Information recommendation method and device, electronic equipment and computer readable storage medium

Abstract

The invention provides an information recommendation method and device, electronic equipment and a computer readable storage medium. The method comprises the following steps: acquiring a heterogeneous graph comprising a plurality of nodes; determining an identification feature of each node, and performing recoding processing on the identification feature of each node to obtain a recoding identification of each node; splitting the recoding identifier of each node to obtain a sub-identifier of each node; determining a sub-identifier of a to-be-recommended object and a sub-identifier of to-be-recommended information from the sub-identifiers corresponding to the plurality of nodes; and based on the sub-identifier of the to-be-recommended object and the sub-identifier of the to-be-recommended information, determining a recommendation index of the to-be-recommended object corresponding to the to-be-recommended information, and executing a recommendation operation based on the recommendation index. Through the information recommendation method and device, the information recommendation effect can be improved in the information recommendation process in combination with the object identifier of the to-be-recommended object and the information identifier of the to-be-recommended information.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular, to an information recommendation method, apparatus, electronic device, and computer-readable storage medium. Background Art

[0002] Artificial Intelligence (AI) technology is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines, enabling the machines to have the functions of perception, reasoning, and decision-making.

[0003] In some information recommendation scenarios, modeling is performed through "object-information matching". The semantic description features of the object to be recommended and the semantic description features of the information to be recommended are respectively mined, and then the matching degree between the semantic description features of the object to be recommended and the semantic description features of the information is calculated to perform targeted information recommendation for the object to be recommended.

[0004] Based on the semantic description features, the related technology further introduces the object identity (ID) of the object to be recommended and the information identity (ID) of the information to be recommended to promote matching, and calls a pre-trained language model to complete the "object-information matching" recommendation task. However, on the one hand, there may be some duplicate ID fields in these IDs, resulting in semantic noise. On the other hand, the number of these IDs may reach hundreds of millions or even billions. Directly adding them to the vocabulary of the pre-trained language model will increase the training difficulty of the model, thereby affecting the subsequent semantic matching between the object to be recommended and the information to be recommended, and ultimately reducing the information recommendation effect of the "object-information matching" model. Summary of the Invention

[0005] The embodiments of this application provide an information recommendation method, apparatus, electronic device, and computer-readable storage medium, which can improve the information recommendation effect during the process of information recommendation by combining the object identity of the object to be recommended and the information identity of the information to be recommended.

[0006] The technical solution of the embodiments of this application is implemented as follows:

[0007] The embodiments of this application provide an information recommendation method, and the method includes:

[0008] Obtain a heterogeneous graph including multiple nodes, where each of the nodes is used to represent either an object identifier of an object to be recommended or an information identifier of information to be recommended;

[0009] Determine the identification features of each of the nodes, and perform re-encoding processing on the identification features of each of the nodes to obtain the re-encoded identifier of each of the nodes;

[0010] Perform splitting processing on the re-encoded identifier of each of the nodes to obtain the sub-identifiers of each of the nodes, where each of the nodes corresponds to multiple sub-identifiers;

[0011] Determine the sub-identifiers of the object to be recommended and the sub-identifiers of the information to be recommended from the sub-identifiers respectively corresponding to the multiple nodes;

[0012] Based on the sub-identifiers of the object to be recommended and the sub-identifiers of the information to be recommended, determine the recommendation index of the information to be recommended corresponding to the object to be recommended, and perform a recommendation operation based on the recommendation index.

[0013] An embodiment of the present application provides an information recommendation device, including:

[0014] An acquisition module, configured to acquire a heterogeneous graph including multiple nodes, where each of the nodes is used to represent either an object identifier of an object to be recommended or an information identifier of information to be recommended;

[0015] An encoding module, configured to determine the identification features of each of the nodes, and perform re-encoding processing on the identification features of each of the nodes to obtain the re-encoded identifier of each of the nodes;

[0016] A splitting module, configured to perform splitting processing on the re-encoded identifier of each of the nodes to obtain the sub-identifiers of each of the nodes, where each of the nodes corresponds to multiple sub-identifiers;

[0017] A determination module, configured to determine the sub-identifiers of the object to be recommended and the sub-identifiers of the information to be recommended from the sub-identifiers respectively corresponding to the multiple nodes;

[0018] A recommendation module, configured to determine the recommendation index of the information to be recommended corresponding to the object to be recommended based on the sub-identifiers of the object to be recommended and the sub-identifiers of the information to be recommended, and perform a recommendation operation based on the recommendation index.

