of the present application, includes a processing device 110 and a client device 120 .

[0078] In an embodiment of the present application, the processing device 110 trains the graph neural network model as a target model, and implements point of interest recommendations in a specified business scenario based on target embedding vectors constructed according to the target model corresponding to each target object and each business object.

[0079] In an embodiment of the present application, the processing device 110 and the client device 120 can communicate through a wired network or a wireless network; in various business scenarios, the processing device 110 can respond to an access indication triggered by the access object on the operational page of the client device 120, determine the business object of interest to the access object, and present the determined business object on the client device 120, wherein the operational page can belong to a developed application, web page, or mini-program, and the present application does not impose any specific restrictions on this.

[0080] The processing device 110 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms; or, it may be a terminal device with certain processing capabilities.

[0081] The client device 120 may be, for example, a mobile phone, a computer, an e-book reader, a smart home appliance, an intelligent voice interaction device, a vehicle terminal, and an aircraft.

[0082] It should be noted that the technical solution proposed in this application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, etc., and information processing can be realized in various scenarios.

[0083] The following describes the relevant application process in combination with several possible application scenarios:

[0084] Scenario 1: Information processing in travel navigation scenario.

[0085] In the travel navigation scenario illustrated in scenario 1, from the perspective of information processing, the processing device obtains the historical operation status of each target object for the travel location, and constructs graph data by analyzing the historical operation status of each target object, wherein corresponding nodes are constructed for each target object and each travel location in the travel navigation scenario, and connecting edges are constructed between corresponding nodes based on the operation status of the target object for the travel location;

[0086] Moreover, the graph data includes: a node information set constructed based on the nodes corresponding to each target object and each travel location, a node feature constructed based on each node, a time information set constructed based on the operation time corresponding to each connection edge, and an edge interaction feature corresponding to each connection edge constructed based on the time sequence operation of the object node (node ​​corresponding to the target object) on the business node (node ​​corresponding to the travel location) corresponding to each connection edge;

[0087] Afterwards, the processing device uses a graph neural network model training method based on graph data to analyze the historical operation status of each target object in the travel navigation scenario, and obtains the target embedding vector corresponding to each node.

[0088] From the perspective of the application of processing results, based on the target embedding vector corresponding to each node, the target embedding vector corresponding to each target object and each travel location can be determined. In the travel navigation scenario, the travel locations that the target object may be interested in can be determined by calculating the similarity of the target embedding vector between the target object and the travel location.

[0089] Scenario 2: Information processing in multimedia data playback scenario.

[0090] In the multimedia data playback scenario illustrated in Scenario 2, from the perspective of information processing, the processing device obtains the historical operation status of each target object on the multimedia data, and constructs graph data by analyzing the historical operation status of each target object, wherein corresponding nodes are constructed for each target object and each multimedia data in the multimedia data playback scenario, and connection edges are constructed between corresponding nodes based on the operation status of the target object on the multimedia data; the multimedia data can specifically be obtained from at least one or a combination of audio data, video data, and text data.

[0091] Moreover, the graph data includes: a node information set constructed based on the nodes corresponding to each target object and each multimedia data, a node feature constructed based on each node, a time information set constructed based on the operation time corresponding to each connection edge, and an edge interaction feature corresponding to each connection edge constructed based on the time sequence operation of the object node (node ​​corresponding to the target object) on the service node (node ​​corresponding to the multimedia data) corresponding to each connection edge;

[0092] Afterwards, the processing device uses a graph neural network model training method based on graph data to analyze the historical operation status of each target object in the multimedia data playback scenario, and obtains the target embedding vector corresponding to each node.

[0093] From the perspective of the application of information processing results, based on the target embedding vector corresponding to each node, the target embedding vector corresponding to each target object and each multimedia data can be determined, so that in the multimedia data playback scenario, the multimedia data that the target object may be interested in can be determined by calculating the similarity of the target embedding vector between the target object and the multimedia data.

[0094] Scenario 3: Information processing in e-commerce shopping scenarios.

[0095] In the e-commerce shopping scenario shown in scenario 3, from the perspective of information processing, the processing device obtains the historical operation status of each target object on the e-commerce product, and constructs graph data by analyzing the historical operation status of each target object, wherein corresponding nodes are constructed for each target object and each e-commerce product in the e-commerce shopping scenario, and connecting edges are constructed between corresponding nodes based on the operation status of the target object on the e-commerce product;

[0096] Moreover, the graph data includes: a node information set constructed based on the nodes corresponding to each target object and each e-commerce product, a node feature constructed based on each node, a time information set constructed based on the operation time feature corresponding to each connection edge, and an edge interaction feature corresponding to each connection edge constructed based on the time sequence operation of the target object point on the e-commerce product.

[0097] Afterwards, the processing device uses a graph neural network model training method based on graph data to analyze the historical operation status of each target object in the e-commerce shopping scenario, and obtains the target embedding vector corresponding to each node.

[0098] From the perspective of the application of processing results, based on the target embedding vector corresponding to each node, the target embedding vector corresponding to each target object and each e-commerce product can be determined. In the e-commerce shopping scenario, the e-commerce products that the target object may be interested in can be determined by calculating the similarity between the target embedding vectors between the target object and the e-commerce product.

[0099] The information processing process is described below from the perspective of the processing device 110 in conjunction with the accompanying drawings:

[0100] See also Figure 2A As shown, it is a schematic diagram of the information processing process in the embodiment of the present application. Figure 2A , the information processing process is explained in detail:

[0101] Step 201: The processing device constructs graph data based on the historical operation status of each target object on the business object in the specified business scenario.

[0102] In an embodiment of the present application, the processing device can, in various business scenarios, respectively construct corresponding graph data for the historical operation status of each target object for the business object in the corresponding business scenario, and then train the corresponding target model for each business scenario, and obtain each target embedding vector constructed by the target model for each target object and each business object. In the following description of the present application, only the process of the processing device constructing graph data for a specified business scenario and generating each target embedding vector for each node in the specified business scenario is used as an example for schematic explanation.

[0103] When the processing device constructs graph data based on the historical operations of each target object on the business object in the specified business scenario, the constructed graph data includes: a node information set constructed based on the nodes corresponding to each target object and each business object, a time information set constructed by the operation time corresponding to the connecting edges between the nodes, the node characteristics of each node and the edge interaction characteristics of each connecting edge, wherein the edge interaction characteristics represent: the temporal operation of the target object on the business object.

[0104] Specifically, in order to construct graph data, the processing device obtains each historical operation data associated with each target object in a specified business scenario, wherein, in a possible implementation method, each historical operation data of the target object can specifically be an operation log authorized to be disclosed by the target object; each historical operation data records the historical time when the target object operates the business object, as well as the operation process of the business object.

[0105] The following describes the process of constructing the node information set, time information set, edge interaction features, and node features based on each historical operation data:

[0106] 1) Construction of node information set

[0107] When the processing device constructs a node information set, it first obtains the historical operation data associated with each target object, and then constructs the corresponding nodes for each target object and each business object covered in the historical operation data under the specified business scenario; then, based on the object identification information corresponding to each target object and each business object, the node information set corresponding to each node is obtained, wherein the node information set includes the object identification information corresponding to each node, and the object identification information is used to uniquely identify an object.

