Object resource recommendation method and device, storage medium and electronic equipment

Abstract

The invention discloses an object resource recommendation method and device, a storage medium and electronic equipment. The method comprises the steps that based on object resources issued by a target platform and candidate user accounts stored in a cold start pool corresponding to the target platform, a target local relation graph matched with the target platform is constructed, and the target local relation graph comprises an evaluation relation used for indicating the candidate user accounts and the object resources; a recommendation coefficient between each candidate user account and each object resource is determined by using the target local relation graph and a global relation graph, and the global relation graph is a relation graph jointly constructed based on the target local relation graph and a reference local relation graph matched with a reference platform except the target platform; and according to the recommendation coefficient, determining a target object resource to be recommended for each candidate user account. According to the resource recommendation method and device, the technical problem that the resource recommendation accuracy is low in a resource recommendation method provided by the related technology is solved.

Description

Technical Field

[0001] The present application relates to the field of computers, and in particular to a method and device for recommending object resources, a storage medium, and an electronic device. Background Art

[0002] Many application platforms often experience a cold start period during their initial launch. This refers to the limited number of users who directly engage with and use the published resources on the platform, without any prior promotional traffic. Currently, to address this cold start issue, technologies are using historical user data from other similar platforms to quickly recommend each published resource to interested users.

[0003] However, when referencing historical user data generated on other similar platforms in this approach, user accounts that have performed resource interactions on these platforms are typically retrieved from this historical user data and then directly used as candidate user accounts for resource recommendations on the current platform in a cold start state. In other words, the resource recommendation methods provided by related technologies are relatively simple in their approach, resulting in low accuracy in resource recommendations.

[0004] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0005] The embodiments of the present application provide a method and apparatus for recommending object resources, a storage medium, and an electronic device, so as to at least solve the technical problem of low accuracy of resource recommendation existing in the resource recommendation methods provided by related technologies.

[0006] According to one aspect of an embodiment of the present application, a method for recommending object resources is provided, comprising: constructing a target local relationship graph matching the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes a graph for indicating the evaluation relationship between the candidate user accounts and the object resources; determining a recommendation coefficient between each candidate user account and each object resource by using the target local relationship graph and the global relationship graph, wherein the global relationship graph is a relationship graph jointly constructed based on the target local relationship graph and a reference local relationship graph matching a reference platform other than the target platform; and determining the target object resource to be recommended for each candidate user account according to the recommendation coefficient.

[0007] According to another aspect of an embodiment of the present application, a device for recommending object resources is also provided, including: a construction unit for constructing a target local relationship graph matching the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship for indicating the relationship between the candidate user accounts and the object resources; a first determination unit for determining the recommendation coefficient between each candidate user account and each object resource using the target local relationship graph and the global relationship graph, wherein the global relationship graph is a relationship graph jointly constructed based on the target local relationship graph and a reference local relationship graph matching a reference platform other than the target platform; a second determination unit for determining the target object resource to be recommended for each candidate user account according to the recommendation coefficient.

[0008] According to another aspect of the embodiments of the present application, a computer-readable storage medium is provided, in which a computer program is stored. The computer program is configured to execute the above-mentioned object resource recommendation method at runtime.

[0009] According to another aspect of an embodiment of the present application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the above-described method for recommending object resources.

[0010] According to another aspect of an embodiment of the present application, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the object resource recommendation method through the computer program.

[0011] In an embodiment of the present application, a target local relationship graph that matches the target platform is constructed based on the object resources released by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship for indicating the candidate user accounts and the object resources. Then, the target local relationship graph and the global relationship graph are used to determine the recommendation coefficient between each candidate user account and each object resource, wherein the global relationship graph is a relationship graph jointly constructed based on the target local relationship graph and the reference local relationship graph matched by the reference platform other than the target platform. Then, the target object resource to be recommended is determined for each candidate user account according to the recommendation coefficient. In other words, using an embodiment of the present application, the target local relationship graph of the target platform constructed based on the object resource body in the target platform and the candidate user accounts in the cold start pool, and the global relationship graph constructed by combining the target local relationship graph and the reference local relationship graph corresponding to the reference platform other than the target platform, is used to obtain the recommendation coefficient between each candidate user account corresponding to the target platform and each object resource. This makes the resulting recommendation coefficient more reliable, allowing the target resource to be recommended for each candidate user account based on the recommendation coefficient, adapting to the actual needs of the candidate user account. This avoids the problem of low resource recommendation accuracy caused by the relatively simple recommendation methods provided by related technologies, thereby achieving the technical effect of improving the accuracy of resource recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0013] Figure 1 is a schematic diagram of an application environment of an optional object resource recommendation method according to an embodiment of the present application;

[0014] Figure 2 is a flowchart of an optional method for recommending object resources according to an embodiment of the present application;

[0015] Figure 3 is a schematic diagram of an optional method for recommending object resources according to an embodiment of the present application;

[0016] Figure 4 is a schematic diagram of an optional method for recommending object resources according to an embodiment of the present application;

[0017] Figure 5 is a schematic diagram of an optional method for recommending object resources according to an embodiment of the present application;

[0018] Figure 6 is a schematic diagram of an optional method for recommending object resources according to an embodiment of the present application;

[0019] Figure 7 is a schematic diagram of an optional method for recommending object resources according to an embodiment of the present application;

[0020] Figure 8 is a flowchart of an optional method for recommending object resources according to an embodiment of the present application;

[0021] Figure 9 is a schematic structural diagram of an optional device for recommending object resources according to an embodiment of the present application;

[0022] Figure 10 It is a schematic structural diagram of an optional electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] According to one aspect of the embodiment of the present application, a method for recommending object resources is provided. Optionally, as an optional implementation, the method for recommending object resources can be applied to, but is not limited to, Figure 1 In the environment shown. Figure 1As shown, terminal device 102 includes a memory 104 for storing various data generated during the operation of terminal device 102, a processor 106 for processing and calculating the aforementioned data, and a display 108 for displaying the target local relationship graph and the global relationship graph. Terminal device 102 can exchange data with server 112 via network 110. Server 112 is connected to database 114, which is used to store various data. Terminal device 102 can run an application for executing the aforementioned object resource recommendation method.

[0026] Furthermore, the above method Figure 1 The specific application process in the environment shown is as follows:

[0027] Step S102 , upon receiving a resource recommendation request, the terminal device 102 sends a request message to the server 112 via the network 110 requesting the server to determine a target resource to be recommended for each candidate user account.

[0028] In step S104, upon receiving the request, the server 112 constructs a target local relationship graph that matches the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform. The target local relationship graph includes an evaluation relationship indicating the relationship between the candidate user accounts and the object resources.

[0029] In step S106, the server 112 determines the recommendation coefficient between each candidate user account and each object resource using the target local relationship graph and the global relationship graph, wherein the global relationship graph is a relationship graph constructed based on the target local relationship graph and reference local relationship graphs matched with reference platforms other than the target platform.

[0030] Step S108 , the server 112 determines the target object resource to be recommended for each candidate user account based on the recommendation coefficient;

[0031] In step S110 , the server 112 sends the target object resources to be recommended for each candidate user account to the terminal device 102 via the network 110 .

[0032] In an embodiment of the present application, a target local relationship graph that matches the target platform is constructed based on the object resources released by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship for indicating the candidate user accounts and the object resources. Then, the target local relationship graph and the global relationship graph are used to determine the recommendation coefficient between each candidate user account and each object resource, wherein the global relationship graph is a relationship graph jointly constructed based on the target local relationship graph and the reference local relationship graph matched by the reference platform other than the target platform. Then, the target object resource to be recommended is determined for each candidate user account according to the recommendation coefficient. In other words, using an embodiment of the present application, the target local relationship graph of the target platform constructed based on the object resource body in the target platform and the candidate user accounts in the cold start pool, and the global relationship graph constructed by combining the target local relationship graph and the reference local relationship graph corresponding to the reference platform other than the target platform, is used to obtain the recommendation coefficient between each candidate user account corresponding to the target platform and each object resource. This makes the resulting recommendation coefficient more reliable, allowing the target resource to be recommended for each candidate user account based on the recommendation coefficient, adapting to the actual needs of the candidate user account. This avoids the problem of low resource recommendation accuracy caused by the relatively simple recommendation methods provided by related technologies, thereby achieving the technical effect of improving the accuracy of resource recommendations.

[0033] Optionally, in this embodiment, the above-mentioned terminal device can be a terminal device configured with a target client, which can include but is not limited to at least one of the following: a mobile phone (such as an Android phone, an iOS phone, etc.), a laptop computer, a tablet computer, a PDA, an MID (Mobile Internet Devices), a PAD, a desktop computer, a smart TV, etc. The target client can be a video client, an instant messaging client, a browser client, an education client, etc. The above-mentioned network can include but is not limited to: a wired network, a wireless network, wherein the wired network includes: a local area network, a metropolitan area network and a wide area network, and the wireless network includes: Bluetooth, WIFI and other networks that realize wireless communication. The above-mentioned server can be a single server, or it can be a server cluster composed of multiple servers, or a cloud server. The above is only an example, and this embodiment does not impose any limitation on this.

[0034] Alternatively, as an alternative, Figure 2 As shown, the recommended methods for the above object resources include:

[0035] S202, based on the object resources released by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, construct a target local relationship graph that matches the target platform, wherein the target local relationship graph includes an evaluation relationship for indicating the candidate user account and the object resources.

[0036] Optionally, the above-mentioned object resource recommendation method can be applied to the cold start scenario of the platform. Specifically, when the platform is in the cold start phase and the number of users who directly pay attention to and use the object resources published on the platform is very limited, a targeted recommendation scenario is performed for the object resources published on the platform, that is, the object resources published on the platform are recommended to user accounts that are interested in the object resources. In addition, the above-mentioned object resource recommendation method can be applied to the resource delivery adjustment scenario of the platform. Specifically, the interest level of user accounts that have visited the platform in different object resources is estimated, and then the overall interest of all user accounts that have visited the platform in the object resources on the platform is predicted, and then the object resources that the public is interested in are given priority in the platform.

[0037] It should be noted that the target platform can be, but is not limited to, an application platform in various fields (e.g., video, gaming, shopping, etc., which is not limited in this embodiment). User accounts using the platform can access the object resources published by the platform through the platform. For further example, assuming that the target platform is an application platform in the video field, the object resources can be, but are not limited to, short videos, film and television works, etc., which is not limited in this embodiment.

[0038] Furthermore, the candidate user accounts stored in the cold start pool may include, but are not limited to, candidate user accounts that have only browsed the target platform, candidate user accounts that have only browsed and clicked on object resources on the target platform, and candidate user accounts that have evaluated the target platform and the object resources published on the target platform. Specifically, the number of candidate user accounts stored in the cold start pool that have evaluated the target platform and the object resources published on the target platform is less than the number of candidate user accounts that have only browsed and clicked on object resources on the target platform; and the number of candidate user accounts that have only browsed and clicked on object resources on the target platform is less than the number of candidate user accounts that have only browsed the target platform.

[0039] Optionally, in this embodiment, the above-mentioned target local relationship graph can be constructed with reference to but not limited to the following steps: taking the candidate user account as the account node and the object resource as the resource node; taking the resource node of each object resource as the current resource node in turn, and performing the following operations: taking the account node of the candidate user account that has performed the evaluation operation on the current object resource corresponding to the current resource node as the current associated account node associated with the current resource node; connecting the current resource node with the current associated account node; when the current resource node is not the last resource node in the target local relationship graph, obtaining the next resource node as the current resource node; when the current resource node is the last resource node in the target local relationship graph, determining to construct a target local relationship graph that matches the target platform.

[0040] It should be noted that in this embodiment, the target local relationship graph may also include, but is not limited to, similarity relationships indicating the similarity between the individual account characteristics of the candidate user accounts. Specifically, the individual account characteristics may be, but are not limited to, obtained based on historical evaluation information published by the candidate user account on the target platform, or historical evaluation information published by associated user accounts associated with the candidate user account on platforms other than the target platform.

[0041] S204, using the target local relationship graph and the global relationship graph, determine the recommendation coefficient between each candidate user account and each object resource, wherein the global relationship graph is a relationship graph constructed based on the target local relationship graph and the reference local relationship graph matched by the reference platform other than the target platform.

[0042] Optionally, in this embodiment, after constructing a target local relationship graph that matches the target platform based on the object resources released by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, it also includes: obtaining the reference local relationship graphs that match the reference platforms, wherein, in the reference local relationship graphs, the reference object resources released by the reference platform serve as reference resource nodes, and the user accounts that have performed evaluation operations on at least one reference object resource in the reference platform serve as reference account nodes. The reference account nodes are connected to the reference resource nodes that have an evaluation association relationship with the reference account nodes, and the evaluation association relationship indicates the user corresponding to the reference account node. The account has performed an evaluation operation on the object resource corresponding to the reference resource node; obtain the union of the reference account node contained in the reference local relationship graph and the account node contained in the target local relationship graph to obtain the global account node in the global relationship graph, and obtain the union of the reference resource node contained in the reference local relationship graph and the resource node contained in the target local relationship graph to obtain the global resource node in the global relationship graph; connect the global account node and the global resource node that has an evaluation association relationship with the global account node to construct a global relationship graph, wherein the evaluation association relationship indicates that the user account corresponding to the global account node has performed an evaluation operation on the object resource corresponding to the global resource node.

[0043] It should be further noted that the recommendation coefficient may be, but is not limited to, used to indicate each candidate user account's estimated evaluation score for each resource. In other words, the recommendation coefficient may be, but is not limited to, used to indicate each candidate user account's level of interest in each resource. The greater the candidate user account's interest in the resource, the higher the estimated evaluation score between the candidate user account and the resource.

[0044] Optionally, in this embodiment, the above-mentioned use of the target local relationship graph and the global relationship graph to determine the recommendation coefficient between each candidate user account and each object resource may include, but is not limited to: determining the node state characteristics based on the point-to-point connection relationship and the point-to-edge connection relationship in the target local relationship graph, and the point-to-point connection relationship and the point-to-edge connection relationship in the global relationship graph, wherein the node state characteristics include the account node state characteristics corresponding to each candidate user account, and the resource node state characteristics corresponding to each object resource; inputting the node state characteristics into the resource recommendation network to obtain the recommendation coefficient of the candidate user account for the above-mentioned object resources, wherein the resource recommendation network is trained using the reference local relationship graph and the target local relationship graph.

[0045] S206: Determine the target object resource to be recommended for each candidate user account according to the recommendation coefficient.

[0046] It should be noted that, in this embodiment, the above-mentioned determination of the target object resource to be recommended for each candidate user account based on the recommendation coefficient may include, but is not limited to: determining the candidate user account and object resource corresponding to the recommendation coefficient greater than the recommendation coefficient threshold, and recommending the object resource to the candidate user account; or, obtaining the object resource corresponding to the recommendation coefficient greater than the recommendation coefficient threshold; obtaining the object resource feature of the object resource based on the resource description information of the object resource, and determining the object resource feature of the object resource as the target resource feature; determining the object resource corresponding to the object resource feature whose feature similarity between the object resource feature and the target resource feature is greater than the resource feature similarity threshold as the target object resource; and increasing the recommendation priority of the target object resource in the target platform. Specifically, the higher the recommendation priority of the object resource, the more eye-catching the recommendation position corresponding to the object resource in the target platform may be, but is not limited to, set.

[0047] Optionally, as an optional embodiment, assuming that the above object resource recommendation method is applied to a cold start scenario of a platform, the above method can be explained by example based on, but not limited to, the following steps:

[0048] S1, during a cold start period of a target platform, constructing a target local relationship graph that matches the target platform based on object resources published by the target platform and candidate user accounts stored in a cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship indicating between candidate user accounts and object resources;

[0049] S2, using the target local relationship graph and a global relationship graph constructed based on the target local relationship graph and reference local relationship graphs matched with reference platforms other than the target platform, to determine a recommendation coefficient indicating the degree of interest of each candidate user account in each object resource;

[0050] S3: Determine candidate user accounts and object resources corresponding to recommendation coefficients greater than a recommendation coefficient threshold, and recommend the object resources to the candidate user accounts.