[0019] An embodiment of the present application provides an electronic device, including:

[0020] A memory, configured to store computer-executable instructions or computer programs;

[0021] A processor, when executing computer-executable instructions or a computer program stored in the memory, implements the information recommendation method provided by the embodiments of the present application.

[0022] The embodiments of the present application provide a computer-readable storage medium storing computer-executable instructions or a computer program, which is used to implement the information recommendation method provided by the embodiments of the present application when being executed by a processor.

[0023] The embodiments of the present application provide a computer program product, including computer-executable instructions or a computer program, which implements the information recommendation method provided by the embodiments of the present application when the computer-executable instructions or the computer program are executed by a processor.

[0024] The embodiments of the present application have the following beneficial effects:

[0025] During the recommendation process, the object identifier of the object to be recommended and the information identifier of the information to be recommended are used as nodes to construct a heterogeneous graph, and the identifier features of the nodes in the heterogeneous graph are extracted. Then, the identifier features of the nodes are re-encoded to obtain re-encoded identifiers. Thus, by re-encoding the object identifier or the information identifier, it can be ensured that there will no longer be duplicate fields between the identifiers of the object identifier or the information identifier. Next, the obtained re-encoded identifiers are split by fields, which can reduce the field length of the object identifier or the information identifier, so as to reduce the number of fields of the object identifier or the information identifier, and further reduce the vocabulary of the object identifier or the information identifier. Finally, according to the sub-identifiers of the object to be recommended and the sub-identifiers of the information to be recommended obtained by splitting, the recommendation metrics are determined to perform information recommendation. In this way, information recommendation is combined with the object identifier and the information identifier to improve the effect of information recommendation. Description of the Drawings

[0026] Figure 1 is a schematic structural diagram of the information recommendation system architecture provided by the embodiments of the present application;

[0027] Figure 2 is a schematic structural diagram of the electronic device provided by the embodiments of the present application;

[0028] Figure 3A is a first flowchart of the information recommendation method provided by the embodiments of the present application;

[0029] Figure 3B is a second flowchart of the information recommendation method provided by the embodiments of the present application;

[0030] Figure 3C is a third flowchart of the information recommendation method provided by the embodiments of the present application;

[0031] Figure 3D is a fourth flowchart of the information recommendation method provided by the embodiments of the present application;

[0032] Figure 3E is the fifth process schematic diagram of the information recommendation method provided by the embodiments of the present application;

[0033] Figure 3F is the sixth process schematic diagram of the information recommendation method provided by the embodiments of the present application;

[0034] Figure 3G is the seventh process schematic diagram of the information recommendation method provided by the embodiments of the present application;

[0035] Figure 3H is the eighth process schematic diagram of the information recommendation method provided by the embodiments of the present application;

[0036] Figure 3I is the ninth process schematic diagram of the information recommendation method provided by the embodiments of the present application;

[0037] Figure 3J is the tenth process schematic diagram of the information recommendation method provided by the embodiments of the present application;

[0038] Figure 3K is the eleventh process schematic diagram of the information recommendation method provided by the embodiments of the present application;

[0039] Figure 4 is the schematic diagram of user game matching recommendation provided by the embodiments of the present application;

[0040] Figure 5 is the recommended process diagram of multi-game placement provided by the embodiments of the present application;

[0041] Figure 6 is the schematic principle diagram of the user game matching recommendation method provided by the embodiments of the present application. Detailed implementation manners

[0042] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be construed as limiting the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present application.

[0043] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments. However, it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0044] In the following description, the terms "first", "second", and "third" are only used to distinguish similar objects and do not represent a specific order for the objects. Understandably, "first", "second", and "third" can be interchanged in a specific order or sequence when permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0045] In the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one 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 a part of the overall module or unit that includes the function of that module or unit.

[0046] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meaning as commonly understood by those skilled in the art to which the present application belongs. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0047] Before further elaborating on the embodiments of the present application, the nouns and terms involved in the embodiments of the present application are described. The nouns and terms involved in the embodiments of the present application are subject to the following explanations.

[0048] 1) Object identifier, which represents the unique identity identifier (Identity, ID) of the object to be recommended, and is set or randomly assigned by the system where the recommendation scenario is located. For example, in an advertisement recommendation scenario, a user ID is assigned to each user, and this user ID is the unique identity identifier of the user in the advertisement recommendation scenario. Subsequently, all operations of the user in the advertisement recommendation scenario will be recorded based on this user ID, and advertisement recommendations will be made based on this user ID.