[0108] It should be noted that, in the embodiment of the present application, when the processing device determines each business object covered in each historical operation data, it can search for a matching business object in each historical operation data based on the field name used to indicate the business object.

[0109] For example, assuming that the field name "pname" is used to identify the business object in each historical operation data, then the field name "pname" can be used to search in the historical operation data, and ultimately each business object identified by "pname" can be found from each historical operation data.

[0110] For example, see Figure 2B As shown, it is a schematic diagram of the process of determining a business object from each historical operation data in an embodiment of the present application. Figure 2B It can be seen from the schematic content that when determining the business objects operated by target object 1 based on the historical operation data of target object 1, by matching based on the field name "pname" that identifies the business object, it can be determined that the business objects operated by target object 1 are: business object 1 corresponding to P1 and business object 2 corresponding to P2.

[0111] 2) Construction of time information set

[0112] When the processing device constructs a time information set, it determines the historical time when each target object operates a business object recorded in the historical operation data associated with each target object according to the historical operation data corresponding to each target object; and then, for each target object, based on the historical times when the target object operates the same business object, a corresponding operation time is statistically determined.

[0113] Specifically, the processing device obtains the corresponding operation time for each historical operation data associated with each target object based on the historical times of the target object operating the same business object recorded in different historical operation data; and then obtains the time information set based on the operation time of each target object on the business object.

[0114] It should be noted that, in an embodiment of the present application, for a target object and a business object operated by the target object, when it is determined that the target object has operated on the same business object multiple times based on the historical operation data associated with the target object, the historical times of the target object for a business object can be obtained, wherein one historical time corresponds to an operation occurrence time recorded in a piece of historical operation data; further, in order to analyze the operating habits of a target object for a business object, the historical times of the target object operating the business object can be counted, and the operation time corresponding to the operation tendency can be determined at the time level.

[0115] When determining the corresponding operation time based on the analysis of each historical time, in some feasible technical solutions, the distribution of each historical time can be determined first, and then a candidate historical time can be selected as the corresponding operation time from among the candidate historical times that are distributed and clustered.

[0116] Specifically, when determining the candidate historical times for distribution clustering, first map each historical time to one day to obtain the time sequence corresponding to each historical time after mapping; then, when the time interval between each historical time after mapping is determined to be lower than the first set value for a number of consecutive times to reach the second set value, it is determined that there is a distribution clustering situation, and then the historical times that meet the distribution clustering judgment are determined as candidate historical times; thereafter, one can be selected from each candidate historical time as the corresponding operation time, wherein the selected candidate historical time can be arbitrarily selected, or can be the earliest occurrence among the candidate historical times, or can be the latest occurrence among the candidate historical times, or can be a candidate historical time with a time sequence position in the middle among the candidate historical times; the values ​​of the first set value and the second set value are set according to actual processing needs, and this application does not impose any specific restrictions.

[0117] For example, see Figure 2C As shown, it is a schematic diagram of the process of determining the operation time in an embodiment of the present application. Figure 2B On the basis of Figure 2C To illustrate, taking determining the operation time between target object 1 and business object 1 as an example, we can first determine the historical times of target object 1 operating business object 1 from the historical operation data associated with target object 1; then map each historical time within a day to determine the corresponding temporal relationship of each historical time on the time axis, and then perform time interval analysis on each historical time on the time axis; assuming that the first setting value is 20 minutes and the second setting value is 3, then 10:50:13, 10:57:21, and 11:13:10 mapped on the time axis can be determined as candidate historical times, and a candidate historical time in the middle position is selected as the corresponding operation time.

[0118] When determining the corresponding operation time based on the analysis of each historical time, in other feasible technical solutions, when the historical operation times included in each historical operation data correspond to the same day, one can be directly selected from the historical times as the operation time, wherein the selected historical time can be arbitrarily selected, or can be the earliest one among the historical times, or can be the latest one among the historical times, or can be a historical time with a chronological position in the middle among the historical times.

[0119] Alternatively, even if the historical operation times included in the historical operation data do not correspond to the same day, one of the historical times may be selected as the operation time, and this application does not impose any specific limitation on this.

[0120] Similarly, for each target object, the business object of each target object operation can be determined separately according to the historical operation data associated with the target object, and the corresponding operation time can be determined between the target object and the business object where there is an operation relationship in accordance with the same determination method as described above; then, based on the determined operation times, a time information set can be obtained.

[0121] In this way, the corresponding operation time can be determined between the target object and the business object that have an operation relationship, so that from the perspective of time, the time that represents the most likely time for the target object to operate the business object can be selected. This is equivalent to introducing time as a description content when describing the operation status of the target object and the business object, which helps to enhance time sensitivity in the subsequent processing results.

[0122] 3) Construction of edge interaction features

[0123] When the processing device constructs edge interaction features in graph data, it aggregates the operation information of each target object operating the business object at different historical times based on the historical operation data representing the historical operation conditions associated with each target object in the specified business scenario; then, corresponding to each target object and each business object covered by each historical operation data, it constructs each node, and constructs connecting edges between corresponding nodes based on the operational relationship between the target object and the business object; thereafter, based on the aggregated operation information groups corresponding to each connecting edge, it generates corresponding edge interaction features.

[0124] Specifically, when constructing edge interaction features in graph data, the connecting edges in the graph data are determined, that is, corresponding nodes are constructed for each target object and each business object covered in each historical operation data, and then connecting edges are constructed between the nodes corresponding to the target object and the business object where there is an operation relationship, wherein the connecting edges in the implementation of the present application are undirected edges; at the same time, for each target object, the operation information when the target object operates the same business object at different historical times is aggregated, which is equivalent to content aggregation of different historical operation data corresponding to the same target object and the same business object, and can correspond to the connecting edges between the target object and the business object to obtain the aggregated operation information group; further, according to the operation information groups corresponding to each connecting edge, corresponding edge interaction features are generated respectively.

[0125] In this way, by generating corresponding edge interaction features for the aggregated operation information groups corresponding to each connecting edge, it is possible to describe all operations of the target object on the business object at different historical times in the form of edge interaction features, thereby enabling a comprehensive analysis of the operations that have existed between the target object and the business object.

[0126] It should be noted that in the implementation of the present application, in the process of aggregating the operation information of each target object operating a business object at different historical times, the processing device performs data content splitting on each historical operation data associated with each target object according to a preset segmentation identifier to determine the operation information of the target object on the business object at different historical times; then, for the target object and each business object operated, the operation information at different historical times is aggregated according to the information type to obtain an aggregated operation information group.

[0127] It should be noted that when aggregating the operation information group, in some feasible embodiments, the processing device may aggregate different contents according to the information type; in other feasible embodiments, the processing device may directly splice different operation information.