[0051] As can be seen from the above examples, by adopting the embodiment of the present application, a target local relationship graph of the target platform constructed based on the object resource body in the target platform and the candidate user accounts in the cold start pool, and a global relationship graph constructed by combining the target local relationship graph and the reference local relationship graph corresponding to the reference platform outside the target platform, is used to obtain the recommendation coefficient between each candidate user account corresponding to the target platform and each object resource. This makes the obtained recommendation coefficient more reliable, and further enables the target object resource to be recommended to each candidate user account to be determined separately according to the recommendation coefficient, which can adapt to the actual needs of the above candidate user accounts. This achieves the purpose of effectively recommending object resources in the target platform during the cold start process, so that more user accounts use the object resources in the target platform. This solves the technical problem that during the cold start process, there are fewer user accounts that actually use the object resources in the target platform, resulting in the inability to fully utilize the functions and values ​​of the object resources.

[0052] Optionally, as an optional embodiment, assuming that the above-mentioned object resource recommendation method is applied to a resource placement adjustment scenario of a platform, the above-mentioned method can be explained by example based on, but not limited to, the following steps:

[0053] S1, constructing a target local relationship graph that matches the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship indicating the relationship between the candidate user accounts and the object resources;

[0054] S2, using the target local relationship graph and a global relationship graph constructed based on the target local relationship graph and reference local relationship graphs matched with reference platforms other than the target platform, to determine a recommendation coefficient indicating the degree of interest of each candidate user account in each object resource;

[0055] S3, obtaining the object resource corresponding to the recommendation coefficient greater than the recommendation coefficient threshold; based on the resource description information of the above object resource, obtaining the object resource characteristics of the above object resource, and determining the object resource characteristics of the object resource as the target resource characteristics; determining the object resource corresponding to the object resource characteristics whose feature similarity between the object resource characteristics and the target resource characteristics is greater than the resource feature similarity threshold as the target object resource; and improving the recommendation priority of the above target object resource in the target platform.

[0056] In an embodiment of the present application, a target local relationship graph that matches the target platform is constructed based on the object resources released by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship for indicating the candidate user accounts and the object resources. Then, the target local relationship graph and the global relationship graph are used to determine the recommendation coefficient between each candidate user account and each object resource, wherein the global relationship graph is a relationship graph jointly constructed based on the target local relationship graph and the reference local relationship graph matched by the reference platform other than the target platform. Then, the target object resource to be recommended is determined for each candidate user account according to the recommendation coefficient. In other words, using an embodiment of the present application, the target local relationship graph of the target platform constructed based on the object resource body in the target platform and the candidate user accounts in the cold start pool, and the global relationship graph constructed by combining the target local relationship graph and the reference local relationship graph corresponding to the reference platform other than the target platform, is used to obtain the recommendation coefficient between each candidate user account corresponding to the target platform and each object resource. This makes the resulting recommendation coefficient more reliable, allowing the target resource to be recommended for each candidate user account based on the recommendation coefficient, adapting to the actual needs of the candidate user account. This avoids the problem of low resource recommendation accuracy caused by the relatively simple recommendation methods provided by related technologies, thereby achieving the technical effect of improving the accuracy of resource recommendations.

[0057] Optionally, as an optional solution, based on the object resources released by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, constructing a target local relationship graph matching the target platform includes:

[0058] The candidate user account is used as the account node, and the object resource is used as the resource node;

[0059] Optionally, in this embodiment, after the candidate user account is used as an account node, the process further includes: generating the initial node state features of the candidate user account based on the platform-related features corresponding to the candidate user account and the account personality features of the candidate user account. Specifically, the platform-related features may be, but are not limited to, features that indicate the candidate user account's corresponding features in the target platform. For example, assuming that the target platform is a gaming platform, the platform-related features may be, but are not limited to, features that indicate the gaming level of the candidate user account. The account personality features may be, but are not limited to, features that indicate the preference features corresponding to the candidate user account after it leaves the target platform. Among these, in the initial node state features, the features that correspond to the candidate user account in the target platform are randomly selected, and the account personality features of the candidate user account are obtained based on the historical evaluation information published by the candidate user account in the target platform, or the historical evaluation information published by the associated user accounts associated with the candidate user account in other platforms other than the target platform.

[0060] For example, the above process can refer to the following example:

[0061]

[0062] Among them, the above Used to represent the individual characteristics of the candidate user account. Used to indicate platform-related features. The above G lv It is used to represent the initial node status characteristics of the account node corresponding to the candidate user account. lv [0:m] is used to represent the 0th to mth dimension data of the initial node state characteristics. lv (m:emb_size] is used to represent the m+1th to emb_sizeth dimension data of the above-mentioned initial node state features. In other words, the above-mentioned initial node state features include emb_size-dimensional data, wherein the 0th-mth dimensions are used to represent the account personality characteristics of the candidate user account, and the m+1th to emb_sizeth dimensions are used to represent the platform-related characteristics of the candidate user account.

[0063] Optionally, in this embodiment, after the reference object resource is used as a resource node, the following steps may be performed, but are not limited to: obtaining description information of the reference object resource; and determining, based on the description information of the reference object resource, initial node state features of the resource node corresponding to the reference object resource. The following steps may be performed, but are not limited to, using a Word to Vector (Word2Vec) model to extract features from the description information of the reference object resource to obtain the initial node state features of the resource node corresponding to the reference object resource.

[0064] Set the resource node of each object resource as the current resource node in turn, and perform the following operations:

[0065] The account node of the candidate user account that has performed the evaluation operation on the current object resource corresponding to the current resource node is used as the current associated account node associated with the current resource node;

[0066] Connect the current resource node with the current associated account node;

[0067] It should be noted that, in this embodiment, the edge connecting the current resource node and the currently associated account node also carries an initial edge feature. Specifically, this initial edge feature can be obtained, but is not limited to, based on the following steps: determining the evaluation score of the currently associated account node for the current resource node; and encoding the evaluation score to obtain an edge feature corresponding to the edge connecting the current resource node and the currently associated account node. For example, in this embodiment, the evaluation score can be divided into five levels, one to five, with different levels corresponding to different evaluation score ranges. If the evaluation score is at level one and the corresponding evaluation score range is equal, then the initial edge feature can be [1, 0, 0, 0, 0]; if the evaluation score is at level two and the corresponding evaluation score range is equal, then the initial edge feature can be [0, 1, 0, 0, 0, 0]; if the evaluation score is at level three and the corresponding evaluation score range is equal, then the initial edge feature can be [0, 0, 1, 0, 0, 0], and so on. It should be noted that the above example is merely an optional example in this embodiment. The specific implementation method for determining the initial connection edge features can be flexibly adjusted according to needs and is not limited to this in this embodiment. The above encoding processing method can be, but is not limited to, one-hot encoding (abbreviated as one-hot encoding) or other similar methods, and this embodiment is not limited to this in any way.

[0068] When the current resource node is not the last resource node in the target local relationship graph, obtain the next resource node as the current resource node;

[0069] In a case where the current resource node is the last resource node in the target local relationship graph, it is determined to construct a target local relationship graph that matches the target platform.

[0070] For example, assuming that the candidate accounts include candidate user account 1, candidate user account 2, candidate user account 3, candidate user account 4, candidate user account 5, and candidate user account 6, and the object resources include object resource a, object resource b, object resource c, object resource d, object resource e, and object resource f, wherein candidate user account 1 has performed evaluation operations on the object resources a and object resource f, and candidate user account 2 has performed evaluation operations on the object resource c, the above process can refer to the following example:

[0071] like Figure 3 As shown in (a), account node 1, account node 2, account node 3, account node 4, account node 5, and account node 6 are determined based on candidate user account 1, candidate user account 2, candidate user account 3, candidate user account 4, candidate user account 5, and candidate user account 6 respectively; and resource node a, resource node b, resource node c, resource node d, resource node e, and resource node f are determined based on object resource a, object resource b, object resource c, object resource d, object resource e, and object resource f respectively.

[0072] Then, if Figure 3 As shown in (b), the account node 1 corresponding to the candidate user account 1 is connected to the resource node a and resource node f corresponding to the object resource a and object resource f, respectively, and the account node 2 corresponding to the candidate user account 2 is connected to the resource node c corresponding to the object resource c to obtain the target local relationship graph.

[0073] In an embodiment of the present application, the candidate user account is used as the account node, and the object resource is used as the resource node; the resource node of each object resource is used as the current resource node in turn, and the following operations are performed: the account node of the candidate user account that has performed the evaluation operation on the current object resource corresponding to the current resource node is used as the current associated account node associated with the current resource node. Then, the current resource node and the current associated account node are connected. Furthermore, in the case that the current resource node is not the last resource node in the target local relationship graph, the next resource node is obtained as the current resource node. Thus, in the case that the current resource node is the last resource node in the target local relationship graph, it is determined to construct a target local relationship graph that matches the target platform. In other words, using an embodiment of the present application, the target local relationship graph is constructed by determining the candidate user account and the object resource as nodes and connecting the nodes with the evaluation operation relationship. The relationship between candidate user accounts and object resources can be intuitively and fully reflected in the target local relationship graph, so that each candidate user account and object resource can learn richer content by utilizing the relationship presented in the target local relationship graph, so as to improve the reliability of the recommendation coefficient between each candidate user account and each object resource determined by utilizing the target local relationship graph and the global relationship graph.

[0074] Optionally, as an optional solution, after constructing a target local relationship graph matching the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, the method further includes:

[0075] S1. Obtain the reference local relationship graphs matched by the reference platforms, wherein, in the reference local relationship graphs, the reference object resources released by the reference platform serve as reference resource nodes, and the user accounts that have performed evaluation operations on at least one reference object resource in the reference platform serve as reference account nodes. The reference account nodes are connected to the reference resource nodes that have evaluation association relationships with the reference account nodes, and the evaluation association relationships indicate that the user account corresponding to the reference account nodes has performed evaluation operations on the object resources corresponding to the reference resource nodes.

[0076] It should be noted that in this embodiment, the aforementioned reference platforms may include, but are not limited to, other platforms belonging to the same field as the target platform, and other platforms belonging to different fields than the target platform. This embodiment does not impose any restrictions on this. It should also be noted that the reference platforms may be, but are not limited to, platforms that are already widely used. In this embodiment, it is necessary to construct at least one reference local relationship graph based on at least one reference platform.

[0077] Optionally, in this embodiment, the reference local relationship graph can be constructed based on but not limited to the following steps: taking the reference user account as the reference account node, and taking the reference object resource as the reference resource node; taking the reference resource node of each reference object resource as the current reference resource node in turn, and performing the following operations: taking the reference account node of the reference user account that has performed an evaluation operation on the current reference object resource corresponding to the reference current resource node as the current associated reference account node associated with the current reference resource node; connecting the current reference resource node with the current reference associated account node; when the current reference resource node is not the last resource node in the reference local relationship graph, obtaining the next reference resource node as the current reference resource node; when the current reference resource node is the last resource node in the reference local relationship graph, determining to construct a reference local relationship graph that matches the reference platform.

[0078] Optionally, in this embodiment, after the reference user account is used as the reference account node, the process further includes: generating the initial node state features of the reference user account based on the platform-related features corresponding to the reference user account and the account personality features of the reference user account. Specifically, the platform-related features may be, but are not limited to, features that indicate the reference user account's corresponding features in the target platform. For example, assuming that the target platform is a gaming platform, the platform-related features may be, but are not limited to, features that indicate the gaming level of the reference user account. The account personality features may be, but are not limited to, features that indicate the preference features corresponding to the reference user account after it leaves the target platform. Among these, in the initial node state features, the features that correspond to the reference user account in the target platform are randomly selected, and the account personality features of the reference user account are obtained based on the historical evaluation information published by the reference user account in the target platform, or the historical evaluation information published by associated user accounts associated with the reference user account in platforms other than the target platform.

[0079] For example, the above process can refer to the following example:

[0080]

[0081] Among them, the above Used to represent the account personality characteristics of the reference user account, the above Used to indicate platform-related features. The above G gv Used to represent the initial node status characteristics of the account node corresponding to the reference user account. The above G gv [0:m] is used to represent the 0th to mth dimension data of the initial node state characteristics. gv(m:emb_size] is used to represent the m+1th to emb_sizeth dimension data of the above-mentioned initial node state features. In other words, the above-mentioned initial node state features include emb_size-dimensional data, wherein the 0th-mth dimensions are used to represent the account personality characteristics of the reference user account, and the m+1th to emb_sizeth dimensions are used to represent the platform-related characteristics of the reference user account.

[0082] Optionally, in this embodiment, after the object resource is used as a resource node, the following steps may be included but are not limited to: obtaining description information of the object resource; and determining initial node state characteristics of the resource node corresponding to the object resource based on the object resource description information.

[0083] S2, obtain the union of the reference account node contained in the reference local relationship graph and the account node contained in the target local relationship graph, obtain the global account node in the global relationship graph, and obtain the union of the reference resource node contained in the reference local relationship graph and the resource node contained in the target local relationship graph, obtain the global resource node in the global relationship graph.

[0084] S3, connecting the global account node and the global resource node having an evaluation association relationship with the global account node to construct a global relationship graph, wherein the evaluation association relationship indicates that the user account corresponding to the global account node has performed an evaluation operation on the object resource corresponding to the global resource node.

[0085] For example, suppose Figure 4 The target local relationship diagram shown in (a) includes: account node 1 corresponding to candidate user account 1, account node 2 corresponding to candidate user account 2, and account node 3 corresponding to candidate account 3, object resource node a corresponding to object resource a, object resource node b corresponding to object resource b, and object resource node c corresponding to object resource c; Figure 4 As shown in (a), it is assumed that the reference local relationship graph includes: reference user account node 4 corresponding to reference user account 4, reference user account node 5 corresponding to reference user account 5, and reference user account node 6 corresponding to reference user account 6, reference object resource node d corresponding to reference object resource d, reference object resource node e corresponding to reference object resource e, and reference object resource node f corresponding to reference object resource f. The above steps can refer to the following example:

[0086] Obtain account information for candidate user accounts for each account node in the target local relationship graph, as well as account information for reference user accounts for each reference account node in the reference local relationship graph. Determine that the account information corresponding to candidate user account 1 for account node 1 is the same as the account information corresponding to the reference user account for reference account node 6.

[0087] Obtain resource information of the object resources of each resource node in the target local relationship graph and resource information of the reference object resources of each reference resource node in the reference local relationship graph. Determine that the resource information corresponding to the object resource a of resource node a is the same as the resource information of the reference object resource of reference resource node f.

[0088] Then, if Figure 4 As shown in (b), all nodes in the target local relationship graph and all nodes in the reference local relationship graph are determined as global account nodes in the global relationship graph. Account node 1 is merged with reference account node 6 to obtain a merged account node, and resource node a is merged with reference resource node f to obtain a merged resource node. The global account node and the global resource nodes that have an evaluation association with the global account node are then connected to construct a global relationship graph.