[0049] 2) Information identifier, which represents the unique identity identifier of the information to be recommended, and is generally set or randomly assigned by the information database. A large amount of information is stored in the information database, and a corresponding unique identity identifier is assigned to each piece of information. For example, databases such as application databases and game databases assign an application ID or a game ID to the stored application software or games. When information recommendation is required, the corresponding application or game is recommended through ID matching.

[0050] The embodiments of the present application provide an information recommendation method, device, electronic device, computer-readable storage medium, and computer program product, which can improve the effect of information recommendation in the process of combining the object identifier of the object to be recommended and the information identifier of the information to be recommended.

[0051] The following describes exemplary applications of the electronic device provided in the embodiments of the present application. The device provided in the embodiments of the present application can be implemented as various types of user terminals such as laptop computers, tablet computers, desktop computers, set-top boxes, mobile devices (for example, mobile phones, portable music players, personal digital assistants, dedicated messaging devices, portable gaming devices), smart phones, smart speakers, smart watches, smart televisions, and vehicle-mounted terminals, and can also be implemented as servers.

[0052] Taking the server provided in the embodiment of the present application as an example, it can be a server cluster deployed in the cloud, and the server in the cloud is encapsulated with the program of the information recommendation method provided in the embodiment of the present application. The user calls the information recommendation service in the cloud service through the terminal (the terminal runs an APP, such as an instant messaging APP, a reading APP, etc.), so that the server deployed in the cloud calls the encapsulated information recommendation method program. When the user initiates an information recommendation request at the terminal, in response to the information recommendation request initiated by the terminal, the user will be used as the object to be recommended, and then the object identifier of the object to be recommended and any one of the information identifiers of the information to be recommended in the APP will be used as a node to obtain a heterogeneous graph including multiple nodes. For the heterogeneous graph, the identification feature of each node is determined, and the identification feature of each node is re-encoded to obtain the re-encoded identifier of each node. Next, the re-encoded identifier of each node is split to obtain the sub-identifier of each node, and the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended are determined from the sub-identifiers corresponding to multiple nodes. Finally, based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, the recommendation index of the object to be recommended corresponding to the information to be recommended is determined, and the recommendation operation is performed based on the recommendation index to generate the information recommendation result, which is returned to the terminal. The user can then preview the information recommendation result returned by the server in the terminal.

[0053] See also Figure 1 , Figure 1 It is a schematic diagram of the architecture of the information recommendation system 100 provided in an embodiment of the present application, including a terminal 500, a network 300 and a server 200. The terminal 500 is connected to the server 200 via the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0054] In some embodiments, the information recommendation method provided by the embodiments of the present application can be implemented on the terminal 500. When a user initiates an information recommendation request through an application program running on the terminal, the terminal can respond to the user's information recommendation request and directly display the information recommendation result in the application program of the terminal for the user to preview. For example, when the object to be recommended (which can be a user or other artificial intelligence program) initiates an information recommendation request on the terminal 500, the terminal 500 uses either the object identifier of the object to be recommended or the information identifier of the information to be recommended in the application program as a node, and obtains a heterogeneous graph including multiple nodes. For the heterogeneous graph, determine the identification features of each node, and perform re-encoding processing on the identification features of each node to obtain the re-encoded identifier of each node. Next, perform splitting processing on the re-encoded identifier of each node to obtain the sub-identifier of each node, and determine the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended from the sub-identifiers respectively corresponding to multiple nodes. Finally, based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, determine the recommendation index of the information to be recommended corresponding to the object to be recommended, and perform a recommendation operation based on the recommendation index. After generating the information recommendation result, directly display the information recommendation result in the terminal 500 for the object to be recommended to preview.