[0128] For example, see Figure 2D As shown, it is a schematic diagram of the process of aggregating the operation information of a target object for a business object in the embodiment of the present application, according to Figure 2D As can be seen from the schematic content, based on the historical operation data associated with target object 1, the historical operation data of target object 1 operating business object 1 can be determined; then, by splitting the data content according to the segmentation mark “&”, the value corresponding to different keys can be obtained, that is, the information content of different information types can be obtained; then, the different information contents corresponding to the same information type recorded in different historical operation data are aggregated, and finally the Figure 2D The aggregated operation information group shown, wherein: Figure 2D The labels for the values ​​corresponding to the same key in different historical operation data are only for distinguishing the value content.

[0129] In this way, the data content representing the operation process can be separated from the historical operation data, and by aggregating the content of different information types, the different information contents under the same information type in the historical operation data corresponding to the same target object and the same business object can be brought together, which is equivalent to organizing the various types of operations and related information between the target object and the business object as a whole.

[0130] In addition, in an embodiment of the present application, when the processing device generates corresponding edge interaction features based on the aggregated operation information groups corresponding to each connecting edge, a preset encoding technology is used to encode different types of information content in the operation information groups for each aggregated operation information group to obtain corresponding edge interaction features.

[0131] Specifically, the processing device may use one-hot encoding technology to encode different types of content in each aggregated operation information group to obtain corresponding edge interaction features.

[0132] For example, assuming that a type of information content obtained through aggregation includes "likes, favorites", then, for this type of information content, encoding positions can be reserved according to possible information content, and then encoding and filling can be performed based on the information content to obtain an encoding result; specifically, assuming that for information type key1, the corresponding value range is "likes, comments, forwarding, playback, favorites", and for the historical operation process of target object 1 on business object 1, the information content corresponding to key1 obtained through aggregation is "likes, favorites", then, bits 0-4 in the edge interaction feature can be set to the encoding area corresponding to key1, and one type of information content corresponds to one bit of encoding result, then, when generating the edge interaction feature between target object 1 and business object 1, the feature content corresponding to key1 in the edge interaction feature is 10001.

[0133] In this way, by encoding the operation information for the aggregated operation information group, the corresponding edge interaction features can be obtained. Therefore, in the obtained edge interaction features, corresponding encoding representation forms exist for different types of information content, respectively, thereby being able to obtain features that integrate operation situations at different historical times.

[0134] 4) Construction of node features.

[0135] When the processing device constructs the corresponding node features for each node, it obtains the corresponding public attribute information of each target object for each target object, wherein, in some possible implementations, the public attribute information is included in the historical operation data associated with each target object. In other possible embodiments, the processing device collects and obtains the corresponding public attribute information for the target object.

[0136] It should be noted that when the target object's attribute information is not included in each historical operation data, or when less attribute information is included, the processing device can obtain the target object's publicly displayed attribute information, wherein the publicly displayed attribute information obtained can be any one or combination of the following: when the target object performs business processing in a specified business scenario based on an account in other applications, the target object sets the publicly displayed attribute information in other applications; when the target object obtains the attribute information with the authorization of the target object, the target object sets the displayed attribute information in other applications. The publicly displayed attribute information can specifically be the target object's publicly displayed gender information, age information, and region information.

[0137] In the embodiment of the present application, the acquisition of data related to the object is carried out with the authorization of the relevant object, and the acquisition and use of the data complies with the provisions of relevant laws and regulations.

[0138] Furthermore, after the processing device obtains the public attribute information corresponding to each target object, it encodes each public attribute information respectively to obtain the encoding result of the set dimension corresponding to each target object; finally, in order to take into account the nodes corresponding to the business objects, the encoding results of the target objects are dimensionally expanded to obtain the corresponding node features.

[0139] It should be noted that in an embodiment of the present application, when constructing node features for nodes corresponding to target objects and business objects, the constructed node features are used to initialize the embedding vectors corresponding to the nodes in the subsequent model training process. Optionally, the node features corresponding to each business object can be set to 0.

[0140] For example, the constructed node features can form a corresponding node feature matrix, and the constructed node feature matrix is ​​as follows:

[0141]

[0142] Among them, X0-X M-1 represents the node features corresponding to each target object, λi represents the encoding result obtained by encoding the public attribute information of the target object, and X M -X M+N-1 Represents the node features corresponding to each business object, M is the total number of target objects, N is the total number of business objects; for the node feature matrix X composed of each node feature, X∈R (M+N)×d1 , d1 is the node feature dimension, and the value of d1 is greater than or equal to the dimension of the encoding result.

[0143] In this way, the corresponding node feature vector can be constructed for each node, which provides a processing basis for feature aggregation in the initial stage when subsequently training the graph neural network model.

[0144] It should be understood that, in the process of generating corresponding graph data for each historical operation data associated with each target object in a specified business scenario, the generation of different contents in the graph data can be carried out in parallel, and this application does not impose any specific restrictions on this.

[0145] Step 202: The processing device performs multiple rounds of iterative training on the initial graph neural network model based on the graph data to obtain a trained target model and a target embedding vector constructed by the target model for each node.

[0146] After the processing device constructs the graph data, it performs multiple rounds of unsupervised iterative training on the initial graph neural network model based on the graph data until the preset convergence conditions are met, and obtains the trained target model and the target embedding vector constructed by the target model for each node, wherein the preset convergence conditions may be: the number of training rounds reaches a set value, or the number of times the loss value is continuously lower than the first threshold value reaches a second threshold value, etc., wherein the values ​​of the first threshold value and the second threshold value are set according to actual processing needs.

[0147] Specifically, when constructing the initial graph neural network model, the graph neural network model can be constructed based on the time feature extraction subnetwork and the graph neural subnetwork.

[0148] For the time feature extraction subnetwork, what is realized is to convert the time information into time features, that is, to extract features for the operation time and obtain the corresponding operation time features. The feature extraction logic implemented inside the time feature extraction subnetwork is as follows:

[0149]

[0150] Among them, t2v(t) is an m+1-dimensional vector, that is, the vector obtained by feature extraction for time t, F is a periodic activation function (such as the sin function); W i and Refers to the trainable model parameters.

[0151] For the graph neural sub-network, it can be specifically built based on the GraghSage architecture, which is used to aggregate the features of neighboring nodes for each node, and finally obtain the corresponding embedding vector; it should be understood that in the embodiment of the present application, the time feature extraction sub-network and the graph neural sub-network are jointly trained in the process of training the graph neural network model.

[0152] It should be noted that from the perspective of mathematical processing, when the processing device inputs graph data into the graph neural network model, in order to make the constructed graph data recognizable by the model, it is necessary to establish the corresponding adjacency matrix and feature matrix to carry the graph data based on the operational relationship between the nodes.

[0153] With the help of the constructed adjacency matrix, the subsequent graph neural sub-network can perform neighborhood sampling based on the adjacency matrix, where the adjacency matrix represents the connection status between nodes; at the same time, a node feature matrix is ​​constructed for each node feature obtained, an operation time feature matrix is ​​constructed for each operation time feature, and an edge interaction feature matrix is ​​constructed for the edge interaction features corresponding to each connection edge; then the constructed adjacency matrix, operation time feature matrix, node feature matrix, and edge interaction feature matrix are input into the corresponding graph neural sub-network to realize node feature aggregation.