[0089] In an embodiment of the present application, a reference local relationship graph matching each reference platform is obtained, wherein in the reference local relationship graph, the reference object resources published by the reference platform serve as reference resource nodes, and the user accounts that have performed evaluation operations on at least one reference object resource in the reference platform serve as reference account nodes. The reference account nodes are connected to the reference resource nodes that have an evaluation association relationship with the reference account nodes, and the evaluation association relationship indicates that the user account corresponding to the reference account node has performed an evaluation operation on the object resource corresponding to the reference resource node. Then, the union of the reference account nodes contained in the reference local relationship graph and the account nodes contained in the target local relationship graph is obtained to obtain a global account node in the global relationship graph, and the union of the reference resource nodes contained in the reference local relationship graph and the resource nodes contained in the target local relationship graph is obtained to obtain a global resource node in the global relationship graph. Next, the global account node and the global resource nodes that have an evaluation association relationship with the global account node are connected to construct a global relationship graph, wherein the evaluation association relationship indicates that the user account corresponding to the global account node has performed an evaluation operation on the object resource corresponding to the global resource node. In other words, by using the target local relationship graph and the reference local relationship graph to construct a global relationship graph, the embodiments of the present application enrich the content presented in the global local relationship graph, thereby improving the reliability of the recommendation coefficients between each candidate user account and each object resource determined using the target local relationship graph and the global relationship graph. This achieves the technical effect of improving the accuracy of resource recommendations.

[0090] Optionally, as an optional solution, after constructing a target local relationship graph matching the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, the method further includes:

[0091] S1, obtaining historical evaluation information published by each candidate user account in the target local relationship graph on the associated platform, wherein the associated platform uses the identity information of the candidate user account to log in to the platform other than the target platform.

[0092] S2: extracting the individual account characteristics corresponding to each candidate user account from the historical evaluation information.

[0093] It should be noted that the aforementioned associated platform may be, but is not limited to, used by a user using the candidate user account to log in to a platform other than the target platform using their identity information using an associated account associated with the candidate user account. The aforementioned historical evaluation information may be, but is not limited to, indicating resource evaluation information, comments, or likes posted by the associated account on the aforementioned associated platform, and this is not limited in this embodiment.

[0094] Optionally, in this embodiment, extracting the account personality characteristics corresponding to each candidate user account from the historical evaluation information may include, but is not limited to, converting the historical evaluation information into a historical evaluation vector and extracting the account personality characteristics from the historical evaluation vector. This may be accomplished using, but is not limited to, a Word to Vector (Word2Vec) model or other models with similar functionality.

[0095] S3, based on the feature similarity between the individual features of the accounts, similar account pairs are determined from the candidate user accounts.

[0096] For example, the above steps may refer to, but are not limited to, the following examples:

[0097] S u,v =sim(G(U) u , G(U) v ) (3)

[0098] Among them, the above S u,v It is used to represent the feature similarity between the personality features of two candidate user accounts. The above G(U) u 、G(U) v ) are used to represent the account personality characteristics of the two candidate user accounts for similarity comparison.

[0099] S4, connecting the corresponding account nodes of similar account pairs to obtain the updated target local relationship graph.

[0100] It should be noted that in this embodiment, the connection edges connecting the account nodes corresponding to similar accounts carry similar connection edge features. Specifically, these similar connection edge features can be obtained based on, but not limited to, the following steps: encoding the similarity between similar accounts to obtain the similar connection edge features. For example, assuming the similarity between similar accounts is 0.8, the similar connection edge features can be represented by, but not limited to, [0, 0, 0, 0, 0.8].

[0101] As an optional embodiment, it is possible but not limited to adopt the following Figure 5 The following steps are shown to illustrate the above method:

[0102] like Figure 5 As shown in (a), the target local relationship graph includes: account node 1, account node 2, and account node 3.

[0103] Next, the account personality features of the candidate user accounts corresponding to each of the account nodes in account node 1, account node 2, and account node 3 are obtained respectively, and it is determined that the feature similarity between the account personality features of the candidate user accounts corresponding to account node 1 and account node 2 is greater than the similarity threshold, so as to Figure 5 As shown in (b), account node 1 and account node 2 are connected to obtain the updated target local relationship graph.

[0104] Optionally, in this embodiment, after the account personality features corresponding to each candidate user account are extracted from the historical evaluation information, the above method also includes: S1, obtaining the reference historical evaluation information published by each reference user account on the associated platform, wherein the associated platform is a platform other than the target platform logged in using the identity information of the reference user account; S2, referring to the reference account personality features corresponding to each reference user account extracted from the historical evaluation information; S3, based on the feature similarity between the account personality features, determining the account nodes corresponding to the candidate user accounts with similarity association relationships and the account nodes corresponding to the reference user accounts in the global relationship graph; S4, connecting the account nodes corresponding to the candidate user accounts with similarity association relationships and the account nodes corresponding to the reference user accounts to obtain an updated global relationship graph.

[0105] It should be noted that, for a specific example of updating the global relationship graph, please refer to the above-mentioned example of updating the local relationship graph, which will not be further described in this embodiment.

[0106] In an embodiment of the present application, historical evaluation information published by each candidate user account in a target local relationship graph on an associated platform is obtained, where the associated platform uses the identity information of the candidate user account to log in to a platform other than the target platform. Then, the account personality characteristics corresponding to each candidate user account are extracted from the historical evaluation information. Furthermore, based on the feature similarity between the account personality characteristics, similar account pairs are identified from the candidate user accounts. The account nodes corresponding to the similar account pairs are then connected to obtain an updated target local relationship graph. In other words, using an embodiment of the present application, the connection relationships in the target local relationship graph are increased based on the feature similarity between the account personality characteristics of the accounts. This allows the relationships between the candidate user accounts and the object resources to be intuitively and fully reflected in the target local relationship graph. This facilitates each candidate user account and the object resource to learn richer content by leveraging the relationships presented in the target local relationship graph, thereby improving the reliability of the recommendation coefficients between each candidate user account and each object resource determined using the target local relationship graph and the global relationship graph. This achieves the technical effect of improving the accuracy of resource recommendations.

[0107] Optionally, as an optional solution, using the target local relationship graph and the global relationship graph, determining the recommendation coefficient between each candidate user account and each object resource includes:

[0108] S1, based on the point-to-point connection relationships and point-edge connection relationships in the target local relationship graph, and the point-to-point connection relationships and point-edge connection relationships in the global relationship graph, determining node state features, wherein the node state features include account node state features corresponding to each candidate user account, and resource node state features corresponding to each object resource;

[0109] Optionally, in this embodiment, the above-mentioned determination of node status features based on the point-to-point connection relationship and the point-to-edge connection relationship in the target local relationship graph, and the point-to-point connection relationship and the point-to-edge connection relationship in the global relationship graph includes: determining the account node of the key user account with the same account information, and the resource node of the key object resource with the same resource identifier from the target local relationship graph and the global relationship graph; taking the account node of the key user account and the resource node of the key object resource as the current node respectively, and performing the following operations in the feature aggregation network: determining the first current neighbor resource node connected to the current node and the first current neighbor account node connected to the current node from the target local relationship graph, and determining the second current neighbor resource node connected to the current node and the second current neighbor account node connected to the current node from the global relationship graph; obtaining the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current node, and the first connection edge aggregation result and the second connection edge aggregation result corresponding to the current node, wherein, The first neighbor node aggregation result is used to indicate the result of feature aggregation processing on the first current neighbor resource node and the first current neighbor account node, the second neighbor node aggregation result is used to indicate the result of feature aggregation processing on the second current neighbor resource node and the second current neighbor account node, the first connection edge aggregation result is used to indicate the result of feature aggregation processing on the connection edge between the current node and the first current neighbor resource node, and the connection edge between the current node and the first current neighbor account node in the target local relationship graph, the second connection edge aggregation result is used to indicate the result of feature aggregation processing on the connection edge between the current node and the second current neighbor resource node, and the connection edge between the current node and the second current neighbor account node in the global relationship graph; the first neighbor node aggregation result, the second neighbor node aggregation result, the first connection edge aggregation result and the second connection edge aggregation result are used to determine the current node state feature corresponding to the current node; wherein, the feature aggregation network is obtained by joint training with the resource recommendation network.

[0110] S2, inputting the node state features into the resource recommendation network to obtain the recommendation coefficient of the candidate user account for the above-mentioned object resource, wherein the resource recommendation network is trained using the reference local relationship graph and the target local relationship graph.

[0111] Optionally, in this embodiment, the above-mentioned inputting of node status characteristics into the resource recommendation network to obtain the recommendation coefficient of the candidate user account for the above-mentioned object resource may include, but is not limited to: inputting the node status characteristics corresponding to the account node of the candidate user account and the node status characteristics corresponding to the resource node of the object resource into the resource recommendation network to obtain the recommendation coefficient of the candidate user account for the above-mentioned object resource.

[0112] Furthermore, the node status features corresponding to the account node of the candidate user account and the node status features corresponding to the resource node of the target resource are input into the resource recommendation network to obtain the recommendation coefficient of the candidate user account for the target resource, which may include but is not limited to:

[0113] S1: Input the node status features corresponding to the account node of the candidate user account and the node status features corresponding to the resource node of the target resource into the resource recommendation network.

[0114] S2, in the first layer of the resource recommendation network included in the resource recommendation network, performs the following operations:

[0115] The node status features corresponding to the account node and the node status features corresponding to the resource node are concatenated to obtain a connection result. Next, the connection result is multiplied by the first resource recommendation network parameter to obtain a first multiplication result. Furthermore, the first multiplication result and the second resource recommendation network parameter are weighted and summed to obtain a summed result. The summed result is then subjected to a first target rule calculation to obtain the output of the first-layer resource recommendation network. The first target rule calculation may include, but is not limited to, squaring the summed result, taking the absolute value of the summed result, or taking the logarithm of the summed result.

[0116] S3: Input the output result of the first-layer resource recommendation network into the second-layer resource recommendation network included in the resource recommendation network.

[0117] S4, performing the following operations in the second-layer resource recommendation network included in the resource recommendation network:

[0118] The third resource recommendation network parameter is multiplied by the output of the first-layer resource recommendation network to obtain a second multiplication result. Then, a weighted sum calculation is performed on the second multiplication result and the fourth resource recommendation network parameter to obtain a recommendation coefficient for the candidate user account for the target resource.

[0119] For example, the above steps may be implemented with reference to, but not limited to, the following examples:

[0120]

[0121]

[0122] Among them, formula 4 can be used but not limited to represent the first-layer resource recommendation network, the above formula 5 can be used but not limited to represent the second-layer resource recommendation network, and the above formula It can be used, but is not limited to, to represent the output results of the first-layer resource recommendation network. The above G(U) v , G(I) kIt can be but not limited to the node status characteristics corresponding to the account node of the candidate user account and the node status characteristics corresponding to the resource node of the object resource. n1 It can be used, but not limited to, to represent the first resource recommendation network parameters, the above b n1 It can be used, but not limited to, to represent the second resource recommendation network parameters, the above It can be used, but is not limited to, to represent the output result of the second-layer resource recommendation network (i.e., the above-mentioned recommendation coefficient). n2 It can be used, but not limited to, to indicate the third resource recommended network parameter. n2 It can be used, but is not limited to, to represent the fourth resource recommendation network parameter. The above relu can be used, but is not limited to, to represent the first target rule calculation.

[0123] In an embodiment of the present application, node state features are determined based on the point-to-point connection relationship and point-edge connection relationship in the target local relationship graph, as well as the point-to-point connection relationship and point-edge connection relationship in the global relationship graph, wherein the node state features include the account node state features corresponding to each candidate user account, and the resource node state features corresponding to each object resource. Then, the node state features are input into the resource recommendation network to obtain the recommendation coefficient of the candidate user account for the above-mentioned object resource, wherein the resource recommendation network is trained using the reference local relationship graph and the target local relationship graph. In other words, using an embodiment of the present application, by using the target local relationship graph and the global relationship graph to jointly obtain the node state features of the account node corresponding to the candidate user account, and the node state features of the resource node corresponding to the object resource, the content learned by the account node and resource node corresponding to the target platform not only depends on the data in the reference platform, but is also affected by the data in the target platform. This ensures that the content learned by the above-mentioned account node and resource node will not be excessively affected by the data in the reference platform, but will retain the characteristics of the target platform while learning the data in the reference platform. This makes the node status characteristics of the account nodes corresponding to the candidate user accounts and the node status characteristics of the resource nodes corresponding to the target resources more accurate, thereby improving the accuracy of the recommendation coefficients obtained based on the node status characteristics of the account nodes corresponding to the candidate user accounts and the node status characteristics of the resource nodes corresponding to the target resources. This achieves the technical effect of improving the accuracy of resource recommendations.

[0124] Optionally, as an optional solution, based on the point-to-point connection relationship and the point-to-edge connection relationship in the target local relationship graph, and the point-to-point connection relationship and the point-to-edge connection relationship in the global relationship graph, determining the node state feature includes:

[0125] Determine, from the target local relationship graph and the global relationship graph, account nodes of key user accounts having identical account information and resource nodes of key object resources having identical resource identifiers;

[0126] It should be noted that the above-mentioned key user account nodes with the same account information may include, but are not limited to, nodes corresponding to candidate user accounts in the target local relationship graph and nodes corresponding to candidate user accounts in the global relationship graph. Correspondingly, the above-mentioned key object resource resource nodes with the same resource identifier may include, but are not limited to, nodes corresponding to object resources in the target local relationship graph and nodes corresponding to object resources in the global relationship graph.

[0127] Set the account node of the key user account and the resource node of the key object resource as the current node, and perform the following operations in the feature aggregation network:

[0128] Determine a first current neighbor resource node connected to the current node and a first current neighbor account node connected to the current node from the target local relationship graph, and determine a second current neighbor resource node connected to the current node and a second current neighbor account node connected to the current node from the global relationship graph;

[0129] Obtain the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current node, and the first connection edge aggregation result and the second connection edge aggregation result corresponding to the current node, wherein the first neighbor node aggregation result is used to indicate the result obtained by performing feature aggregation processing on the first current neighbor resource node and the first current neighbor account node, and the second neighbor node aggregation result is used to indicate the result obtained by performing feature aggregation processing on the second current neighbor resource node and the second current neighbor account node. The first connection edge aggregation result is used to indicate the result obtained by performing feature aggregation processing on the connection edge between the current node and the first current neighbor resource node, and the connection edge between the current node and the first current neighbor account node in the target local relationship graph. The second connection edge aggregation result is used to indicate the result obtained by performing feature aggregation processing on the connection edge between the current node and the second current neighbor resource node, and the connection edge between the current node and the second current neighbor account node in the global relationship graph;

[0130] Determine a current node state feature corresponding to the current node using the first neighbor node aggregation result, the second neighbor node aggregation result, the first connection edge aggregation result, and the second connection edge aggregation result;

[0131] Among them, the feature aggregation network is jointly trained with the resource recommendation network.

[0132] Optionally, in this embodiment, obtaining the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current node, and the first connection edge aggregation result and the second connection edge aggregation result corresponding to the current node includes:

[0133] S1, aggregating node status features of a first current neighbor resource node and node status features of a first current neighbor account node to obtain a first neighbor node aggregation result, wherein the node status features of the first current neighbor account node include account personality features of a neighbor account corresponding to the first current neighbor account node, and the node status features corresponding to the first current neighbor resource node are generated based on resource description information of a neighbor object resource corresponding to the first current neighbor resource node;

[0134] S2, aggregating the node status characteristics of the second current neighbor resource node and the node status characteristics of the second current neighbor account node to obtain a second neighbor node aggregation result, wherein the node status characteristics of the second current neighbor account node include the account personality characteristics of the neighbor account corresponding to the second current neighbor account node, and the node status characteristics corresponding to the second current neighbor resource node are generated based on resource description information of the neighbor object resource corresponding to the second current neighbor resource node;

[0135] S3. Classify and aggregate the first side state feature of the connection edge between the current node and the first current neighbor resource node, and the second side state feature of the connection edge between the current node and the first current neighbor account node to obtain a first connection edge aggregation result, wherein, when the current node is an account node of a key user account, the first side state feature is obtained based on the evaluation information of the key user account corresponding to the current node on the object resource corresponding to the first current neighbor resource node, and the second side state feature is obtained based on the feature similarity between the account personality feature of the key user account corresponding to the current node and the account personality feature of the user account corresponding to the first current neighbor account node; when the current node is a resource node of a key object resource, the first side state feature is obtained based on the similarity between the object resource corresponding to the current node and the object resource corresponding to the first current neighbor resource node, and the second side state feature is obtained based on the evaluation information of the key user account corresponding to the first current neighbor account node on the object resource corresponding to the current node;

[0136] S4. Classify and aggregate the third side state feature of the connection edge between the current node and the second current neighbor resource node, and the fourth side state feature of the connection edge between the current node and the second current neighbor account node to obtain a second connection edge aggregation result, wherein, when the current node is an account node of a key user account, the third side state feature is obtained based on the evaluation information of the key user account corresponding to the current node on the object resource corresponding to the second current neighbor resource node, and the fourth side state feature is obtained based on the feature similarity between the account personality feature of the key user account corresponding to the current node and the account personality feature of the user account corresponding to the second current neighbor account node; when the current node is a resource node of a key object resource, the third side state feature is obtained based on the similarity between the object resource corresponding to the current node and the object resource corresponding to the second current neighbor resource node, and the fourth side state feature is obtained based on the evaluation information of the key user account corresponding to the second current neighbor account node on the object resource corresponding to the current node.