[0055] In some embodiments, the information recommendation method provided by the embodiments of the present application can be jointly implemented by the terminal 500 and the server 200. For example, various application programs (Application, APP) are running on the terminal 500, such as an instant messaging APP, a reading APP, a video APP, a game APP, or other software programs. When the object to be recommended (which can be a user or other artificial intelligence program) initiates an information recommendation request on the terminal 500 and sends it to the server 200 through the network 300, the server 200 responds to the information recommendation request initiated by the object to be recommended on the terminal 500, and uses either the object identifier of the object to be recommended or the information identifier of the information to be recommended in the application program as a node, and obtains a heterogeneous graph including multiple nodes. For the heterogeneous graph, determine the identification features of each node, and perform re-encoding processing on the identification features of each node to obtain the re-encoded identifier of each node. Next, perform splitting processing on the re-encoded identifier of each node to obtain the sub-identifier of each node, and determine the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended from the sub-identifiers respectively corresponding to multiple nodes. Finally, based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, determine the recommendation index of the information to be recommended corresponding to the object to be recommended, and perform a recommendation operation based on the recommendation index to generate an information recommendation result. The information recommendation result will be returned to the terminal 500 through the network 300, and the object to be recommended (which can be a user or other artificial intelligence program) can view the information recommendation result returned by the server 200 in the application program of the terminal 500.

[0056] In some embodiments, the server 200 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It may also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. The terminal 500 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, a vehicle-mounted terminal, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected through wired or wireless communication methods, which are not limited in the embodiments of the present application.

[0057] See Figure 2 , Figure 2 is a schematic structural diagram of the electronic device 400 provided by the embodiments of the present application. Figure 2 The illustrated electronic device 400 includes: at least one processor 410, a memory 450, and at least one network interface 420. Each component in the terminal 500 is coupled together through a bus system 440. It can be understood that the bus system 440 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clear illustration, in Figure 2 all kinds of buses are labeled as the bus system 440.

[0058] The processor 410 may be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a Digital Signal Processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or any conventional processor, etc.

[0059] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memories, hard disk drives, optical disc drives, etc. The memory 450 optionally includes one or more storage devices that are physically remote from the processor 410.

[0060] The memory 450 includes volatile memory or non-volatile memory, and may also include both volatile and non-volatile memory. The non-volatile memory may be a Read Only Memory (ROM), and the volatile memory may be a Random Access Memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0061] In some embodiments, the memory 450 is capable of storing data to support various operations. Examples of such data include programs, modules, and data structures, or subsets or supersets thereof, which will be exemplarily described below.

[0062] The operating system 451 includes system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, the core library layer, the driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0063] The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420. Exemplary network interfaces 420 include: Bluetooth, Wireless Fidelity (WiFi), and Universal Serial Bus (USB), etc.;

[0064] In some embodiments, the device provided by the embodiments of the present application can be implemented in software. Figure 2 Shown is an information recommendation device 453 stored in the memory 450, which can be software in the form of programs and plugins, etc., including the following software modules: an acquisition module 4531, an encoding module 4532, a splitting module 4533, a determination module 4534, and a recommendation module 4535. These modules are logical, and thus can be arbitrarily combined or further split according to the functions to be implemented. The functions of each module will be described below.

[0065] In some embodiments, the terminal or the server can implement the information recommendation method provided by the embodiments of the present application by running various computer-executable instructions or computer programs. For example, the computer-executable instructions can be commands at the microprogram level, machine instructions, or software instructions. The computer program can be a native program or a software module in the operating system; it can be a native application (APP), that is, a program that needs to be installed in the operating system to run, such as a live broadcast APP or an instant messaging APP; it can also be a small program that can be embedded in any APP, that is, a program that only needs to be downloaded to the browser environment to run. In short, the above-mentioned computer-executable instructions can be instructions in any form, and the above-mentioned computer programs can be application programs, modules, or plugins in any form.

[0066] The information recommendation method provided by the embodiments of the present application will be described in combination with the exemplary applications and implementations of the server provided by the embodiments of the present application.

[0067] See Figure 3A , Figure 3A is a schematic flowchart of the information recommendation method provided by the embodiments of the present application, taking Figure 1The server 200 shown in FIG. 1 is the execution subject, which will be combined with Figure 3A The steps shown are explained.

[0068] In step 101, a heterogeneous graph including a plurality of nodes is obtained.