[0154] The following is a description of the processing involved in the initial round of iterative training with reference to the accompanying drawings:

[0155] See also Figure 3A As shown, it is a schematic diagram of the process of performing a round of iterative training in the embodiment of the present application. Figure 3A , describing the steps performed during the initial round of training:

[0156] Step 301: The processing device uses a graph neural network model to extract corresponding operation time features for each operation time, and performs the following operations for each node: aggregate the node features, operation time features and edge interaction features corresponding to each associated neighborhood node to obtain the corresponding prediction embedding vector.

[0157] In the processing process corresponding to step 301, the processing device trains the initial graph neural network model based on the graph data constructed based on the specified business scenario, and in the model training process, extracts the corresponding operation time features for each operation time in the time information set;

[0158] Afterwards, for each node, the node features, operation time features and edge interaction features corresponding to each associated neighboring node are aggregated to obtain the corresponding prediction embedding vector. According to the preset model depth and neighborhood sampling function, the neighboring nodes associated with the node are determined; then, the temporal edge aggregator corresponding to the model depth is used to aggregate the comprehensive features corresponding to each associated neighboring node for the node to obtain the corresponding prediction embedding vector, wherein the comprehensive features are obtained by concatenating the node features, operation time features, and edge interaction features associated with the neighboring nodes.

[0159] Specifically, a graph neural network model is used to determine the corresponding neighborhood nodes for each node indicated in the node information set using a preset neighborhood sampling function; then, according to the preset model depth, the comprehensive features associated with each neighborhood node are aggregated to obtain a predicted embedding vector, wherein the node features associated with the neighborhood nodes refer to the node features corresponding to the neighborhood nodes themselves; the edge interaction features associated with the neighborhood nodes refer to: the edge interaction features corresponding to the connecting edges between the neighborhood nodes and the current node; the operation time features associated with the domain nodes refer to: the operation time features corresponding to the connecting edges between the neighborhood nodes and the current node.

[0160] The following is a description of the process of aggregating neighborhood node features within the graph neural network model to obtain the features of each node, in conjunction with Table 1:

[0161] Table 1

[0162]

[0163] According to the node aggregation process shown in Table 1, the graph data provided to the graph neural network model is recorded as G(V, E, X, F), where V refers to the node information set that identifies each node, E refers to the information set composed of each operation time feature, X refers to the feature matrix composed of each node feature, and F refers to the edge interaction feature matrix constructed by the edge interaction features of each connecting edge; before model training, a specific value of the depth K is set, W k It represents a learnable weight matrix, and corresponding aggregation functions are set for different depth values. In addition, the depth K in the algorithm refers to the number of time-series edge aggregators, the number of weight matrices, and the number of network layers. The aggregators and weight matrices are shared in each layer of the network. The number of network layers can be understood as the number of hops in the neighborhood that requires maximum access.

[0164] Continuing with the aggregation process illustrated in the result table 1, for each depth value in the present application, there is a corresponding temporal edge aggregator algorithm to implement feature aggregation, wherein the main goal of the temporal edge aggregator in the present application is: for each node, aggregate the features of neighboring nodes and the edge features between nodes to obtain a vector representation, and the edge features include operation time features and edge interaction features.

[0165] The outer loop in the temporal edge aggregator algorithm is executed as follows:

[0166] First, take each node u∈V as the central node and aggregate its neighboring nodes Comprehensive characteristics of Forming vectors It can be seen that in the process of obtaining the embedding vector corresponding to node u in the kth aggregation, it depends on the representation information generated by the k-1th aggregation as well as When k = 0, the following relationship holds: x u Represents the node features constructed for node u in the graph data.

[0167] Next, after aggregating the neighborhood node information of the central node and its edge information, the original features of the central node are Neighbor representations with which it is aggregated concatenate them, and input this concatenated representation into a fully connected layer with a nonlinear factor σ to obtain and will As the result after K aggregations.

[0168] Finally, after the Kth aggregation, the obtained results are normalized, and the corresponding predicted embedding vector can be obtained for the node u as the central node Similarly, the corresponding prediction embedding vector (or node representation) can be obtained for each node, and then the corresponding node representation matrix Z can be obtained by aggregating the prediction embedding vectors.

[0169] For example, see Figure 3B As shown, it is a schematic diagram of the direction of sampling neighborhood nodes and the characteristics of aggregation neighborhood nodes in the embodiment of the present application. Figure 3B As shown in the figure, assuming that the depth value is set to 2 in the graph neural network model, then when the preset neighborhood sampling function is used and node 1 is used as the center node for sampling, we can get Figure 3B The neighborhood node sampling results marked by the thick solid line points are shown in FIG. 1 , where the depth is 2, and the neighborhood nodes obtained by sampling include nodes that are one hop away from node 1 and nodes that are two hops away from node 1;

[0170] Furthermore, continue to combine Figure 3B From the schematic content, we can see that when generating the corresponding prediction embedding vector for node 1, the comprehensive features associated with each neighboring node are aggregated. The relevant aggregation method is as follows: Figure 3B As shown, it can be seen that the sampling direction and the aggregation direction are completely opposite.

[0171] For example, see Figure 3C As shown, it is a schematic diagram of the process of aggregating the comprehensive features of neighborhood nodes in the embodiment of the present application, combined with Figure 3B and Figure 3C The contents of Figure 3B Based on the schematic content, when performing feature aggregation on node 1, the specific aggregation process is as follows Figure 3C As shown in the figure, the larger the depth value, the features farther from the central node can be aggregated; when K is 2, at the first layer, node 1 aggregates the comprehensive features of two neighboring nodes (node ​​2 and node 5) with a distance of 1 hop in the 0th layer. For node 2, the comprehensive features of nodes 3 and 4 in the 0th layer are aggregated. For node 5, the comprehensive features of node 6 in the 0th layer are aggregated.

[0172] In the second layer, node 1 aggregates the comprehensive features of nodes 2 and 5 in the first layer. For node 2, the comprehensive features of nodes 3 and 4 in the first layer are aggregated. For node 5, the comprehensive features of node 6 in the first layer are aggregated.

[0173] In this way, the processing device can obtain the predicted embedding vectors obtained by the model for each node with the help of the graph neural network model, which is equivalent to performing feature aggregation for each node in the model content to obtain the corresponding embedding vector representation. Moreover, the aggregated features include edge interaction features, operation time features, and node features, which is equivalent to considering the influence of time and operation conditions when generating the embedding vector of the node, which can enhance the sensitivity of the generated prediction embedding vector to time and help to better analyze and explore the operation conditions between the target object and the business object.

[0174] Step 302: The processing device determines the node pairs with neighboring relationships connected to each connection edge, and determines the reference nodes in each node pair, and adjusts the model parameters based on the similarity of the predicted embedding vectors between the reference nodes corresponding to each node pair and the domain nodes, and between the reference nodes and other non-neighboring nodes.

[0175] Specifically, after the processing device uses the graph neural network model to generate a corresponding predicted embedding vector for each node, in order to realize unsupervised training of the graph neural network model, the processing device samples non-neighborhood nodes corresponding to the reference node in the node pair for each node pair connected by each connecting edge and having a neighborhood relationship; then adjusts the model parameters based on the similarity of the predicted embedding vectors between the reference node and the neighboring nodes, and between the reference node and the non-neighboring nodes, where the reference node is an arbitrarily selected node from the two nodes connected by the connecting edge.