[0137] Furthermore, the above-mentioned determination of the current node state feature corresponding to the current node using the first neighbor node aggregation result, the second neighbor node aggregation result, the first connection edge aggregation result, and the second connection edge aggregation result may include, but is not limited to: aggregating the first neighbor node aggregation result and the first connection edge aggregation result to obtain a target point-edge aggregation result that matches the target local relationship graph. Then, aggregating the second neighbor node aggregation result and the second connection edge aggregation result to obtain a global point-edge aggregation result that matches the global relationship graph. Next, fusing the target point-edge aggregation result with the global point-edge aggregation result to obtain the current node state feature.

[0138] In an embodiment of the present application, the neighbor nodes and neighbor edges of each node in the local relationship graph are aggregated in sequence, and the neighbor nodes and neighbor edges of each node in the global relationship graph are aggregated, and then the aggregation results obtained by the above two aggregation operations are fused again to obtain the node state characteristics of each node. This allows each node to fully learn the corresponding data in the local relationship graph and the corresponding data in the global relationship graph, thereby achieving the purpose of improving the richness and accuracy of the node state characteristics of each node. This achieves the technical effect of improving the accuracy of resource recommendations.

[0139] Optionally, as an optional solution, obtaining the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current node, and the first connection edge aggregation result and the second connection edge aggregation result corresponding to the current node includes:

[0140] S1. Aggregate the node status characteristics of the first current neighbor resource node and the node status characteristics of the first current neighbor account node to obtain the first neighbor node aggregation result, wherein the node status characteristics of the first current neighbor account node include the account personality characteristics of the neighbor account corresponding to the first current neighbor account node, and the node status characteristics corresponding to the first current neighbor resource node are generated based on the resource description information of the neighbor object resource corresponding to the first current neighbor resource node.

[0141] It should be noted that, in this embodiment, the above-mentioned current node can be an account node in a local relationship graph, or a resource node in a local relationship graph, or an account node in a global relationship graph, or a resource node in a global relationship graph.

[0142] Specifically, the first current neighbor resource node and the first current neighbor account node are the neighbor resource nodes and neighbor account nodes of the current node in the local relationship graph when the current node is an account node or a resource node in the local relationship graph. The second current neighbor resource node and the second current neighbor account node are the neighbor resource nodes and neighbor account nodes of the current node in the local relationship graph when the current node is an account node or a resource node in the global relationship graph.

[0143] Optionally, in this embodiment, the node status feature of the first current neighbor resource node may be, but is not limited to, used to indicate the initial node status feature of the first current neighbor resource node. The node status feature of the first current neighbor account node may be, but is not limited to, used to indicate the initial node status feature of the first current neighbor account node.

[0144] Furthermore, for example, the above steps may be implemented with reference to, but not limited to, the following examples:

[0145] N v =[G vnl ; G vn2 ;...;G vnq ] (6)

[0146]

[0147] T v =N v a v (8)

[0148] Among them, the above N v ∈R d×q It can be used, but not limited to, to represent the initial node state feature aggregation result of the first current neighbor resource node and the first current account resource node. The above G vnl ; G vn2 ;...;G vnqIt may include, but is not limited to, the node status characteristics of the first current neighbor resource node and the initial node status characteristics of the first current neighbor account node, the above q is the neighbor number of node v, the above a v Can be used, but not limited to, to represent the aggregation weight matrix, W1∈R d , W2∈R d×d It can be used, but is not limited to, to represent the weight matrix in the feature aggregation network, that is, the first aggregation parameter and the second aggregation parameter included in the feature aggregation network. The above T v It can be used, but is not limited to, to indicate the aggregation result of the first neighbor node.

[0149] S2. Aggregate the node status characteristics of the second current neighbor resource node and the node status characteristics of the second current neighbor account node to obtain the second neighbor node aggregation result, wherein the node status characteristics of the second current neighbor account node include the account personality characteristics of the neighbor account corresponding to the second current neighbor account node, and the node status characteristics corresponding to the second current neighbor resource node are generated based on the resource description information of the neighbor object resource corresponding to the second current neighbor resource node.

[0150] It should be noted that the above-mentioned aggregation processing of the node status characteristics of the second current neighbor resource node and the node status characteristics of the second current neighbor account node to obtain the second neighbor node aggregation result can be referred to the relevant example of obtaining the first neighbor node aggregation result above. This will not be repeated in this embodiment.

[0151] S3. Classify and aggregate the first side state feature of the connection edge between the current node and the first current neighbor resource node, and the second side state feature of the connection edge between the current node and the first current neighbor account node to obtain a first connection edge aggregation result, wherein, when the current node is an account node of a key user account, the first side state feature is obtained based on the evaluation information of the key user account corresponding to the current node on the object resource corresponding to the first current neighbor resource node, and the second side state feature is obtained based on the feature similarity between the account personality feature of the key user account corresponding to the current node and the account personality feature of the user account corresponding to the first current neighbor account node; when the current node is a resource node of a key object resource, the first side state feature is obtained based on the similarity between the object resource corresponding to the current node and the object resource corresponding to the first current neighbor resource node, and the second side state feature is obtained based on the evaluation information of the key user account corresponding to the first current neighbor account node on the object resource corresponding to the current node.

[0152] Optionally, in this embodiment, the connection edge between the current node and the first current neighbor resource node is, when the current node is an account node or a resource node in the local relationship graph, the connection edge between the neighbor resource nodes connected to the current node in the local relationship graph. The connection edge between the current node and the first current neighbor account node is, when the current node is an account node or a resource node in the local relationship graph, the connection edge between the neighbor account nodes connected to the current node in the local relationship graph.

[0153] Furthermore, for example, the above steps may be implemented with reference to, but not limited to, the following examples:

[0154] P v，r =mean(E vl，r , E v2，r ,...,E vb，r ) (9)

[0155] P v =[P v，1 ;P v，2 ;...P v，s ] (10)

[0156]

[0157] H v =P v β v (12)

[0158] Among them, the above P v，r It can be used, but is not limited to, to represent the initial aggregation results corresponding to the same type of connection edges. vl，r , E v2，r ,...,E vb，r ∈R e It can be used, but not limited to, to represent the edge features of all connected edges of the same type. The above P v，1 ;P v，2 ;...P v，s ∈R e It can be used, but is not limited to, to represent all initial aggregation results corresponding to all types of connected edges. The above P v It can be used, but is not limited to, to represent the result obtained by aggregating all the initial aggregation results corresponding to all types of connected edges again. The above β v It can be used, but is not limited to, to represent the weight matrix corresponding to all initial aggregation results of all types of connection edges. The above W3∈R e , W4∈R eThey are used to represent the weight matrix in the feature aggregation network, that is, the third feature aggregation network parameter and the fourth feature aggregation network parameter in the feature aggregation network. The above H v Used to represent the first connection edge aggregation result. Specifically, in the case where the above-mentioned current node is a user node in the local relationship graph, the type of the above-mentioned connection edge may include, but is not limited to: connection edges corresponding to different evaluation score levels, and connection edges corresponding to different account personality feature similarities. In the case where the above-mentioned current node is a resource node in the local relationship graph, the type of the above-mentioned connection edge may include, but is not limited to: connection edges corresponding to different evaluation score levels, and connection edges corresponding to different resource similarities, wherein the above-mentioned resource similarity is used to indicate the similarity between the description information of the object resources corresponding to different resource nodes.

[0159] Furthermore, based on the above examples, it can be seen that in this embodiment, during the aggregation process, the following Figure 6 The attention mechanism shown in the following example is Figure 6 In the first attention layer shown in (a), referring to the node attention stage, the neighbor nodes of the current node are aggregated to obtain the aggregation results of the neighbor nodes (e.g., Figure 6 Feature a in the reference edge attention stage, the neighbor edges of the current node are aggregated to obtain the aggregated results of the neighbor edges. Figure 6 In the second attention layer shown in (b), the aggregation results of neighbor nodes (such as Figure 6 The feature a) in and the aggregation results of neighbor edges (such as, Figure 6 The feature b) in the ,is aggregated to obtain the node embedding of the current node (such as, Figure 6 Feature C in the edge attention stage. In the edge attention stage, a dual-process attention mechanism is also used. Specifically, different types of edges are first aggregated to obtain the aggregation results corresponding to the edges of different types (e.g., Figure 6 The features b1…bn in the image are then aggregated again to obtain the total aggregation results corresponding to the neighbor edges (e.g., Figure 6 Feature b).

[0160] S4. Classify and aggregate the third side state feature of the connection edge between the current node and the second current neighbor resource node, and the fourth side state feature of the connection edge between the current node and the second current neighbor account node to obtain a second connection edge aggregation result, wherein, when the current node is an account node of a key user account, the third side state feature is obtained based on the evaluation information of the key user account corresponding to the current node on the object resource corresponding to the second current neighbor resource node, and the fourth side state feature is obtained based on the feature similarity between the account personality feature of the key user account corresponding to the current node and the account personality feature of the user account corresponding to the second current neighbor account node; when the current node is a resource node of a key object resource, the third side state feature is obtained based on the similarity between the object resource corresponding to the current node and the object resource corresponding to the second current neighbor resource node, and the fourth side state feature is obtained based on the evaluation information of the key user account corresponding to the second current neighbor account node on the object resource corresponding to the current node.

[0161] It should be noted that the above classification and aggregation processing of the third edge state feature of the connection edge between the current node and the second current neighbor resource node, and the fourth edge state feature of the connection edge between the current node and the second current neighbor account node, is performed. For an example of obtaining the second connection edge aggregation result, please refer to the relevant example of obtaining the first connection edge aggregation result above. This will not be repeated in this embodiment.

[0162] Optionally, in this embodiment, the connection edge between the current node and the second current neighbor resource node is, when the current node is an account node or a resource node in the local relationship graph, the connection edge between the neighbor resource nodes connected to the current node in the local relationship graph. The connection edge between the current node and the second current neighbor account node is, when the current node is an account node or a resource node in the local relationship graph, the connection edge between the neighbor account nodes connected to the current node in the local relationship graph.

[0163] In an embodiment of the present application, the neighbor nodes and neighbor edges of each node in the local relationship graph are aggregated in sequence, and the neighbor nodes and neighbor edges of each node in the global relationship graph are aggregated, and then the aggregation results obtained by the above two aggregation operations are fused again to obtain the node state characteristics of each node. This allows each node to fully learn the corresponding data in the local relationship graph and the corresponding data in the global relationship graph, thereby achieving the purpose of improving the richness and accuracy of the node state characteristics of each node. This achieves the technical effect of improving the accuracy of resource recommendations.

[0164] Optionally, as an optional solution, using the first neighbor node aggregation result, the second neighbor node aggregation result, the first connection edge aggregation result, and the second connection edge aggregation result to determine the current node state feature corresponding to the current node includes:

[0165] S1, aggregate the first neighbor node aggregation result and the first connection edge aggregation result to obtain a target point edge aggregation result that matches the target local relationship graph;

[0166] Optionally, in this embodiment, aggregating the first neighbor node aggregation result and the first connection edge aggregation result to obtain a target point-edge aggregation result that matches the target local relationship graph may include, but is not limited to, the following steps:

[0167] Step S1-1, normalizing the fifth feature aggregation network parameter in the feature aggregation network to obtain a normalized result, and performing element-by-element multiplication on the sixth feature aggregation network parameter in the feature aggregation network and the first connection edge aggregation result to obtain a multiplication result;

[0168] Step S1-2, obtaining a first intermediate result based on the normalization result and the multiplication result;

[0169] Step S1-3, obtaining a second intermediate result based on the normalization result and the first neighbor node aggregation result;

[0170] Step S1-4: performing a weighted sum operation on the first intermediate result and the second intermediate result to obtain a target point-edge aggregation result that matches the target local relationship graph.

[0171] For example, the above steps may be implemented with reference to, but not limited to, the following examples:

[0172]

[0173] Among them, G gv Can be used but not limited to represent the target point edge aggregation result, W∈R d×e It can be used, but is not limited to, to represent the weight matrix of the aggregation network, which is also the fifth characteristic aggregation network parameter, M∈R d×e To represent but not limited to the transformation matrix, that is, the sixth feature aggregation network parameter, T v It can be used, but is not limited to, to represent the aggregation result of the first neighbor node. v It can be used, but is not limited to, to represent the first connection edge aggregation result.

[0174] S2, aggregate the second neighbor node aggregation result and the second connection edge aggregation result to obtain a global point-edge aggregation result that matches the global relationship graph;

[0175] It should be noted that for the specific example of obtaining the global point-edge aggregation result matched with the global relationship graph, please refer to the above example of obtaining the target point-edge aggregation result matched with the local relationship graph, which will not be described in detail in this embodiment.

[0176] S3, fuse the target point-edge aggregation results and the global point-edge aggregation results to obtain the current node state features.

[0177] Optionally, in this embodiment, the above-mentioned fusion of the target point-edge aggregation result and the global point-edge aggregation result to obtain the current node state feature may include, but is not limited to, the following steps:

[0178] Step S3-1, normalizing the seventh feature aggregation network parameter in the feature aggregation network to obtain a first normalization result;

[0179] Step S3-2, performing element-by-element multiplication calculation based on the above normalization result and the target point-edge aggregation result to obtain a third intermediate result;

[0180] Step S3-3: obtaining a fourth intermediate result based on the first normalization result and the global point-edge aggregation result;

[0181] Step S3-4: Perform a weighted sum operation on the third intermediate result and the fourth intermediate result to obtain the current node state feature.

[0182] For example, the above steps may be implemented with reference to, but not limited to, the following examples:

[0183]

[0184] Among them, G v It can be used, but not limited to, to represent the current node state characteristics. n ∈R d It can be used, but is not limited to, to represent the weight matrix of the aggregation network, that is, the seventh characteristic aggregation network parameter. The above G lv It can be used, but not limited to, to represent the target point edge aggregation result. The above G gv It can be used, but is not limited to, to represent global point-edge aggregation results.

[0185] In an embodiment of the present application, the neighbor nodes and neighbor edges of each node in the local relationship graph are aggregated in sequence, and the neighbor nodes and neighbor edges of each node in the global relationship graph are aggregated, and then the aggregation results obtained by the above two aggregation operations are fused again to obtain the node state characteristics of each node. This allows each node to fully learn the corresponding data in the local relationship graph and the corresponding data in the global relationship graph, thereby achieving the purpose of improving the richness and accuracy of the node state characteristics of each node. This achieves the technical effect of improving the accuracy of resource recommendations.