[0069] Among them, the heterogeneous graph includes nodes for object identifiers of objects to be recommended and nodes for information identifiers representing information to be recommended, that is, a node in the heterogeneous graph is used to represent the object identifier of the object to be recommended, or is used to represent the information identifier of the information to be recommended, that is, each node is used to represent any one of the object identifier of the object to be recommended and the information identifier of the information to be recommended. Incorporating object identifiers and information identifiers into the information recommendation process is mainly to further promote the matching of objects to be recommended and information to be recommended by identifying the association between the object identifier of the object to be recommended and the information identifier of the information to be recommended, so as to infer whether to recommend the information to be recommended to the object to be recommended. Since there can be multiple objects to be recommended, there are also multiple information to be recommended. Therefore, here, obtaining a heterogeneous graph including multiple nodes is to regard the object identifier of the object to be recommended or the information identifier of the information to be recommended as a node to construct a heterogeneous graph including multiple nodes, so that some nodes in the heterogeneous graph are used to represent the object identifier of the object to be recommended, and other nodes are used to represent the information identifier of the information to be recommended. It is hoped that the heterogeneous graph can be used to mine the association between nodes, that is, to mine the association between the object identifier of the object to be recommended and the information identifier of the information to be recommended. In this process, the identification features corresponding to each node are extracted to identify the mined association relationships.

[0070] In step 102, the identification feature of each node is determined, and the identification feature of each node is re-encoded to obtain the re-encoded identification of each node.

[0071] After obtaining a heterogeneous graph including multiple nodes, the identification features of each node can be determined based on the heterogeneous graph, and the identification features of each node can be re-encoded to obtain the re-encoded identification of each node. Because when extracting the identification features of the nodes, it is necessary to mine the associations between the nodes, so the embodiment of the present application collects node sequences in the heterogeneous graph, mines the associations between the nodes according to the context of the nodes in the node sequence, and extracts the identification features of each node. Next, the identification features of each node are classified according to the node representation, and the classified identification features are clustered again. Finally, the node identification features in the cluster cluster are re-encoded to obtain the re-encoded identification of each node, which is explained in detail below.

[0072] In some embodiments, see Figure 3B , Figure 3A"Determining the identification features of each node" in step 102 shown can be implemented through the following steps 1021A to 1022A, which are specifically described below.

[0073] In step 1021A, path collection is performed on the heterogeneous graph to obtain a node sequence of the heterogeneous graph.

[0074] To obtain the node sequence in the heterogeneous graph, path collection needs to be performed on the heterogeneous graph, and the nodes obtained during the path collection process are used as the node sequence. Since for each object to be recommended, any information to be recommended can be recommended, there are many paths between nodes in the heterogeneous graph, and different paths selected will result in different node sequences. Therefore, by performing different path collections on the heterogeneous graph, multiple different node sequences can be correspondingly obtained.

[0075] In some embodiments, referring to Figure 3C , Figure 3B step 1021A shown can be implemented through the following steps 10211A to 10213A, which are specifically described below.

[0076] In step 10211A, an arbitrary node is selected from the heterogeneous graph as the starting wandering node.

[0077] In some embodiments, path collection on the heterogeneous graph can be implemented by node wandering. First, a node needs to be selected from the heterogeneous graph as the starting wandering node. Here, the starting wandering node can ignore the node representation and can be arbitrarily selected in the heterogeneous graph.

[0078] In step 10212A, starting from the starting wandering node, random wandering processing is performed along the edges in the heterogeneous graph to obtain multiple wandering nodes in sequence.

[0079] Continuing the above embodiments, in the heterogeneous graph, there are also edges for connecting two nodes, and during the path collection process, the edges of the nodes can be used as paths. Then, starting from the starting wandering node, random wandering processing is performed along the edges in the heterogeneous graph to obtain multiple wandering nodes in sequence, where the number of wandering nodes can be preset. For example, if the number is preset to 2, then starting from the starting wandering node, only two random wandering processes need to be performed along the edges in the heterogeneous graph to obtain 2 wandering nodes.

[0080] In step 10213A, the starting wandering node and the multiple wandering nodes are sorted according to the order of wandering to obtain a node sequence.

[0081] After the node walk ends, the starting walk node and multiple walk nodes can be sorted according to the order of walking to obtain a node sequence. Since the selected starting walk node is random and the walk process is also random, multiple different node sequences can be obtained here.

[0082] In a heterogeneous graph, since two nodes connected by an edge may have different representations, that is, one node represents the object identifier of the object to be recommended, and the other node represents the information identifier of the information to be recommended. Therefore, when performing a node walk, the starting walk node and the first walk node may have different representations, and the first walk node and the second walk node may also have different representations. It can be inferred from this that the representations of the starting walk node and the second walk node are the same, and this cycle continues when continuing the walk. Therefore, in some embodiments, the number of walk nodes is generally preset to 2, that is, two random walks are performed along the edges in the heterogeneous graph to obtain 2 walk nodes, and finally the starting walk node and the 2 walk nodes are sorted according to the order of walking to obtain a node sequence including 3 nodes.