[0176] In the embodiment of the present application, the processing device may use the following formula to calculate the unsupervised loss, and then adjust the model parameters according to the unsupervised loss. In the embodiment of the present application, the theory of unsupervised learning is based on the assumption that the neighboring nodes have similar predicted embedding vectors, while the nodes without intersection (i.e., non-neighboring nodes) have dissimilar predicted embedding vectors:

[0177]

[0178] Among them, z u is the predicted embedding vector corresponding to node u, z p It can be the predicted embedding vector corresponding to node p after node u determines the neighboring node p through random walk, v n ~P n (v) represents negative sampling (i.e., sampling to obtain non-neighborhood nodes), Q is the number of samples (i.e., the total number of non-neighborhood nodes); s() represents an inner product operation or a full connection operation.

[0179] Furthermore, based on the obtained unsupervised loss, the processing device adjusts the model parameters in the graph neural network model; similarly, the processing device uses the predicted embedding vector currently generated for each node as the initialized embedding vector corresponding to each node in the next round of training, and continues to iteratively train the adjusted neural network model until the preset convergence conditions are met, and finally obtains the trained target model and the target embedding vector constructed by the target model for each node.

[0180] Afterwards, the processing device may make a business recommendation for the access object according to the target embedding vectors corresponding to each target object and each business object in the specified business scenario.

[0181] It should be noted that in the embodiments of the present application, when implementing business recommendations based on the target embedding vectors corresponding to the business object and the target object, the recommended business object can be determined by directly calculating the similarity between the vectors, or a corresponding recommendation network can be constructed for the business recommendation task, and sample labels can be constructed based on the target object's historical operations on the business object, and the recommendation network can be trained to learn the target embedding vectors based on the target object and the business object to determine whether the target object will operate on the business object, and then determine the business object recommended to the target object based on the trained target recommendation model.

[0182] For example, the processing device may use the following function to predict whether the target object will operate the business object, that is, to predict whether the recommended business object is effective for the target object, wherein the evaluation method of effectiveness is: whether the recommended business object can be operated by the target object:

[0183]

[0184] Among them, s() is the inner product operation, σ is the nonlinear factor, Zu is the target embedding vector corresponding to the access object, and Zh is the target embedding vector of a business object obtained.

[0185] The following description only takes the example of determining the recommended business object by calculating the similarity between the embedded vectors to illustrate the relevant processing process.

[0186] In some feasible embodiments, the processing device may search for a target embedding vector that matches the access object in response to an access indication of the access object; when a target embedding vector that matches the access object is found, based on the target embedding vector corresponding to the access object and the vector similarity between the target embedding vectors of each business object, determine the target business object that is adapted to the access object, and recommend the target business object to the access object.

[0187] Specifically, when a corresponding target embedding vector is generated in advance for the access object, the vector similarity of the target embedding vector between the access object and the business object can be calculated first, and then the business object whose corresponding vector similarity exceeds a set threshold can be determined as the target business object that is compatible with the access object, and the target business object can be recommended to the access object.

[0188] In this way, since the generated target embedding vector integrates the operation time features, edge interaction features, and node features, it can accurately characterize the characteristics of each business object and each target object in the specified business scenario. Therefore, with the help of the target embedding vector corresponding to each object, it is possible to accurately determine the content that the access object may be interested in within the object range composed of various business objects in the specified business scenario in the specific business processing process, targeting the access object, thereby avoiding the content limitation caused by only considering the historical data of its own operations.

[0189] In particular, when the processing device fails to find a target embedding vector that matches the access object, it can determine the historical business objects operated by the access object based on the historical operations of the access object in the specified business scenario, and determine the target embedding vector that matches the historical business objects in each target embedding vector constructed for each node in the target model; then, based on the target embedding vector of the historical business object, generate a target embedding vector corresponding to the access object, and based on the target embedding vector of the access object and the vector similarity between the target embedding vector of each business object, determine the target business object recommended to the access object.

[0190] Specifically, when there is no target embedding vector corresponding to the access object in each target embedding vector constructed by the target model, but there is a target embedding vector corresponding to some or all of the historical business objects operated by the access object, the processing device determines the target embedding vector corresponding to the historical business object operated by the access object in each target embedding vector constructed for each node; and then performs weighted average processing on the target embedding vectors of the historical business objects to obtain the target embedding vector corresponding to the access object;

[0191] Afterwards, by calculating the vector similarity, a target business object that can be recommended to the access object is determined from among the business objects.

[0192] In this way, when the corresponding target embedding vector is not constructed for the access object, the target embedding vector of the access object can be equivalently obtained based on the target embedding vectors of the historical business objects operated by the access object, so that the recommended business content can be determined for this type of access object.

[0193] In other feasible embodiments, the processing device responds to an access indication of an access object and obtains the access time carried in the access indication; when it is determined that the access object is a target object in the graph data, a time constraint determined according to the access time is added to the neighborhood sampling function in the target model, and the adjusted target model is used to update the target embedding vectors corresponding to the access object and each business object; thereafter, based on the updated target embedding vector and the vector similarity between the target embedding vectors of each business object, the target business object recommended to the access object is determined.

[0194] Specifically, when the processing device determines the target business object that can be recommended for the access object, it can add a time constraint determined according to the access time in the neighborhood sampling function within the target model based on the access time of the access object, so that the method of determining the neighborhood nodes in the target model is adjusted; then, the processing device re-updates the target embedding vector corresponding to each node based on the model parameters obtained by previous training and the adjusted neighborhood sampling function, and obtains the updated target embedding vector corresponding to each node under the condition of the access time constraint;

[0195] Afterwards, based on the updated target embedding vector, by calculating the vector similarity, the target business object that can be recommended to the access object is determined from among the business objects.

[0196] In this way, according to actual processing needs, the method of determining the neighborhood nodes can be adjusted according to the access time of the access object, and the neighborhood nodes whose corresponding operation time is adapted to the access time can be selected. Then, by aggregating the comprehensive features of the appropriately matched neighborhood nodes, the target embedding vector corresponding to each node can be retrieved, so that the content that the access object may be interested in at the current access time can be recommended to the access object, which helps to improve the effectiveness of the recommendation.

[0197] It should be noted that in the embodiment of the present application, the time feature extraction subnetwork for generating operation time features is included in the graph neural network model and is jointly trained with other networks. The present application compares the joint training of the time feature extraction subnetwork and the graph neural network subnetwork, and finally obtains the target model, as well as the pre-trained time feature extraction subnetwork. When the target model is finally obtained, the recommendation accuracy obtained based on each target embedding vector constructed based on the target model is found. The target model trained by the joint training method adopted in the present application has the highest recommendation accuracy, where the recommendation accuracy refers to the probability of the user operating the recommended business object.

[0198] See Table 2, which is a schematic diagram of the difference in recommendation accuracy of the models in the embodiments of the present application.