[0186] Optionally, as an optional solution, before constructing a target local relationship graph matching the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, the method further includes:

[0187] Determine a first sample account node and a first sample resource node from the sample local relationship graph, and determine a second sample account node and a second sample resource node from the sample global relationship graph;

[0188] Optionally, in this embodiment, the first sample account node and the first sample resource node are the nodes corresponding to the candidate user account and the object resource in the target platform, respectively. The second sample account node includes the node corresponding to the reference user account in the reference platform and the node corresponding to the candidate user account in the target platform, and the second sample resource node includes the node corresponding to the reference object resource in the reference platform and the node corresponding to the object resource in the target platform. In other words, the sample local relationship graph is constructed based on the data in the target platform, and the sample global relationship graph is constructed based on the data in both the target platform and the reference platform.

[0189] Furthermore, before determining the first sample account node and the first sample resource node from the sample local relationship graph and determining the second sample account node and the second sample resource node from the sample global relationship graph, the method further includes:

[0190] The first sample user account (i.e., the candidate user account in the target platform) is used as the account node, and the first sample object resource (i.e., the object resource in the target platform) is used as the resource node;

[0191] S1. Take the sample resource node of each first sample object resource as the current sample resource node in turn, and perform the following operations: take the account node of the first sample user account that has performed the evaluation operation on the current sample object resource corresponding to the current sample resource node as the current associated sample account node associated with the current sample resource node; connect the current sample resource node with the current associated sample account node; when the current sample resource node is not the last sample resource node in the target local relationship graph, obtain the next sample resource node as the current sample resource node; when the current sample resource node is the last sample resource node in the sample local relationship graph, determine to construct a sample local relationship graph that matches the target platform.

[0192] S2. Obtain the reference sample local relationship graphs matched by each reference platform, wherein, in the reference sample local relationship graph, the reference object resources published by the reference platform serve as reference resource nodes, and the user accounts that have performed evaluation operations on at least one reference object resource in the reference platform serve as reference account nodes. The reference account nodes are connected to the reference resource nodes having an evaluation association relationship with the reference account nodes, and the evaluation association relationship indicates that the user account corresponding to the reference account node has performed an evaluation operation on the object resource corresponding to the reference resource node; obtain the union of the reference sample account nodes contained in the reference sample local relationship graph and the sample account nodes contained in the sample local relationship graph to obtain the sample global account node in the sample global relationship graph, and obtain the union of the reference sample resource nodes contained in the reference sample local relationship graph and the sample resource nodes contained in the sample local relationship graph to obtain the sample global resource node in the sample global relationship graph; connect the sample global account nodes and the sample global resource nodes having an evaluation association relationship with the sample global account nodes to construct a sample global relationship graph, wherein the evaluation association relationship indicates that the sample user account corresponding to the sample global account node has performed an evaluation operation on the sample object resource corresponding to the sample global resource node.

[0193] In the feature aggregation network being trained, do the following:

[0194] S1, determining a sample key account node having the same account information from the first sample account node and the second sample account node, and determining a sample resource node of a sample key object resource having the same resource identifier from the first sample resource node and the second sample resource node.

[0195] It should be noted that the aforementioned sample key user account nodes with the same account information may include, but are not limited to, a first sample account node corresponding to a first sample user account (i.e., a candidate user account) in the sample local relationship graph, and a second sample account node corresponding to a second sample user account with the same account information as the first sample account in the global relationship graph. In other words, the aforementioned key user account nodes may be, but are not limited to, account nodes corresponding to the same user account in both the sample local relationship graph and the sample global relationship graph.

[0196] Accordingly, the sample resource nodes of the sample key object resources having the same resource identifier may include, but are not limited to, a first sample resource node corresponding to a first sample object resource (i.e., an object resource in the target platform) in the sample local relationship graph, and a second sample resource node corresponding to a second sample object resource having the same resource identifier as the first sample object resource in the global relationship graph. In other words, the key resource nodes may be, but are not limited to, resource nodes corresponding to the same object resource in both the sample local relationship graph and the sample global relationship graph.

[0197] S2, traverse the sample key account nodes and sample resource nodes, and determine the sample node state characteristics obtained after aggregation processing in the sample local relationship graph and the sample global relationship graph.

[0198] It should be noted that, in this embodiment, the above-mentioned sample node status characteristics may include, but are not limited to: sample node status characteristics of sample resource nodes corresponding to sample object resources, and sample node status characteristics of sample account nodes corresponding to sample user accounts.

[0199] Optionally, in this embodiment, the above-mentioned traversal of the sample key account nodes and the sample resource nodes to determine the sample node state features obtained after aggregation processing in the sample local relationship graph and the sample global relationship graph may include, but is not limited to: taking the account node of the key user account and the resource node of the key object resource as the current node respectively, and performing the following operations in the feature aggregation network:

[0200] Step S2-1, determining a first current sample neighbor resource node connected to the current sample node and a first current sample neighbor account node connected to the current sample node from the sample local relationship graph, and determining a second current sample neighbor resource node connected to the current sample node and a second current sample neighbor account node connected to the current sample node from the sample global relationship graph;

[0201] Step S2-2, obtaining the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current sample node, and the first sample connection edge aggregation result and the second sample connection edge aggregation result corresponding to the current sample node, wherein the first neighbor node aggregation result is used to indicate the result obtained by performing feature aggregation processing on the first current sample neighbor resource node and the first current sample neighbor account node, and the second sample neighbor node aggregation result is used to indicate the result obtained by performing feature aggregation processing on the second current sample neighbor resource node and the second current sample neighbor account node. The first sample connection edge aggregation result is used to indicate the result obtained by performing feature aggregation processing on the sample connection edge between the current sample node and the first current sample neighbor resource node, and the sample connection edge between the current sample node and the first current sample neighbor account node in the sample local relationship graph. The second sample connection edge aggregation result is used to indicate the result obtained by performing feature aggregation processing on the sample connection edge between the current sample node and the second current sample neighbor resource node, and the sample connection edge between the current sample node and the second current sample neighbor account node in the sample global relationship graph;

[0202] Step S2-3: using the first sample neighbor node aggregation result, the second sample neighbor node aggregation result, the first sample connection edge aggregation result and the second sample connection edge aggregation result, determine the current sample node state feature corresponding to the current sample node.

[0203] For specific examples of the above steps, please refer to the relevant examples in the process of generating node status features during the application of this method, which will not be described in detail in this embodiment.

[0204] S3, obtain the first account personality feature obtained during the aggregation processing of the sample key account node in the sample local relationship graph, and the second account personality feature obtained during the aggregation processing of the sample key account node in the sample global relationship graph, and use the first account personality feature and the second account personality feature to calculate the aggregation training loss of the feature aggregation network.

[0205] It should be noted that, in this embodiment, the above-mentioned acquisition of the first account personality feature obtained in the process of aggregation processing of the sample key account node in the sample local relationship graph, and the second account personality feature obtained in the process of aggregation processing of the sample key account node in the sample global relationship graph may include, but is not limited to: obtaining the first sample point edge aggregation result and the second sample point edge aggregation result corresponding to the sample key account node, wherein the first sample point edge aggregation result is obtained by aggregating the first sample neighbor node aggregation result and the first sample connection edge aggregation result corresponding to the sample key account node in the sample local relationship graph, the first sample neighbor node aggregation result is the result of feature aggregation processing of the first sample neighbor resource node and the first sample neighbor account node corresponding to the sample key account in the sample local relationship graph, and the first sample connection edge aggregation result is the feature aggregation processing of the connection edge between the sample key account and the first sample neighbor resource node, and the connection edge between the sample key account and the first sample neighbor account node in the sample local relationship graph. The result obtained is that the second sample point edge aggregation result is obtained by aggregating the second sample neighbor node aggregation result and the second sample connection edge aggregation result corresponding to the sample key account node in the sample global relationship graph. The second sample neighbor node aggregation result is the result obtained by feature aggregation processing of the second sample neighbor resource node and the second sample neighbor account node corresponding to the sample key account in the sample global relationship graph. The second sample connection edge aggregation result is the result obtained by feature aggregation processing of the connection edge between the sample key account and the second sample neighbor resource node, and the connection edge between the sample key account and the second sample neighbor account node in the sample local relationship graph; the first sample point edge aggregation result and the second sample point edge aggregation result are fused to determine the sample node state feature corresponding to the sample key account node; the first account personality feature of the sample key account node in the sample local relationship graph is obtained from the first sample point edge aggregation result, and the second account personality feature of the sample key account node in the sample global relationship graph is obtained from the second sample point edge aggregation result.

[0206] Optionally, in this embodiment, the aggregate training loss of the feature aggregation network calculated using the first account personality feature and the second account personality feature may include, but is not limited to:

[0207] Step S3-1, obtain the first feature similarity between the first account personality feature and the second account personality feature, and obtain the second feature similarity between the negative sample account personality features of the first account personality feature, wherein the above-mentioned negative sample account personality feature can be but is not limited to being used to indicate the account personality features of any other sample account node in the above-mentioned sample local relationship graph or sample global relationship graph except the sample account nodes corresponding to the above-mentioned first account personality feature and the second account personality feature.

[0208] Step S3-2: Obtain a node set corresponding to the key account node, and use the node set and the first feature similarity and the second feature similarity to obtain the aggregate training loss of the feature aggregation network.

[0209] For example, the specific implementation of the above steps can be, but is not limited to, referring to the following examples:

[0210]

[0211] Among them, the above l c Used to represent the aggregate training loss, the above Used to indicate the personality characteristics of the first account. Used to indicate the personality characteristics of the second account. represents the negative sample account personality characteristics, μ represents the node set corresponding to the key account node, and τ represents a hyperparameter that can be flexibly adjusted. In this example, using Formula 15, the distance between the first account personality characteristics and the negative sample account personality characteristics is widened, while the distance between the first account personality characteristics and the second account personality characteristics is narrowed, ensuring that the account personality characteristics corresponding to the same key account node remain unchanged in both the local graph and the global graph.

[0212] S4, input the sample node state features into the resource recommendation network in training, and obtain the recommendation training loss of the resource recommendation network.

[0213] Step S4-1, obtaining a resource recommendation evaluation coefficient output by a resource recommendation network in training that matches the state characteristics of a sample node;

[0214] Step S4-2, comparing the resource recommendation evaluation coefficient and the recommendation information indicated by the node reference tag corresponding to the sample node state feature to obtain a recommendation difference result;

[0215] Step S4-3: Calculate the recommendation training loss based on the recommendation difference result.

[0216] S5, perform weighted sum calculation on the aggregation training loss and the recommendation training loss to obtain the joint training loss.

[0217] For example, the specific implementation of the above steps can be, but is not limited to, referenced to the following examples:

[0218] l=l r +λ*l c (16)

[0219] Among them, the above l r It can be used, but is not limited to, to represent the above-mentioned recommendation training loss. cIt can be used, but not limited to, to represent the above-mentioned aggregate training loss, the above-mentioned l can be used, but not limited to, to represent the above-mentioned joint training loss, and the above-mentioned λ can be used, but not limited to, to represent the weight corresponding to the aggregate training loss.

[0220] S6, when the joint training loss does not reach the threshold condition, adjust the aggregation network parameters in the feature aggregation network and the recommendation network parameters in the resource recommendation network.

[0221] It should be noted that the aforementioned aggregated network parameters may include, but are not limited to, all network parameters included in the feature aggregation network, such as the aforementioned first aggregation parameter, second aggregation parameter, third aggregation parameter, fourth aggregation parameter, fifth aggregation parameter, sixth aggregation parameter, and seventh aggregation parameter. The aforementioned recommended network parameters may include, but are not limited to, all network parameters included in the resource recommendation network, such as the aforementioned first resource recommendation network parameter, second resource recommendation network parameter, third resource recommendation network parameter, and fourth resource recommendation network parameter.

[0222] S7, when the joint training loss reaches a threshold condition, determining that the feature aggregation network and the resource recommendation network meet the network convergence condition.

[0223] In the embodiment of the present application, by using the sample local relationship graph constructed by the sample accounts and sample object resources corresponding to the target platform, and the sample global relationship graph constructed by the sample accounts and sample object resources corresponding to the target platform and the reference platform, the feature aggregation network and the resource recommendation network are jointly trained. This allows the feature aggregation network and the resource recommendation network to learn more comprehensive and rich content, thereby achieving the purpose of improving the robustness of the feature aggregation network and the resource recommendation network. In addition, in the embodiment of the present application, by using the sample local relationship graph and the sample global relationship graph to train the feature aggregation network and the resource recommendation network for obtaining the recommendation coefficients of the candidate user accounts in the target platform for the above-mentioned object resources, the feature aggregation network and the resource recommendation network are not only dependent on the data in the reference platform, but also affected by the data in the target platform. This ensures that the content learned by the above-mentioned feature aggregation network and resource recommendation network is not excessively affected by the data in the reference platform. Instead, while learning the data in the reference platform, the characteristics of the target platform are retained, thereby improving the accuracy of the recommendation coefficient acquisition. In other words, the embodiment of the present application solves the problem that the resource recommendation method provided by the related art has a relatively single recommendation method, which leads to low accuracy of resource recommendation. The technical effect of improving the accuracy of resource recommendations is achieved.

[0224] Optionally, as an optional solution, obtaining the first account personality feature obtained during the aggregation processing of the sample key account nodes in the sample local relationship graph and the second account personality feature obtained during the aggregation processing of the sample key account nodes in the sample global relationship graph includes:

[0225] S1, obtain the first sample point edge aggregation result and the second sample point edge aggregation result corresponding to the sample key account node, wherein the first sample point edge aggregation result is obtained by aggregating the first sample neighbor node aggregation result and the first sample connection edge aggregation result corresponding to the sample key account node in the sample local relationship graph, the first sample neighbor node aggregation result is the result of feature aggregation processing on the first sample neighbor resource node and the first sample neighbor account node corresponding to the sample key account in the sample local relationship graph, and the first sample connection edge aggregation result is the connection edge between the sample key account and the first sample neighbor resource node in the sample local relationship graph, and the connection edge between the sample key account and the first sample neighbor account node. The second sample point edge aggregation result is the result of feature aggregation processing on the connection edges between the points. The second sample point edge aggregation result is obtained by aggregating the second sample neighbor node aggregation result corresponding to the sample key account node in the sample global relationship graph and the second sample connection edge aggregation result. The second sample neighbor node aggregation result is the result of feature aggregation processing on the second sample neighbor resource node and the second sample neighbor account node corresponding to the sample key account in the sample global relationship graph. The second sample connection edge aggregation result is the result of feature aggregation processing on the connection edge between the sample key account and the second sample neighbor resource node, and the connection edge between the sample key account and the second sample neighbor account node in the sample local relationship graph.

[0226] Optionally, in this embodiment, the step of obtaining the first sample point edge aggregation result and the second sample point edge aggregation result corresponding to the sample key account node may include, but is not limited to, the following steps:

[0227] Set the account node of the key user account and the resource node of the key object resource as the current node, and perform the following operations in the feature aggregation network:

[0228] Step S1-1, determining a first current sample neighbor resource node connected to the current sample node and a first current sample neighbor account node connected to the current sample node from the sample local relationship graph, and determining a second current sample neighbor resource node connected to the current sample node and a second current sample neighbor account node connected to the current sample node from the sample global relationship graph;

[0229] Step S1-2, obtaining the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current sample node, and the first sample connection edge aggregation result and the second sample connection edge aggregation result corresponding to the current sample node, wherein the first neighbor node aggregation result is used to indicate the result obtained by performing feature aggregation processing on the first current sample neighbor resource node and the first current sample neighbor account node, and the second sample neighbor node aggregation result is used to indicate the result obtained by performing feature aggregation processing on the second current sample neighbor resource node and the second current sample neighbor account node. The first sample connection edge aggregation result is used to indicate the result obtained by performing feature aggregation processing on the sample connection edge between the current sample node and the first current sample neighbor resource node, and the sample connection edge between the current sample node and the first current sample neighbor account node in the sample local relationship graph. The second sample connection edge aggregation result is used to indicate the result obtained by performing feature aggregation processing on the sample connection edge between the current sample node and the second current sample neighbor resource node, and the sample connection edge between the current sample node and the second current sample neighbor account node in the sample global relationship graph;

[0230] Step S1-3: Perform feature aggregation processing on the first sample neighbor node aggregation result and the first sample connection edge aggregation result to obtain the above-mentioned first sample point edge aggregation result, and perform feature aggregation processing on the second sample neighbor node aggregation result and the second sample connection edge aggregation result to obtain the second sample point edge aggregation result.