[0083] Continue to refer to Figure 3B , in step 1022A, based on the node sequence of the heterogeneous graph, determine the identification feature of each node.

[0084] After obtaining multiple node sequences of the heterogeneous graph, based on the node sequence of the heterogeneous graph, determine the identification feature of each node. Among them, the identification feature of a node can be the word vector representation of the node, and the method for determining the feature can adopt the Word2Vec word vector algorithm, that is, by predicting the adjacent nodes of the target node, the correlation relationship between nodes is mined and the identification feature of the node is extracted, which will be specifically described below.

[0085] In some embodiments, refer to Figure 3D , Figure 3B The step 1022A shown can be implemented by the following steps 10221A to 10222A, which will be specifically described below.

[0086] In step 10221A, from the node sequence of the heterogeneous graph, obtain a target node sequence including the target node.

[0087] Since it is necessary to determine the identification features of each node in the heterogeneous graph, first, from the node sequence of the heterogeneous graph, a target node sequence including the target node is obtained, where the target node is any node in the heterogeneous graph. Because there are multiple node sequences obtained by path collection, and each node sequence may only include some nodes in the heterogeneous graph. Here, for each target node in the heterogeneous graph, a target node sequence including the target node is determined from multiple node sequences, and then the target node sequence is processed to obtain the identification features of the target node, thereby obtaining the identification features of each node in the heterogeneous graph

[0088] In step 10222A, based on the adjacent nodes of the target node in the target node sequence, the initial features of the target node are updated to obtain the identification features of the target node

[0089] Considering that mining the correlation relationship between nodes can be regarded as learning the context relationship between nodes and adjacent nodes in the node sequence. Therefore, in the embodiment of the present application, the initial features of a target node are first determined here. The initial features of the target node are initialization features, which can be obtained by directly encoding the target node through one-hot encoding, or can be randomly set, such as a zero vector. Next, the word vector model is used to predict the adjacent nodes of the target node in the target node sequence through the initial features of the target node, and then based on the adjacent nodes, the initial features of the target node are updated to obtain the identification features of the target node. Among them, the word vector model can be the Word2Vec word vector model. The following specifically describes the specific process of the update process

[0090] In some embodiments, refer to Figure 3E , Figure 3D The step 10222A shown can be implemented through the following steps 102221A to step 102223A. The following specifically describes

[0091] In step 102221A, based on the initial features of the target node in the target node sequence, adjacent node prediction processing is performed on the target node to obtain the predicted distribution probability of the adjacent nodes of the target node

[0092] After first determining the target node sequence including the target node, adjacent node prediction processing is performed on the target node based on the initial features of the target node in the target node sequence to obtain the predicted probability of the adjacent nodes of the target node. Here, although the adjacent nodes of the target node are known and marked with real labels, that is, the marking of the distribution probability of the adjacent nodes. However, it is still necessary to predict its adjacent nodes through the initial features of the target node to obtain the predicted distribution probability of the adjacent nodes

[0093] In the target node sequence, it is necessary to determine K adjacent nodes of the target node, that is, to determine the first K / 2 nodes and the last K / 2 nodes in front of the target node. When the target node is the first node or the last node of the target node sequence, only the last K / 2 nodes or the first K / 2 nodes of the target node are determined. When the target node sequence only includes three nodes and the target node is the second node, K is 2.

[0094] After determining the K adjacent nodes of the target node in the target node sequence, first obtain the initial features of the target node through one-hot encoding, and then input them into the word vector model. There are two weight matrices in the hidden layer of the word vector model. First, perform a linear transformation on the initial features of the target node through the first weight matrix to obtain the updated initial features of the target node. Then, input the updated initial features into the second weight matrix for linear transformation, and map the result of the linear transformation through an activation function to obtain the predicted distribution probability of the adjacent nodes of the target node. This predicted distribution probability includes the predicted probability of each adjacent node.

[0095] In step 102222A, determine the error between the predicted distribution probability and the distribution probability label.

[0096] Since in the target node sequence, the adjacent nodes of the target node all have true labels, that is, distribution probability labels, where the distribution probability label is the label of the distribution probability of the adjacent nodes of the target node. After predicting the predicted distribution probability of the adjacent nodes of the target node, the predicted distribution probability can be compared with the distribution probability label to determine the error between the predicted distribution probability and the distribution probability label, so as to verify whether the prediction of each adjacent node by the word vector model is accurate. Among them, the method of calculating the error can be the difference between the distribution probability label and the predicted distribution probability, that is, the sum of the differences between the predicted probability of each adjacent node and the corresponding true label (the label of the distribution probability).