[0199] Table 2

[0200] Recommendation accuracy Most frequently recommended model 6.91％ Frequency Top3 Recommended Model 11.62％ Static graph model 50.15％ Model for pre-training temporal feature extraction subnetwork 50.92％ This application (joint training of temporal feature extraction subnetwork) 51.84％

[0201] Among them, the highest frequency recommendation model and the frequency Top3 recommendation model are implemented using the processing method described in the background technology; the static graph model, the model of the pre-trained time feature extraction subnetwork, and the model of the jointly trained time feature extraction subnetwork are model structures that are continuously explored and determined by this application, among which the static graph model refers to a graph neural network model that does not consider the influence of time features.

[0202] In the following, firstly, in conjunction with the accompanying drawings, taking the travel navigation business scenario as an example, the historical operation data for the travel navigation scenario is analyzed, the target embedding vectors corresponding to each user and each location in the travel navigation scenario are constructed, and the process of recommending travel destinations based on each target embedding vector is explained.

[0203] See also Figure 4 As shown, it is a schematic diagram of the process of constructing graph data for the travel navigation scenario in an embodiment of the present application. When performing information processing for the travel navigation scenario, the processing device can obtain each user log for the users authorized to perform log analysis, and filter the target field from each user log to construct a user information data set and a user travel record data set, and then construct the corresponding graph data based on the user information data and the user travel record data set.

[0204] The original log content format:

[0205]

[0206] in, is the log length, k is the log record number, uin is the user's unique identifier, features is the user's public feature information (or user's public attribute information), event_value is the event value of the user generating the travel log at time, where the event value contains the user's travel information and is composed of a large number of information strings connected by &. The event_value format is: key1=v1&key2=v2&key3=v3..., pname represents the location value, which is used to identify the travel location selected by the user and is included in event_value.

[0207] Combined with Figure 4 As shown, attached Figure 4Each historical operation data corresponds to each user log. To extract valid fields, & is used as a segmentation mark to segment event_value and obtain a field sequence (pname,…,field), where field represents the travel feature value of a travel event that occurred to the user (i.e., user uin interacted with location pname once), and each location is determined in each historical operation data based on pname. The travel location and corresponding operation information (or travel information) are obtained in the form of (pname: 'pname', key1: v1, key2: v2,…), where field is a non-pname part and a useful keyi set that is filtered out. The method of defining whether keyi is useful can be determined based on actual processing experience.

[0208] At the same time, considering that a uin and the corresponding pname may appear in multiple historical operation data, the operation information in multiple historical operation data can be aggregated based on the condition that (uin, pname) is the same to obtain the travel information sequence sequence, that is, the collection of the contents corresponding to the field in multiple historical operation data.

[0209] In addition, the processing device filters the user's feature information features to obtain λ, that is, the features disclosed by the user.

[0210] After filtering and aggregation, we can use Ui to represent user uin, Pi to represent location pname, and t i represents the operation time when user Ui interacts with location Pi. Considering that a uin and pname may appear in multiple historical operation data, there will be multiple historical times. The operation time is taken from a series of interactions that are close in time (such as the latest, the middle of the sequence, etc.). i Represents the travel information sequence sequence.

[0211] Based on this, the constructed user travel record dataset is as follows:

[0212] ε{(u 1 ,p 1 ,t 1 ,f 1 ),...,(u M ,p N ,t L ,f L )}

[0213] Among them, M is the number of users, N is the number of locations, and L is the size of the user travel record dataset.

[0214] Taking the user information data set composed of feature sequences λ respectively selected from the feature information data of each user as an example, a user information data set with the following definition is constructed:

[0215] ξ={(u 1 ,λ 1 ),…,(u M ,λ M )}

[0216] Afterwards, the size of the user information data set is M.

[0217] Furthermore, based on the constructed user information data set and user travel record data set, graph data can be constructed.

[0218] In the process of characterizing graph data, an adjacency matrix can be constructed based on user travel record data and user information data set to characterize the connection between nodes.

[0219] Specifically, based on the constructed travel record dataset ε, non-duplicate travel records are screened. According to the actual meaning of the data in ε, if there is a data (u i ,p i ,t i ,f i ), then it represents user u i With location p i In t i An interaction occurs at a time i , that is u i With p i are connected to construct the adjacency matrix A. It should be noted that the node u representing the user i Only with the node p representing the location i Connected, representing the node p of the location i It can only communicate with the node u representing the user i are connected.

[0220] In the process of constructing the feature matrix, based on the constructed user information data set ξ, the location features constructed by the location features and the node features constructed by the attribute information (or feature information) disclosed by the user are set to the same dimension. The value in the location feature can be set to 0, and the value in the node feature corresponding to the user is the value of the user's attribute information after one-hot encoding. The encoded data presents a numerical type, which is convenient for model input. Finally, the node features corresponding to each user are connected with the node features corresponding to the location features to construct the feature matrix Where M is the number of users, N is the number of locations, and d 1 is the node feature dimension. The matrices A and X are expressed as follows:

[0221]

[0222]

[0223] Among them, A is the adjacency matrix, X is the node feature matrix, d1>=the dimension after onehot of user information, the first M rows in X are the node features corresponding to user nodes, and the last N rows are the node features corresponding to location nodes, using the same feature space. Because there is no node feature information initially, some positions are filled with 0.

[0224] When constructing the adjacency matrix and feature matrix to carry the graph data, according to the connection relationship between nodes represented by the adjacency matrix, for the graph data G(V,E,X,F), V represents the node set, v i ∈V, defined as follows:

[0225]

[0226] represents the edge set, where It has the following meanings: There is a record (u i ,p i ,t i ,f i ), there is a record (u i ,p i ),u i ,p i There is a connecting edge in the travel graph G, which actually means that user u i With location p i At time t i An interaction occurs, forming a connecting edge F represents the edge interaction feature matrix, L is the length of the dataset ε, d 2 is the characteristic dimension of the mutual information, where Indicates that there is a record {(u i ,p i ,t i ,f i )},f i is the operation information group of this interaction. After encoding based on fi, the corresponding edge interaction feature is obtained, which is edge features.

[0227] Furthermore, the adjacency matrix, node feature matrix, operation time feature matrix, and edge interaction feature matrix are input into the graph neural sub-network to construct the corresponding embedding vector for each node.

[0228] In this way, by analyzing and constructing graph data from a large amount of location-related spatiotemporal data, and using graph neural network technology to learn the embedded representation of each user and each location in a low-dimensional space based on the graph data, it is possible to model and predict the user's travel behavior with the user's knowledge and authorization, identify their travel behavior intentions, and recommend places that may be of interest to the user.

[0229] For further information, see Figure 5 As shown, it is a schematic diagram of presenting location recommendations for visiting objects in an embodiment of the present application. Figure 5 As can be seen from the schematic content, after the visiting object accesses the travel navigation application Q, the processing device determines the target location that meets the required similarity among various locations based on the target embedding vector corresponding to the visiting object, and determines the target location as the location recommended to the visiting object; further, in response to the visiting object's indication of starting to type in the search box on page 1, the recommended travel location can be displayed.