[0231] S2, obtaining the first account personality feature of the sample key account node in the sample local relationship graph from the first sample point edge aggregation result, and obtaining the second account personality feature of the sample key account node in the sample global relationship graph from the second sample point edge aggregation result.

[0232] In the embodiment of the present application, by obtaining the node account personality features of the user account corresponding to the same user in the sample local relationship graph and the sample global relationship graph, and then using the node account personality features of the user account corresponding to the same user in the sample local relationship graph and the sample global relationship graph to generate the constraint conditions for training the feature aggregation network and the resource recommendation network, it is ensured that after the account node is processed by the feature aggregation network, the account personality features used to represent the preferences of the account corresponding to the account node can still be retained in the account node. This achieves the purpose of improving the robustness of the feature aggregation network and the resource recommendation network.

[0233] Optionally, as an optional solution, the sample node state features are input into the resource recommendation network in training, and the recommendation training loss of the resource recommendation network is obtained, including:

[0234] S1, obtain the resource recommendation evaluation coefficient output by the resource recommendation network in training that matches the state characteristics of the sample node.

[0235] It should be noted that, in this embodiment, the resource recommendation evaluation coefficients output by the resource recommendation network during training that match the state characteristics of the sample nodes may include, but are not limited to:

[0236] Step S1-1: input the node status features corresponding to the sample key account nodes and the node status features corresponding to the sample key resource nodes into the resource recommendation network.

[0237] Step S1-2: In the first-layer resource recommendation network included in the resource recommendation network, perform the following operations:

[0238] The node status features corresponding to the account node and the node status features corresponding to the resource node are concatenated to obtain a connection result. Next, the connection result is multiplied by the first resource recommendation network parameter to obtain a first multiplication result. Furthermore, the first multiplication result and the second resource recommendation network parameter are weighted and summed to obtain a summed result. The summed result is then subjected to a first target rule calculation to obtain the output of the first-layer resource recommendation network. The first target rule calculation may include, but is not limited to, squaring the summed result, taking the absolute value of the summed result, or taking the logarithm of the summed result.

[0239] Step S1-3: input the output result of the first-layer resource recommendation network into the second-layer resource recommendation network included in the resource recommendation network.

[0240] Step S1-4: performing the following operations in the second-layer resource recommendation network included in the resource recommendation network:

[0241] The third resource recommendation network parameter is multiplied by the output of the first-layer resource recommendation network to obtain a second multiplication result. Then, a weighted sum calculation is performed on the second multiplication result and the fourth resource recommendation network parameter to obtain a resource recommendation evaluation coefficient that matches the state characteristics of the sample node.

[0242] It should be noted that, for specific examples of the above steps, please refer to the above examples of obtaining the recommendation coefficients of the candidate user accounts for the above object resources, which will not be described in detail in this embodiment.

[0243] S2, comparing the resource recommendation evaluation coefficient and the recommendation information indicated by the node reference label corresponding to the sample node state feature to obtain the recommendation difference result.

[0244] Optionally, in this embodiment, the recommendation information indicated by the node reference label corresponding to the above-mentioned sample node status feature can be used, but is not limited to, to indicate the user account corresponding to the key user account node corresponding to the above-mentioned sample node status feature, and the real score of the object resource corresponding to the key user resource node corresponding to the above-mentioned sample node status feature in the target platform.

[0245] Furthermore, the above-mentioned comparison of the resource recommendation evaluation coefficient and the recommendation information indicated by the node reference tag corresponding to the sample node state feature to obtain the recommendation difference result can be implemented by referring to, but not limited to, the following examples:

[0246]

[0247] Among them, the above r v，k It can be used, but is not limited to, to indicate the recommended information indicated by the node reference label, the above It can be used, but is not limited to, to represent the resource recommendation evaluation coefficient. The above c can be used, but is not limited to, to represent the above recommendation difference result. The above R1 can be used, but is not limited to, to represent a set of sample key node pairs, wherein the above sample key node pairs include sample key user account nodes and sample key resource nodes with evaluation association relationships. The above evaluation association relationship is used to indicate that the user account corresponding to the sample key user account node has performed an evaluation operation on the object resource corresponding to the sample key resource node.

[0248] S3, calculates the recommendation training loss based on the recommendation difference results.

[0249] Optionally, the above implementation of calculating the recommendation training loss based on the recommendation difference result may refer to, but is not limited to, the following examples:

[0250]

[0251] Among them, the above-mentioned G(U) can be but is not limited to being used to represent a set of sample key user account nodes, the above-mentioned G(I) can be but is not limited to being used to represent a set of sample key resource nodes, and the above-mentioned R can be but is not limited to being used to represent the number of sample key node pairs, wherein the above-mentioned sample key node pairs include sample key user account nodes and sample key resource nodes with evaluation association relationships, and the above-mentioned evaluation association relationships are used to indicate that the user account corresponding to the sample key user account node has performed evaluation operations on the object resources corresponding to the sample key resource node.

[0252] In the embodiments of the present application, by utilizing the resource recommendation evaluation coefficients output by the resource recommendation network and the node reference labels to generate a constraint loss (i.e., recommendation training loss) for training the feature aggregation network and the resource recommendation network, the results output by the resource recommendation network are made more accurate, thereby achieving the goal of improving the robustness of the feature aggregation network and the resource recommendation network.

[0253] Alternatively, as an optional implementation, it can be based on but not limited to the following Figure 7 The following steps are shown to fully illustrate the recommended approach for the above object resources:

[0254] In the aforementioned object resource recommendation method, training of the feature aggregation network and resource recommendation network is primarily accomplished through the embedding learning layer, the embedding optimization layer, and the joint loss learning layer. The embedding learning layer also includes a graph construction layer, a graph enhancement layer, and an embedding generation layer. The joint loss learning layer also includes a feature alignment layer and a joint loss function layer.

[0255] Specifically, the training process of the feature aggregation network and resource recommendation network in the above object resource recommendation method can refer to the following steps:

[0256] The following steps S1-S3 are performed in the embedding learning layer:

[0257] Step S1, perform the following steps in the graph construction layer: Figure 7 As shown, the target local relationship graph is constructed using candidate user accounts and object resources in the target domain (used to represent the target platform under the target domain). At the same time, the reference user accounts and reference object resources in the source domain (used to represent the reference platform under other reference domains other than the target domain) are used to construct the reference local relationship graph. Then, the common nodes included in the target local relationship graph corresponding to the target domain and the reference local relationship graph corresponding to the source domain are merged to obtain a global graph (used to represent the global relationship graph).

[0258] In step S2, the following steps are performed in the graph enhancement layer: obtaining historical evaluation information published by each candidate user account in the target local relationship graph on the associated platform, wherein the associated platform uses the identity information of the candidate user account to log in to the platform other than the target platform; extracting the account personality characteristics corresponding to each candidate user account from the historical evaluation information; determining similar account pairs from the candidate user accounts based on the feature similarity between the account personality characteristics; and connecting the account nodes corresponding to the similar account pairs to obtain an updated target local relationship graph. Furthermore, obtaining reference historical evaluation information published by each reference user account on the associated platform, wherein the associated platform uses the identity information of the reference user account to log in to the platform other than the target platform; extracting the reference account personality characteristics corresponding to each reference user account from the historical evaluation information; determining the account nodes corresponding to the candidate user accounts and the account nodes corresponding to the reference user accounts with similarity relationships in the global relationship graph based on the feature similarity between the account personality characteristics; and connecting the account nodes corresponding to the candidate user accounts and the account nodes corresponding to the reference user accounts with similarity relationships to obtain an updated global relationship graph.

[0259] Step S3, in the embedding generation layer, the following steps are performed: based on the point-to-point connection relationship and the point-to-edge connection relationship in the target local relationship graph in the feature aggregation network, the node embedding of each account node in each target local relationship graph is determined (used to represent the node state features of each account node above, in Figure 7 G lv Representation), and the node embedding of each resource node in the target local relationship graph (used to represent the node state characteristics of each resource node above, in Figure 7 G lk Represented); and based on the point-to-point connection relationship and the point-to-edge connection relationship in the global relationship graph, determine the node embedding of each account node in each global (used to represent the node state characteristics of each account node above, in Figure 7 G gv Representation), and the node embedding of each resource node in the global (used to represent the node state characteristics of each resource node above, in Figure 7 G gk In the above target local relationship graph, the node embedding of each account node (used to represent the node state features of each account node) also includes domain-related features (in Figure 7 Chinese representation) and domain-independent features (in Figure 7 Chinese Represented), accordingly, the node embedding (i.e., node state features) of each account node in the above-mentioned global relationship graphs also includes domain-related features (in Figure 7 Chinese representation) and domain-independent features (in Figure 7 Chinese In which, Figure 7 G(U) in is used to represent user node embedding. Figure 7 G(I) in is used to represent resource node embedding.

[0260] The following steps S1-S2 are performed in the embedding optimization and joint loss learning layers:

[0261] Step S1, in the feature alignment layer, performs the following steps: aligning user nodes, specifically, obtaining the first account personality feature (such as Figure 7 in ), and the second account personality features (such as Figure 7 in ). Then, using the first account personality feature and the second account personality feature, the aggregation training loss of the feature aggregation network is calculated (such as Figure 7 l in c ).

[0262] Step S2, in the joint loss function layer, performs the following steps: Obtain the scoring loss (used to represent the recommendation training loss): S2-1, embed the node of the account node corresponding to the same user account included in the target local relationship graph and the global relationship graph in the feature aggregation network (such as Figure 7 G in lv and G gv ) to obtain the node embedding corresponding to the above user account. And embed the node of the resource node corresponding to the same object resource included in the target local relationship graph and the global relationship graph (such as Figure 7 G in lk and G gk ) are fused to obtain the node embedding corresponding to the above object resource. S2-2, the node embedding corresponding to the above user account and the node embedding corresponding to the above object resource are input into the feature aggregation network to obtain the recommendation coefficient of the above user account for the above object resource. S2-3, the recommendation training loss (such as Figure 7 l in r ). S2-4, obtain the joint loss function: perform weighted sum calculation on the aggregate training loss and the recommendation training loss to obtain the joint training loss (such as Figure 71). S2-5: If the joint training loss does not reach the threshold condition, adjust the aggregation network parameters in the feature aggregation network and the recommendation network parameters in the resource recommendation network. If the joint training loss reaches the threshold condition, determine that the feature aggregation network and the resource recommendation network have reached the network convergence condition.

[0263] It should be noted that the above Figure 7 The corresponding example is an optional example for the purpose of explaining the present embodiment. The present embodiment is not limited to the above example, and no limitation is made to this in the present embodiment.

[0264] When the feature aggregation network and resource recommendation network reach convergence conditions, the feature aggregation network and resource recommendation network can be used to determine the recommendation coefficient of each candidate user account for each object resource in the target platform, and then determine the target object resource to be recommended for each candidate user account based on the recommendation coefficient. Specifically, it includes the following: Figure 8 The following steps are shown:

[0265] S802, build a relationship graph: Use the candidate user accounts and object resources in the target domain (i.e., the target platform under the target domain) to build a target local relationship graph. Simultaneously, use the reference user accounts and reference object resources in the source domain (i.e., the reference platform under other reference domains other than the target domain) to build a reference local relationship graph. Then, merge the common nodes included in the target local relationship graph corresponding to the target domain and the reference local relationship graph corresponding to the source domain to obtain a global graph (i.e., the global relationship graph).

[0266] S804, Enhance the Relationship Graph: Obtain historical evaluation information published by each candidate user account in the target local relationship graph on the associated platform, wherein the associated platform uses the identity information of the candidate user account to log in to the platform other than the target platform; extract the account personality characteristics corresponding to each candidate user account from the historical evaluation information; determine similar account pairs from the candidate user accounts based on the feature similarity between the account personality characteristics; connect the account nodes corresponding to the similar account pairs to obtain an updated target local relationship graph. Also obtain reference historical evaluation information published by each reference user account on the associated platform, wherein the associated platform uses the identity information of the reference user account to log in to the platform other than the target platform; reference account personality characteristics corresponding to each reference user account extracted from the historical evaluation information; determine the account nodes corresponding to the candidate user accounts with similarity relationships and the account nodes corresponding to the reference user accounts in the global relationship graph based on the feature similarity between the account personality characteristics; connect the account nodes corresponding to the candidate user accounts with similarity relationships and the account nodes corresponding to the reference user accounts to obtain an updated global relationship graph.

[0267] S806, feature generation: In the feature aggregation network, based on the point-to-point connection relationship and the point-edge connection relationship in the target local relationship graph, determine the node embedding (i.e., node state feature) of each account node in each target local relationship graph, and the node embedding (i.e., node state feature) of each resource node in the target local relationship graph; and based on the point-to-point connection relationship and the point-edge connection relationship in the global relationship graph, determine the node embedding (i.e., node state feature) of each account node in each global, and the node embedding (i.e., node state feature) of each resource node in the global.

[0268] S808, Feature Fusion: Aggregate the node embeddings (i.e., node state features) of the account nodes corresponding to the same user account in the target local relationship graph and the global relationship graph in the feature aggregation network to obtain the node embeddings (i.e., node state features) corresponding to the user account. Also aggregate the node embeddings (i.e., node state features) of the resource nodes corresponding to the same redemption resource in the target local relationship graph and the global relationship graph to obtain the node embeddings (i.e., node state features) corresponding to the object resource.

[0269] S810, obtaining a recommendation coefficient: inputting the node embedding (i.e., node state feature) corresponding to the above-mentioned user account and the node embedding (i.e., node state feature) corresponding to the above-mentioned object resource into a feature aggregation network to obtain the recommendation coefficient of the above-mentioned user account for the above-mentioned object resource.

[0270] S812, resource recommendation: determining candidate user accounts and object resources corresponding to recommendation coefficients greater than a recommendation coefficient threshold, and recommending the object resources to the candidate user accounts.

[0271] That is to say, by using the embodiment of the present application, by using the target local relationship graph and the global relationship graph to jointly obtain the node status characteristics of the account node corresponding to the candidate user account, and the node status characteristics of the resource node corresponding to the object resource, the content learned by the account node and resource node corresponding to the target platform not only depends on the data in the reference platform, but is also affected by the data in the target platform. This ensures that the content learned by the above-mentioned account node and resource node will not be excessively affected by the data in the reference platform, but while learning the data in the reference platform, the characteristics of the target platform are retained. As a result, the node status characteristics of the account node corresponding to the candidate user account and the node status characteristics of the resource node corresponding to the object resource obtained are more accurate, thereby improving the accuracy of the recommendation coefficient obtained based on the node status characteristics of the account node corresponding to the candidate user account and the node status characteristics of the resource node corresponding to the object resource. Furthermore, by increasing the connection relationships in the target local relationship graph based on the feature similarity between the individual characteristics of the accounts, the relationships between candidate user accounts and object resources can be intuitively and fully reflected in the target local relationship graph. This allows each candidate user account and object resource to learn more content by leveraging the relationships presented in the target local relationship graph, thereby improving the reliability of the recommendation coefficients between each candidate user account and each object resource determined using the target local relationship graph and the global relationship graph. By embedding account individual characteristics in account nodes, the content included in the account nodes in the target local relationship graph and the global relationship graph is richer and better reflects the unique preferences of the user accounts. This further allows each candidate user account and object resource to learn more content by leveraging the relationships presented in the target local relationship graph, thereby improving the reliability of the recommendation coefficients between each candidate user account and each object resource determined using the target local relationship graph and the global relationship graph. This allows the candidate user account to be recommended object resources of interest using these recommendation coefficients. This solves the technical problem of low resource recommendation accuracy due to the overly simplistic resource recommendation methods provided in related technologies, achieving the technical effect of improving the accuracy of resource recommendations. Furthermore, the recommendation coefficient is used to recommend object resources of interest to candidate user accounts. This effectively recommends object resources on the target platform during the cold start process, enabling more user accounts to use these resources. This addresses the technical issue of insufficient user accounts actually using the target platform's object resources during the cold start process, which can hinder the full utilization of the object resources' functionality and value.