[0230] Based on the same inventive concept, see Figure 6 As shown, it is a schematic diagram of the logical structure of the information processing device in the embodiment of the present application. The information processing device 600 includes a construction unit 601 and a training unit 602, wherein:

[0231] The construction unit 601 is used to construct graph data based on the historical operation of each target object on the business object in the specified business scenario; the graph data includes: a node information set constructed based on the nodes corresponding to each target object and each business object, a time information set constructed by the operation time corresponding to the connection edge between the nodes, the node feature of each node and the edge interaction feature of each connection edge; the edge interaction feature represents: the time sequence operation of the target object on the business object;

[0232] The training unit 602 is used to perform multiple rounds of iterative training on the initial graph neural network model based on the graph data to obtain a trained target model and a target embedding vector constructed by the target model for each node, wherein in one round of iterative training, the following operations are performed:

[0233] The graph neural network model is used to extract the corresponding operation time features for each operation time, and the following operations are performed for each node: the node features, operation time features, and edge interaction features corresponding to each associated neighboring node are aggregated to obtain the corresponding prediction embedding vector;

[0234] Determine the node pairs with neighboring relationships connected by each connection edge, determine the reference nodes in each node pair, and adjust the model parameters based on the similarity of the predicted embedding vectors between the reference node and the domain node corresponding to each node pair, and between the reference node and other non-neighboring nodes.

[0235] Optionally, when constructing edge interaction features in graph data, the construction unit 601 is used to:

[0236] Based on the historical operation data representing the historical operation conditions associated with each target object in the specified business scenario, the operation information of each target object operating the business object at different historical times is aggregated;

[0237] Corresponding to each target object and each business object covered by each historical operation data, each node is constructed, and according to the operation relationship between the target object and the business object, a connection edge is constructed between the corresponding nodes;

[0238] Based on the aggregated operation information groups corresponding to each connection edge, the corresponding edge interaction features are generated respectively.

[0239] Optionally, when aggregating the operation information of each target object operating a business object at different historical times, the construction unit 601 is used to include:

[0240] For each historical operation data associated with each target object, the data content of each historical operation data is split according to the preset segmentation identifier to determine the operation information of the target object on the business object at different historical times;

[0241] For the target object and each business object of the operation, the operation information of different historical times is aggregated according to the information type to obtain an aggregated operation information group.

[0242] Optionally, when generating corresponding edge interaction features based on the aggregated operation information groups corresponding to the respective connection edges, the construction unit 601 is used to:

[0243] Using a preset encoding technology, for each aggregated operation information group, different types of information content in the operation information group are encoded to obtain corresponding edge interaction features.

[0244] Optionally, when constructing a time information set in the graph data, the construction unit 601 is used to:

[0245] For each historical operation data associated with each target object, based on the historical times when the target object operates the same business object recorded in different historical operation data, the corresponding operation time is obtained;

[0246] Based on the operation time of each target object on the operated business object, a time information set is obtained.

[0247] Optionally, when aggregating the node features, operation time features, and edge interaction features corresponding to each of the associated neighborhood nodes to obtain a corresponding prediction embedding vector, the training unit 602 is used to:

[0248] According to the preset model depth and neighborhood sampling function, determine each neighboring node associated with the node;

[0249] A temporal edge aggregator corresponding to the model depth is used to aggregate the comprehensive features corresponding to each associated neighboring node for each node to obtain the corresponding prediction embedding vector, where the comprehensive features are obtained by concatenating the node features associated with the neighboring nodes, the operation time features, and the edge interaction features.

[0250] Optionally, after obtaining the trained target model and the target embedding vector respectively constructed by the target model for each node, the device further includes a recommendation unit 603, and the recommendation unit 603 is further used to:

[0251] In response to an access indication of an access object, searching for a target embedding vector matching the access object;

[0252] When a target embedding vector matching the access object is found, based on the vector similarity between the target embedding vector corresponding to the access object and the target embedding vectors of each business object, a target business object that matches the access object is determined and recommended to the access object.

[0253] Optionally, when no target embedding vector matching the access object is found, the recommendation unit 603 is used to:

[0254] Based on the historical operation of the access object in the specified business scenario, determine the historical business object operated by the access object, and determine the target embedding vector matching the historical business object in each target embedding vector constructed by the target model for each node;

[0255] Based on the target embedding vectors of historical business objects, a target embedding vector corresponding to the access object is generated, and based on the vector similarity between the target embedding vector of the access object and the target embedding vectors of each business object, a target business object recommended to the access object is determined.

[0256] Optionally, after obtaining the trained target model and the target embedding vector respectively constructed by the target model for each node, the device further includes a recommendation unit 603, and the recommendation unit 603 is further used to:

[0257] In response to an access instruction for accessing the object, obtaining an access time carried in the access instruction;

[0258] When it is determined that the access object is the target object in the graph data, a time constraint determined according to the access time is added to the neighborhood sampling function in the target model, and the target embedding vectors corresponding to the access object and each business object are updated using the adjusted target model;

[0259] Based on the updated target embedding vector and the vector similarity between the target embedding vectors of each business object, the target business object recommended to the access object is determined.

[0260] After introducing the information processing method and apparatus according to the exemplary embodiment of the present application, next, an electronic device according to another exemplary embodiment of the present application is introduced.

[0261] Those skilled in the art will appreciate that various aspects of the present application may be implemented as a system, method or program product. Therefore, various aspects of the present application may be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software, which may be collectively referred to as "circuit", "module" or "system" herein.

[0262] Based on the same inventive concept as the above method embodiment, an electronic device is also provided in the embodiment of the present application, referring to Figure 7 As shown, it is a schematic diagram of a hardware structure of an electronic device using an embodiment of the present application, and the electronic device 700 may include at least a processor 701 and a memory 702. Among them, the memory 702 stores program codes, and when the program codes are executed by the processor 701, the processor 701 performs any of the above-mentioned information processing steps.

[0263] In some possible implementations, the computing device according to the present application may include at least one processor and at least one memory. The memory stores program code, and when the program code is executed by the processor, the processor performs the steps of information processing according to various exemplary implementations of the present application described above in this specification. For example, the processor may perform the following steps: Figure 2A Follow the steps shown in .

[0264] Refer to the following Figure 8 800 according to this embodiment of the present application. Figure 8 As shown, the computing device 800 is in the form of a general computing device. The components of the computing device 800 may include but are not limited to: at least one processing unit 801, at least one storage unit 802, and a bus 803 connecting different system components (including the storage unit 802 and the processing unit 801).

[0265] Bus 803 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a processor, or a local bus using any of a variety of bus architectures.

[0266] The storage unit 802 may include a readable medium in the form of a volatile memory, such as a random access memory (RAM) 8021 and / or a cache memory 8022 , and may further include a read-only memory (ROM) 8023 .

[0267] The storage unit 802 may also include a program / utility 8025 having a set (at least one) of program modules 8024, such program modules 8024 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0268] The computing device 800 may also communicate with one or more external devices 804 (e.g., keyboards, pointing devices, etc.), may also communicate with one or more devices that enable objects to interact with the computing device 800, and / or communicate with any device that enables the computing device 800 to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface 805. In addition, the computing device 800 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 806. As shown, the network adapter 806 communicates with other modules for the computing device 800 via a bus 803. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the computing device 800, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0269] Based on the same inventive concept as the above method embodiment, various aspects of information processing provided by the present application can also be implemented in the form of a program product, which includes program code. When the program product is run on an electronic device, the program code is used to enable the electronic device to execute the steps of the information processing method according to various exemplary embodiments of the present application described above in this specification. For example, the electronic device can execute the following steps: Figure 2A Steps shown.

[0270] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0271] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0272] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. An information processing method, characterized in that: Used for training graph neural network models, including: Based on the historical operation of each target object on the business object in the specified business scenario, graph data is constructed; the graph data includes: a node information set constructed based on the nodes corresponding to each target object and each business object, a time information set constructed by the operation time corresponding to the connection edge between the nodes, the node features of each node and the edge interaction features of each connection edge; the edge interaction features represent: the time sequence operation of the target object on the business object; Based on the graph data, the initial graph neural network model is iteratively trained for multiple rounds to obtain a trained target model and a target embedding vector constructed by the target model for each node, wherein in one round of iterative training, the following operations are performed: The graph neural network model is used to extract the corresponding operation time features for each operation time, and the following operations are performed for each node: the node features, operation time features and edge interaction features corresponding to each associated neighboring node are aggregated to obtain the corresponding prediction embedding vector; Determine the node pairs with neighborhood relationships connected by each connection edge, determine the reference nodes in each node pair, and adjust the model parameters based on the similarity of the predicted embedding vectors between the reference node and the domain node corresponding to each node pair, and between the reference node and other non-neighborhood nodes.

2. The method according to claim 1, characterized in that When constructing edge interaction features in graph data, we include: Based on the historical operation data representing the historical operation conditions associated with each target object in the specified business scenario, the operation information of the business objects operated by each target object at different historical times is aggregated; Corresponding to each target object and each business object covered by each historical operation data, each node is constructed, and according to the operation relationship between the target object and the business object, a connection edge is constructed between the corresponding nodes; Based on the aggregated operation information groups corresponding to each connection edge, the corresponding edge interaction features are generated respectively.

3. The method according to claim 2, characterized in that The aggregating the operation information of the target objects operating the business objects at different historical times includes: For each historical operation data associated with each target object, data content of each historical operation data is split according to a preset segmentation identifier to determine the operation information of the target object on the business object at different historical times; For the target object and each business object of the operation, the operation information at different historical times is aggregated according to the information type to obtain an aggregated operation information group.

4. The method according to claim 2, characterized in that The generating corresponding edge interaction features based on the aggregated operation information groups corresponding to the respective connection edges includes: By adopting a preset encoding technology, for each aggregated operation information group, different types of information content in the operation information group are encoded to obtain corresponding edge interaction features.

5. The method according to claim 1, characterized in that Construct a time information set in the graph data, including: For each historical operation data associated with each target object, based on each historical time recorded in different historical operation data when the target object operates the same business object, the corresponding operation time is obtained; Based on the operation time of each target object on the operated business object, a time information set is obtained.

6. The method according to claim 1, characterized in that The node features, operation time features and edge interaction features corresponding to each of the neighborhood nodes of the aggregation association are respectively obtained to obtain the corresponding prediction embedding vector, including: Determine each neighboring node associated with the node according to a preset model depth and neighborhood sampling function; A temporal edge aggregator corresponding to the depth of the model is used to aggregate the comprehensive features corresponding to each of the associated neighboring nodes for the node to obtain a corresponding prediction embedding vector, wherein the comprehensive features are obtained by concatenating the node features, operation time features, and edge interaction features associated with the neighboring nodes.

7. The method according to any one of claims 1 to 6, characterized in that: After obtaining the trained target model and the target embedding vectors respectively constructed by the target model for each node, the method further includes: In response to an access indication of an access object, searching for a target embedding vector matching the access object; When a target embedding vector matching the access object is found, a target business object adapted to the access object is determined based on the target embedding vector corresponding to the access object and the vector similarity between the target embedding vectors of each business object, and the target business object is recommended to the access object.

8. The method according to claim 7, characterized in that When a target embedding vector matching the access object is not found, including: Based on the historical operation of the access object in the specified business scenario, determine the historical business object operated by the access object, and determine the target embedding vector matching the historical business object in each target embedding vector constructed by the target model for each node; Based on the target embedding vector of the historical business object, a target embedding vector corresponding to the access object is generated, and based on the vector similarity between the target embedding vector of the access object and the target embedding vectors of each business object, a target business object recommended to the access object is determined.

9. The method according to any one of claims 1 to 6, characterized in that: After obtaining the trained target model and the target embedding vectors respectively constructed by the target model for each node, the method further includes: In response to an access instruction of accessing the object, obtaining an access time carried in the access instruction; When it is determined that the access object is a target object in the graph data, a time constraint determined according to the access time is added to a neighborhood sampling function in the target model, and the target embedding vectors corresponding to the access object and each business object are updated using the adjusted target model; Based on the updated target embedding vector and the vector similarity between the target embedding vectors of each business object, a target business object recommended to the access object is determined.

10. An information processing device, characterized in that: Used for training graph neural network models, including: A construction unit is used to construct graph data based on the historical operation of each target object on the business object in a specified business scenario; the graph data includes: a node information set constructed based on the nodes corresponding to each target object and each business object, a time information set constructed by the operation time corresponding to the connection edge between the nodes, a node feature of each node and an edge interaction feature of each connection edge; the edge interaction feature represents: the time sequence operation of the target object on the business object; A training unit is used to perform multiple rounds of iterative training on the initial graph neural network model based on the graph data to obtain a trained target model and a target embedding vector constructed by the target model for each node, wherein in one round of iterative training, the following operations are performed: The graph neural network model is used to extract the corresponding operation time features for each operation time, and the following operations are performed for each node: the node features, operation time features and edge interaction features corresponding to each associated neighboring node are aggregated to obtain the corresponding prediction embedding vector; Determine the node pairs with neighborhood relationships connected by each connection edge, determine the reference nodes in each node pair, and adjust the model parameters based on the similarity of the predicted embedding vectors between the reference node and the domain node corresponding to each node pair, and between the reference node and other non-neighborhood nodes.

11. The device according to claim 10, characterized in that When constructing edge interaction features in graph data, the construction unit is used to: Based on the historical operation data representing the historical operation conditions associated with each target object in the specified business scenario, the operation information of the business objects operated by each target object at different historical times is aggregated; Corresponding to each target object and each business object covered by each historical operation data, each node is constructed, and according to the operation relationship between the target object and the business object, a connection edge is constructed between the corresponding nodes; Based on the aggregated operation information groups corresponding to each connection edge, the corresponding edge interaction features are generated respectively.

12. The device according to any one of claims 10 to 11, characterized in that: After obtaining the trained target model and the target embedding vectors respectively constructed by the target model for each node, the device further includes a recommendation unit, which is further used to: In response to an access instruction of accessing the object, obtaining an access time carried in the access instruction; When it is determined that the access object is a target object in the graph data, a time constraint determined according to the access time is added to a neighborhood sampling function in the target model, and the target embedding vectors corresponding to the access object and each business object are updated using the adjusted target model; Based on the updated target embedding vector and the vector similarity between the target embedding vectors of each business object, a target business object recommended to the access object is determined.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

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

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images