[0272] Optionally, the above Figure 8The corresponding example is an optional example for the purpose of explaining the present embodiment. The present embodiment is not limited to the above example, and no limitation is made to this in the present embodiment.

[0273] It should be noted that for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0274] It is further noted that the collection and processing of the relevant data in this application should be strictly in accordance with the requirements of relevant national laws and regulations when applied in practice, and the informed consent or separate consent of the personal information subject should be obtained. Subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the personal information subject.

[0275] According to another aspect of the embodiment of the present application, there is also provided an object resource recommendation device for implementing the above object resource recommendation method. Figure 9 As shown, the device includes:

[0276] A construction unit 902 is configured to construct a target local relationship graph that matches the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship indicating the relationship between the candidate user accounts and the object resources;

[0277] A first determining unit 904 is configured to determine a recommendation coefficient between each candidate user account and each object resource using the target local relationship graph and the global relationship graph, wherein the global relationship graph is a relationship graph constructed based on the target local relationship graph and a reference local relationship graph matched with a reference platform other than the target platform;

[0278] The second determining unit 906 is configured to determine the target object resource to be recommended for each candidate user account according to the recommendation coefficient.

[0279] Optionally, in this embodiment, the above-mentioned construction unit includes: a first determination module, which is used to use the candidate user account as the account node and the object resource as the resource node; and the resource node of each object resource is used as the current resource node in turn; a second determination module, which is used to use the account node of the candidate user account that has performed an evaluation operation on the current object resource corresponding to the current resource node as the current associated account node associated with the current resource node; a second connection module, which is used to connect the current resource node and the current associated account node; a first acquisition module, which is used to obtain the next resource node as the current resource node when the current resource node is not the last resource node in the target local relationship graph; and a second acquisition module, which is used to determine and construct a target local relationship graph that matches the target platform when the current resource node is the last resource node in the target local relationship graph.

[0280] Optionally, in this embodiment, the above-mentioned device also includes: a first acquisition unit, used to obtain the reference local relationship graphs matched by each reference platform, wherein, in the reference local relationship graph, the reference object resources published by the reference platform serve as reference resource nodes, and the user accounts that have performed evaluation operations on at least one reference object resource in the reference platform serve as reference account nodes, and the reference account nodes are connected to the reference resource nodes having an evaluation association relationship with the reference account nodes, and the evaluation association relationship indicates that the user account corresponding to the reference account node has performed an evaluation operation on the object resource corresponding to the reference resource node; a second acquisition unit, used to obtain the union of the reference account nodes contained in the reference local relationship graph and the account nodes contained in the target local relationship graph to obtain the global account node in the global relationship graph, and obtain the union of the reference resource nodes contained in the reference local relationship graph and the resource nodes contained in the target local relationship graph to obtain the global resource node in the global relationship graph; a first connection unit, used to connect the global account node and the global resource node having an evaluation association relationship with the global account node to construct a global relationship graph, wherein the evaluation association relationship indicates that the user account corresponding to the global account node has performed an evaluation operation on the object resource corresponding to the global resource node.

[0281] Optionally, in this embodiment, the above-mentioned device also includes: a third acquisition unit, used to obtain historical evaluation information published by each candidate user account in the target local relationship graph on the associated platform, wherein the associated platform is a platform other than the target platform logged in using the identity information of the candidate user account; a first extraction unit, used to extract the account personality characteristics corresponding to each candidate user account from the historical evaluation information; a third determination unit, used to determine similar account pairs from the candidate user accounts based on the feature similarity between the account personality characteristics; a second connection unit, used to connect the account nodes corresponding to each similar account pair to obtain an updated target local relationship graph.

[0282] Optionally, in this embodiment, the above-mentioned first determination unit includes: a third determination module, used to determine the node state characteristics based on the point-to-point connection relationship and the point-edge connection relationship in the target local relationship graph, and the point-to-point connection relationship and the point-edge connection relationship in the global relationship graph, wherein the node state characteristics include the account node state characteristics corresponding to each candidate user account, and the resource node state characteristics corresponding to each object resource; a first input module, used to input the node state characteristics into the resource recommendation network to obtain the recommendation coefficient of the candidate user account for the above-mentioned object resource, wherein the resource recommendation network is trained using the reference local relationship graph and the target local relationship graph.

[0283] Optionally, in this embodiment, the above-mentioned third determination module is also used to determine the account node of the key user account with the same account relationship, and the resource node of the key object resource with the same resource identifier from the target local relationship graph and the global relationship graph; the account node of the key user account and the resource node of the key object resource are respectively used as the current node, and the following operations are performed in the feature aggregation network: determine the first current neighbor resource node connected to the current node and the first current neighbor account node connected to the current node from the target local relationship graph, and determine the second current neighbor resource node connected to the current node and the second current neighbor account node connected to the current node from the global relationship graph; obtain the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current node, and the first connection edge aggregation result and the second connection edge aggregation result corresponding to the current node, wherein the first neighbor node aggregation result is used to indicate the first current neighbor resource node and The result of feature aggregation processing is obtained by performing feature aggregation processing on the first current neighbor account node, the second neighbor node aggregation result is used to indicate the result of feature aggregation processing on the second current neighbor resource node and the second current neighbor account node, the first connection edge aggregation result is used to indicate the result of feature aggregation processing on the connection edge between the current node and the first current neighbor resource node, and the connection edge between the current node and the first current neighbor account node in the target local relationship graph, the second connection edge aggregation result is used to indicate the result of feature aggregation processing on the connection edge between the current node and the second current neighbor resource node, and the connection edge between the current node and the second current neighbor account node in the global relationship graph; the current node state feature corresponding to the current node is determined by using the first neighbor node aggregation result, the second neighbor node aggregation result, the first connection edge aggregation result and the second connection edge aggregation result; wherein, the feature aggregation network is obtained by joint training with the resource recommendation network.

[0284] Optionally, in this embodiment, the above-mentioned third determination module is also used to aggregate the node status characteristics of the first current neighbor resource node and the node status characteristics of the first current neighbor account node to obtain a first neighbor node aggregation result, wherein the node status characteristics of the first current neighbor account node include the account personality characteristics of the neighbor account corresponding to the first current neighbor account node, and the node status characteristics corresponding to the first current neighbor resource node are generated based on the resource description information of the neighbor object resource corresponding to the first current neighbor resource node; aggregate the node status characteristics of the second current neighbor resource node and the node status characteristics of the second current neighbor account node to obtain a second neighbor node aggregation result, wherein the node status characteristics of the second current neighbor account node include the account personality characteristics of the neighbor account corresponding to the second current neighbor account node, and the node status characteristics corresponding to the second current neighbor resource node are generated based on the resource description information of the neighbor object resource corresponding to the second current neighbor resource node; classify and aggregate the first edge status characteristics of the connection edge between the current node and the first current neighbor resource node, and the second edge status characteristics of the connection edge between the current node and the first current neighbor account node to obtain a first connection edge aggregation result, wherein, in the case where the current node is the account node of the key user account, the first edge status characteristics are based on The evaluation information of the key user account corresponding to the current node on the object resource corresponding to the first current neighbor resource node is obtained, and the second side state feature is obtained based on the feature similarity between the account personality feature of the key user account corresponding to the current node and the account personality feature of the user account corresponding to the first current neighbor account node; when the current node is a resource node of the key object resource, the first side state feature is obtained based on the similarity between the object resource corresponding to the current node and the object resource corresponding to the first current neighbor resource node, and the second side state feature is obtained based on the evaluation information of the key user account corresponding to the first current neighbor account node on the object resource corresponding to the current node; the third side state feature of the connection edge between the current node and the second current neighbor resource node, and the fourth side state feature of the connection edge between the current node and the second current neighbor account node are classified and aggregated to obtain a second connection edge aggregation result, wherein, when the current node is an account node of the key user account, the third side state feature is obtained based on the evaluation information of the key user account corresponding to the current node on the object resource corresponding to the second current neighbor resource node, and the fourth side state feature is obtained based on the feature similarity between the account personality feature of the key user account corresponding to the current node and the account personality feature of the user account corresponding to the second current neighbor account node;When the current node is a resource node for a key object resource, the third edge state feature is obtained based on the similarity between the object resource corresponding to the current node and the object resource corresponding to the second current neighbor resource node. The fourth edge state feature is obtained based on the evaluation information of the key user account corresponding to the second current neighbor account node on the object resource corresponding to the current node.

[0285] Optionally, in this embodiment, the above-mentioned third determination module is also used to aggregate the first neighbor node aggregation result and the first connection edge aggregation result to obtain a target point-edge aggregation result that matches the target local relationship graph; aggregate the second neighbor node aggregation result and the second connection edge aggregation result to obtain a global point-edge aggregation result that matches the global relationship graph; and fuse the target point-edge aggregation result and the global point-edge aggregation result to obtain the current node state feature.

[0286] Optionally, in this embodiment, the above-mentioned device also includes: a fourth determination unit, which is used to determine the first sample account node and the first sample resource node from the sample local relationship graph, and determine the second sample account node and the second sample resource node from the sample global relationship graph; a fifth determination unit, which is used to determine the sample key account node with the same account information from the first sample account node and the second sample account node, and the sample resource node of the sample key object resource with the same resource identifier from the first sample resource node and the second sample resource node; a sixth determination unit, which is used to traverse the sample key account node and the sample resource node, and determine the sample node state characteristics obtained after aggregation processing in the sample local relationship graph and the sample global relationship graph; a fourth acquisition unit, which is used to obtain the first sample key account node obtained in the process of aggregation processing in the sample local relationship graph. An account personality feature, and a second account personality feature obtained during the aggregation processing of the sample key account nodes in the sample global relationship graph, and using the first account personality feature and the second account personality feature to calculate the aggregation training loss of the feature aggregation network; a first input unit, used to input the sample node state feature into the resource recommendation network in training, and obtain the recommendation training loss of the resource recommendation network; a first calculation unit, used to perform weighted sum calculation on the aggregation training loss and the recommendation training loss to obtain the joint training loss; a first adjustment unit, used to adjust the aggregation network parameters in the feature aggregation network and the recommendation network parameters in the resource recommendation network when the joint training loss does not reach the threshold condition; a seventh determination unit, used to determine that the feature aggregation network and the resource recommendation network have reached the network convergence condition when the joint training loss reaches the threshold condition.

[0287] Optionally, in this embodiment, the fourth acquisition unit includes: a third acquisition module, configured to acquire a first sample point edge aggregation result and a second sample point edge aggregation result corresponding to the sample key account node, wherein the first sample point edge aggregation result is obtained by aggregating the first sample neighbor node aggregation result and the first sample connection edge aggregation result corresponding to the sample key account node in the sample local relationship graph, the first sample neighbor node aggregation result is the result obtained by performing feature aggregation processing on the first sample neighbor resource node and the first sample neighbor account node corresponding to the sample key account in the sample local relationship graph, the first sample connection edge aggregation result is the result obtained by performing feature aggregation processing on the connection edge between the sample key account and the first sample neighbor resource node, and the connection edge between the sample key account and the first sample neighbor account node in the sample local relationship graph, the second sample point edge aggregation result is the result obtained by performing feature aggregation processing on the second sample neighbor node corresponding to the sample key account node in the sample global relationship graph The aggregation result and the second sample connection edge aggregation result are obtained by aggregating the second sample neighbor node aggregation result, which is the result of feature aggregation processing of the second sample neighbor resource node and the second sample neighbor account node corresponding to the sample key account in the sample global relationship graph. The second sample connection edge aggregation result is the result of feature aggregation processing of the connection edge between the sample key account and the second sample neighbor resource node, and the connection edge between the sample key account and the second sample neighbor account node in the sample local relationship graph; the first fusion module is used to fuse the first sample point edge aggregation result and the second sample point edge aggregation result to determine the sample node state feature corresponding to the sample key account node; the fourth acquisition module is used to obtain the first account personality feature of the sample key account node in the sample local relationship graph from the first sample point edge aggregation result, and obtain the second account personality feature of the sample key account node in the sample global relationship graph from the second sample point edge aggregation result.

[0288] Optionally, in this embodiment, the above-mentioned first input unit includes: a fifth acquisition module, used to obtain the resource recommendation evaluation coefficient output by the resource recommendation network in training that matches the sample node state characteristics; a first comparison module, used to compare the resource recommendation evaluation coefficient and the recommendation information indicated by the node reference label corresponding to the sample node state characteristics to obtain a recommendation difference result; and a first training module, used to calculate the recommendation training loss based on the recommendation difference result.

[0289] In an embodiment of the present application, a target local relationship graph that matches the target platform is constructed based on the object resources released by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship for indicating the candidate user accounts and the object resources. Then, the target local relationship graph and the global relationship graph are used to determine the recommendation coefficient between each candidate user account and each object resource, wherein the global relationship graph is a relationship graph jointly constructed based on the target local relationship graph and the reference local relationship graph matched by the reference platform other than the target platform. Then, the target object resource to be recommended is determined for each candidate user account according to the recommendation coefficient. In other words, using an embodiment of the present application, the target local relationship graph of the target platform constructed based on the object resource body in the target platform and the candidate user accounts in the cold start pool, and the global relationship graph constructed by combining the target local relationship graph and the reference local relationship graph corresponding to the reference platform other than the target platform, is used to obtain the recommendation coefficient between each candidate user account corresponding to the target platform and each object resource. This makes the resulting recommendation coefficient more reliable, allowing the target resource to be recommended for each candidate user account based on the recommendation coefficient, adapting to the actual needs of the candidate user account. This avoids the problem of low resource recommendation accuracy caused by the relatively simple recommendation methods provided by related technologies, thereby achieving the technical effect of improving the accuracy of resource recommendations.

[0290] For a specific embodiment, reference may be made to the example shown in the above-mentioned method for recommending object resources, which will not be described in detail in this embodiment.

[0291] According to another aspect of the embodiment of the present application, an electronic device for implementing the above-mentioned object resource recommendation method is also provided. This embodiment is described using the electronic device as an example. Figure 10 As shown, the electronic device includes a memory 1002 and a processor 1004. The memory 1002 stores a computer program, and the processor 1004 is configured to execute the steps in any of the above method embodiments through the computer program.

[0292] Optionally, in this embodiment, the electronic device may be located in at least one network device among a plurality of network devices of a computer network.

[0293] Optionally, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0294] S1, constructing a target local relationship graph that matches the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship indicating the relationship between the candidate user accounts and the object resources;

[0295] S2, using the target local relationship graph and the global relationship graph, determining the recommendation coefficient between each candidate user account and each object resource, wherein the global relationship graph is a relationship graph constructed based on the target local relationship graph and the reference local relationship graphs matched by the reference platform other than the target platform;

[0296] S3, determining the target object resource to be recommended for each candidate user account according to the recommendation coefficient.

[0297] Alternatively, those skilled in the art will appreciate that Figure 10 The structure shown is for illustration only, and the electronic device may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile Internet device (MID), a PAD, or other terminal devices. Figure 10 It does not limit the structure of the above electronic device. For example, the electronic device may also include Figure 10 More or fewer components (such as network interfaces, etc.) as shown in, or with Figure 10 Different configurations shown.

[0298] Among them, the memory 1002 can be used to store software programs and modules, such as the program instructions / modules corresponding to the method and device for recommending object resources in the embodiments of the present application. The processor 1004 executes various functional applications and data processing by running the software programs and modules stored in the memory 1002, that is, realizes the above-mentioned method for recommending object resources. The memory 1002 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1002 may further include a memory remotely located relative to the processor 1004, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned networks include but are not limited to the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Among them, the memory 1002 can be used specifically but not limited to store target local relationship graphs, global relationship graphs and other information. As an example, such as Figure 10As shown, the memory 1002 may include, but is not limited to, the construction unit 902, the first determination unit 904, and the second determination unit 906 of the object resource recommendation device. In addition, it may also include, but is not limited to, other module units of the object resource recommendation device, which will not be repeated in this example.

[0299] Optionally, the transmission device 1006 is configured to receive or send data via a network. Specific examples of the network may include a wired network and a wireless network. In one embodiment, the transmission device 1006 includes a network interface controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In one embodiment, the transmission device 1006 is a radio frequency (RF) module, which is configured to communicate with the Internet wirelessly.

[0300] In addition, the electronic device further includes: a connection bus 1008 for connecting various module components in the electronic device.

[0301] In other embodiments, the terminal device or server may be a node in a distributed system, wherein the distributed system may be a blockchain system, and the blockchain system may be a distributed system formed by connecting multiple nodes via network communication. The nodes may form a point-to-point network, and any computing device, such as a server, terminal, or other electronic device, may become a node in the blockchain system by joining the point-to-point network.

[0302] According to one aspect of the present application, a computer program product is provided, comprising a computer program / instructions containing program code for executing the above-described method. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component and / or installed from a removable medium. When the computer program is executed by a central processing unit, the various functions provided in the embodiments of the present application are performed.

[0303] According to one aspect of the present application, another computer program product is provided, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method in each embodiment of the present application are implemented.

[0304] According to one aspect of the present application, a computer-readable storage medium is provided. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above method.

[0305] Optionally, in this embodiment, the computer-readable storage medium may be configured to store a computer program for performing the following steps:

[0306] S1, constructing a target local relationship graph that matches the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship indicating the relationship between the candidate user accounts and the object resources;

[0307] S2, using the target local relationship graph and the global relationship graph, determining the recommendation coefficient between each candidate user account and each object resource, wherein the global relationship graph is a relationship graph constructed based on the target local relationship graph and the reference local relationship graphs matched by the reference platform other than the target platform;

[0308] S3, determining the target object resource to be recommended for each candidate user account according to the recommendation coefficient.

[0309] Optionally, in the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0310] Optionally, in this embodiment, a person of ordinary skill in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing the hardware related to the terminal device through a program, and the program may be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0311] If the integrated units in the above embodiments are implemented in the form of software functional units and sold or used as independent products, they can be stored in the above-mentioned computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling one or more computer devices (which can be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application.

[0312] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0313] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. Among them, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, and can be electrical or other forms.

[0314] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0315] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0316] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for recommending object resources, characterized in that: include: Based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, a target local relationship graph matching the target platform is constructed, wherein the target local relationship graph includes an evaluation relationship indicating the relationship between the candidate user accounts and the object resources; Determining a recommendation coefficient between each candidate user account and each object resource using the target local relationship graph and the global relationship graph, wherein the global relationship graph is a relationship graph constructed based on the target local relationship graph and a reference local relationship graph matched with a reference platform other than the target platform; The target object resource to be recommended is determined for each candidate user account according to the recommendation coefficient.

2. The method according to claim 1, characterized in that The step of constructing a target local relationship graph matching the target platform based on the object resources released by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform includes: Taking the candidate user account as an account node and the object resource as a resource node; The resource node of each object resource is used as the current resource node in turn, and the following operations are performed: Using the account node of the candidate user account that has performed the evaluation operation on the current object resource corresponding to the current resource node as the current associated account node associated with the current resource node; Connecting the current resource node with the current associated account node; When the current resource node is not the last resource node in the target local relationship graph, obtaining the next resource node as the current resource node; In a case where the current resource node is the last resource node in the target local relationship graph, it is determined to construct the target local relationship graph that matches the target platform.

3. The method according to claim 2, characterized in that After constructing a target local relationship graph matching the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, the method further includes: Obtaining the reference local relationship graphs matched by the reference platforms, wherein, in the reference local relationship graphs, reference object resources published by the reference platforms serve as reference resource nodes, user accounts that have performed evaluation operations on at least one of the reference object resources in the reference platforms serve as reference account nodes, and the reference account nodes are connected to reference resource nodes that have evaluation associations with the reference account nodes, wherein the evaluation associations indicate that the user account corresponding to the reference account nodes has performed evaluation operations on the object resources corresponding to the reference resource nodes; Obtaining a union of the reference account node contained in the reference local relationship graph and the account node contained in the target local relationship graph to obtain a global account node in the global relationship graph, and obtaining a union of the reference resource node contained in the reference local relationship graph and the resource node contained in the target local relationship graph to obtain a global resource node in the global relationship graph; The global account node and the global resource node having an evaluation association relationship with the global account node are connected to construct the global relationship graph, wherein the evaluation association relationship indicates that the user account corresponding to the global account node has performed an evaluation operation on the object resource corresponding to the global resource node.

4. The method according to claim 2, characterized in that After constructing a target local relationship graph matching the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, the method further includes: Obtaining historical evaluation information published by each of the candidate user accounts in the target partial relationship graph on an associated platform, wherein the associated platform is a platform other than the target platform logged in by the candidate user account using the identity information of the candidate user account; Extracting the account personality characteristics corresponding to each of the candidate user accounts from the historical evaluation information; Determining similar account pairs from the candidate user accounts based on feature similarities between the individual features of the accounts; The account nodes corresponding to the similar account pairs are connected to obtain an updated target local relationship graph.

5. The method according to claim 1, wherein Determining the recommendation coefficient between each candidate user account and each object resource by using the target local relationship graph and the global relationship graph includes: Determining node state features based on the point-to-point connection relationships and point-edge connection relationships in the target local relationship graph, and the point-to-point connection relationships and point-edge connection relationships in the global relationship graph, wherein the node state features include account node state features corresponding to each of the candidate user accounts, and resource node state features corresponding to each of the object resources; The node state feature is input into a resource recommendation network to obtain the recommendation coefficient of the candidate user account for the object resource, wherein the resource recommendation network is trained using the reference local relationship graph and the target local relationship graph.

6. The method according to claim 5, characterized in that The determining of the node state feature based on the point-to-point connection relationship and the point-to-edge connection relationship in the target local relationship graph and the point-to-point connection relationship and the point-to-edge connection relationship in the global relationship graph includes: Determining, from the target local relationship graph and the global relationship graph, account nodes of key user accounts having identical account information, and resource nodes of key object resources having identical resource identifiers; The account node of the key user account and the resource node of the key object resource are respectively used as current nodes, and the following operations are performed in the feature aggregation network: Determine, from the target local relationship graph, a first current neighbor resource node connected to the current node and a first current neighbor account node connected to the current node, and determine, from the global relationship graph, a second current neighbor resource node connected to the current node and a second current neighbor account node connected to the current node; Obtain the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current node, and the first connection edge aggregation result and the second connection edge aggregation result corresponding to the current node, wherein the first neighbor node aggregation result is used to indicate the result obtained by performing feature aggregation processing on the first current neighbor resource node and the first current neighbor account node, and the second neighbor node aggregation result is used to indicate the result obtained by performing feature aggregation processing on the second current neighbor resource node and the second current neighbor account node. The first connection edge aggregation result is used to indicate the result obtained by performing feature aggregation processing on the connection edge between the current node and the first current neighbor resource node in the target local relationship graph, and the connection edge between the current node and the first current neighbor account node. The second connection edge aggregation result is used to indicate the result obtained by performing feature aggregation processing on the connection edge between the current node and the second current neighbor resource node in the global relationship graph, and the connection edge between the current node and the second current neighbor account node; Determine a current node state feature corresponding to the current node by using the first neighbor node aggregation result, the second neighbor node aggregation result, the first connection edge aggregation result, and the second connection edge aggregation result; The feature aggregation network is obtained by joint training with the resource recommendation network.

7. The method according to claim 6, characterized in that The obtaining of the first neighbor node aggregation result and the second neighbor node aggregation result corresponding to the current node, and the first connection edge aggregation result and the second connection edge aggregation result corresponding to the current node includes: aggregating the node status characteristics of the first current neighbor resource node and the node status characteristics of the first current neighbor account node to obtain a first neighbor node aggregation result, wherein the node status characteristics of the first current neighbor account node include account personality characteristics of the neighbor account corresponding to the first current neighbor account node, and the node status characteristics corresponding to the first current neighbor resource node are generated based on resource description information of the neighbor object resource corresponding to the first current neighbor resource node; aggregating the node status characteristics of the second current neighbor resource node and the node status characteristics of the second current neighbor account node to obtain a second neighbor node aggregation result, wherein the node status characteristics of the second current neighbor account node include the account personality characteristics of the neighbor account corresponding to the second current neighbor account node, and the node status characteristics corresponding to the second current neighbor resource node are generated based on the resource description information of the neighbor object resource corresponding to the second current neighbor resource node; Classify and aggregate the first edge state feature of the connection edge between the current node and the first current neighbor resource node, and the second edge state feature of the connection edge between the current node and the first current neighbor account node to obtain the first connection edge aggregation result, wherein, when the current node is the account node of the key user account, the first edge state feature is obtained based on the evaluation information of the key user account corresponding to the current node on the object resource corresponding to the first current neighbor resource node, and the second edge state feature is obtained based on the feature similarity between the account personality feature of the key user account corresponding to the current node and the account personality feature of the user account corresponding to the first current neighbor account node; when the current node is the resource node of the key object resource, the first edge state feature is obtained based on the similarity between the object resource corresponding to the current node and the object resource corresponding to the first current neighbor resource node, and the second edge state feature is obtained based on the evaluation information of the key user account corresponding to the first current neighbor account node on the object resource corresponding to the current node; The third side state feature of the connection edge between the current node and the second current neighbor resource node, and the fourth side state feature of the connection edge between the current node and the second current neighbor account node are classified and aggregated to obtain the second connection edge aggregation result, wherein, when the current node is the account node of the key user account, the third side state feature is obtained based on the evaluation information of the key user account corresponding to the current node on the object resource corresponding to the second current neighbor resource node, and the fourth side state feature is obtained based on the feature similarity between the account personality feature of the key user account corresponding to the current node and the account personality feature of the user account corresponding to the second current neighbor account node; when the current node is the resource node of the key object resource, the third side state feature is obtained based on the similarity between the object resource corresponding to the current node and the object resource corresponding to the second current neighbor resource node, and the fourth side state feature is obtained based on the evaluation information of the key user account corresponding to the second current neighbor account node on the object resource corresponding to the current node.

8. The method according to claim 6, characterized in that Determining the current node state feature corresponding to the current node by using the first neighbor node aggregation result, the second neighbor node aggregation result, the first connection edge aggregation result, and the second connection edge aggregation result includes: Aggregating the first neighbor node aggregation result and the first connection edge aggregation result to obtain a target point-edge aggregation result that matches the target local relationship graph; Aggregating the second neighbor node aggregation result and the second connection edge aggregation result to obtain a global point-edge aggregation result that matches the global relationship graph; The target point-edge aggregation result and the global point-edge aggregation result are fused to obtain the current node state feature.

9. The method according to claim 6, characterized in that Before constructing a target local relationship graph matching the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, the method further includes: Determine a first sample account node and a first sample resource node from the sample local relationship graph, and determine a second sample account node and a second sample resource node from the sample global relationship graph; In the feature aggregation network being trained, do the following: Determine a sample key account node having the same account information from the first sample account node and the second sample account node, and a sample resource node of a sample key object resource having the same resource identifier from the first sample resource node and the second sample resource node; Traversing the sample key account nodes and the sample resource nodes, and determining the sample node state characteristics obtained after aggregation processing in the sample local relationship graph and the sample global relationship graph; Obtaining a first account personality feature obtained during the aggregation process of the sample key account node in the sample local relationship graph, and a second account personality feature obtained during the aggregation process of the sample key account node in the sample global relationship graph, and calculating the aggregation training loss of the feature aggregation network using the first account personality feature and the second account personality feature; Inputting the sample node state features into a resource recommendation network in training, and obtaining a recommendation training loss of the resource recommendation network; Performing weighted sum calculation on the aggregate training loss and the recommended training loss to obtain a joint training loss; When the joint training loss does not reach a threshold condition, adjusting aggregation network parameters in the feature aggregation network and recommendation network parameters in the resource recommendation network; When the joint training loss reaches a threshold condition, it is determined that the feature aggregation network and the resource recommendation network meet a network convergence condition.

10. The method according to claim 9, characterized in that The obtaining of the first account personality feature obtained during the aggregation processing of the sample key account node in the sample local relationship graph and the second account personality feature obtained during the aggregation processing of the sample key account node in the sample global relationship graph includes: Obtain a first sample point edge aggregation result and a second sample point edge aggregation result corresponding to the sample key account node, wherein the first sample point edge aggregation result is obtained by aggregating the first sample neighbor node aggregation result and the first sample connection edge aggregation result corresponding to the sample key account node in the sample local relationship graph, the first sample neighbor node aggregation result is the result of feature aggregation processing of the first sample neighbor resource node and the first sample neighbor account node corresponding to the sample key account in the sample local relationship graph, the first sample connection edge aggregation result is the connection edge between the sample key account and the first sample neighbor resource node in the sample local relationship graph, and the connection edge between the sample key account and the first sample neighbor account node in the sample local relationship graph. The second sample point edge aggregation result is obtained by aggregating the second sample neighbor node aggregation result and the second sample connection edge aggregation result corresponding to the sample key account node in the sample global relationship graph. The second sample neighbor node aggregation result is the result of feature aggregation processing on the second sample neighbor resource node and the second sample neighbor account node corresponding to the sample key account in the sample global relationship graph. The second sample connection edge aggregation result is the result of feature aggregation processing on the connection edge between the sample key account and the second sample neighbor resource node, and the connection edge between the sample key account and the second sample neighbor account node in the sample local relationship graph; The first account personality feature of the sample key account node in the sample local relationship graph is obtained from the first sample point edge aggregation result, and the second account personality feature of the sample key account node in the sample global relationship graph is obtained from the second sample point edge aggregation result.

11. The method according to claim 9, characterized in that Inputting the sample node state features into the resource recommendation network in training and obtaining the recommendation training loss of the resource recommendation network includes: Obtaining a resource recommendation evaluation coefficient output by the resource recommendation network in training that matches the state characteristics of the sample node; Comparing the resource recommendation evaluation coefficient with the recommendation information indicated by the node reference tag corresponding to the sample node state feature to obtain a recommendation difference result; The recommendation training loss is calculated based on the recommendation difference result.

12. A device for recommending object resources, characterized in that: include: a construction unit, configured to construct a target local relationship graph that matches the target platform based on the object resources published by the target platform and the candidate user accounts stored in the cold start pool corresponding to the target platform, wherein the target local relationship graph includes an evaluation relationship indicating the evaluation relationship between the candidate user accounts and the object resources; a first determining unit, configured to determine a recommendation coefficient between each of the candidate user accounts and each of the object resources using the target local relationship graph and a global relationship graph, wherein the global relationship graph is a relationship graph constructed based on the target local relationship graph and a reference local relationship graph matched with a reference platform other than the target platform; The second determining unit is configured to determine the target object resource to be recommended for each candidate user account according to the recommendation coefficient.

13. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored program, wherein the program is executed by a processor to perform the method according to any one of claims 1 to 11.

14. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

15. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to execute the method according to any one of claims 1 to 11 through the computer program.

Images