[0097] In step 102223A, update the initial features of the target node until the error converges, and use the updated initial features of the target node when the error converges as the identification features of the target node.

[0098] After determining the error between the predicted distribution probability and the distribution probability label, then backpropagate the error to the word vector model, update the initial features of the target node until the error converges, and use the updated initial features of the target node when the error converges as the identification features of the target node.

[0099] Specifically, after determining the error between the predicted distribution probability and the distribution probability label, the error is then backpropagated into the hidden layer of the word vector model to update the two weight matrices and related parameters in the hidden layer. Next, the updated initial features of the target node are continuously input into the updated word vector hidden layer for further update, and the error is predicted and calculated again, and the error is continuously backpropagated to update the two weight matrices of the hidden layer, so as to continuously update the initial features of the target node. When the error is not in the minimized convergence state, the initial features of the target node are updated by continuously updating the weight matrix of the word vector hidden layer to obtain the updated initial features until the error converges.

[0100] During the training process, when the error is not in the minimized convergence state, the predicted distribution probability for predicting the adjacent nodes of the target node is constantly changing and will get closer and closer to the distribution probability label. And the error between the predicted probability and the labeled probability will also change and gradually become smaller until it converges. During this process, the initial vector of the target node is also continuously updated.

[0101] After a certain round of update, if the error is in the minimized convergence state, the initial features updated by the target node when the error converges are used as the identification features of the target node. When the error converges, the predicted distribution probability is the highest, even equal to the distribution probability label, indicating that the error between the predicted probability and the labeled probability reaches the minimum at this time, and the prediction accuracy of the adjacent nodes of the target node is the highest. Due to the convergence of the error, the weight matrix of the hidden layer of the word vector model no longer updates, and the initial features updated by the target node obtained at this time are also fixed. Then, the initial features updated by the target node at this time are used as the identification features of the target node. This identification feature is obtained by training the adjacent nodes of the target node and contains the context relationship between the target node and its adjacent nodes. Thus, both the identification features of the target node are determined and the association relationship between the nodes is mined.

[0102] In some embodiments, refer to Figure 3F , Figure 3A As shown, the "re-encoding process for the identification features of each node to obtain the re-encoded identification of each node" in step 102 can be implemented through the following steps 1021B to 1023B. The following is a specific description.

[0103] In step 1021B, the identification features of the node are classified to obtain an object identification feature set and an information identification feature set.

[0104] Since the nodes of the heterogeneous graph have different representations, that is, some are used to represent the object identifier of the object to be recommended, and some are used to represent the information identifier of the information to be recommended. After determining the identification features of each node, the following processing is performed for each node. First, according to the representation of the node, the identification features of the node are classified to obtain an object identifier feature set and an information identifier feature set. Among them, the object identifier feature set includes the identification features of at least one object to be recommended, and the information identifier feature set includes the identification features of at least one piece of information to be recommended.

[0105] Here, the reason for classifying the identification features of the nodes is that the object identifier of the object to be recommended and the information identifier of the information to be recommended both belong to the field sequence, and both are composed of multiple fields spliced together. In order not to confuse the recoded identifiers obtained after recoding the two, it is necessary to classify the identification features of the nodes to separate the object identifier features of the object to be recommended and the information identifier features of the information to be recommended, and then perform subsequent recoding processing separately.

[0106] In step 1022B, the identification features included in the object identifier feature set are clustered to obtain an object identifier cluster, and the identification features included in the information identifier feature set are clustered to obtain an information identifier cluster.

[0107] After classifying the identification features of the nodes, the identification features included in the object identifier feature set are clustered to obtain an object identifier cluster, and the identification features included in the information identifier feature set are clustered to obtain an information identifier cluster. Considering that the object identifier of the object to be recommended is generally randomly assigned by the system where the information recommendation scenario is located, and the information identifier of the information to be recommended is also randomly assigned in the information database, there may be some duplicate identifier fields between different object identifiers (or information identifiers), that is, there is shared information. Semantic noise will be generated during semantic matching, affecting subsequent information matching and recommendation. If the identification features of each node are recoded separately, the obtained recoded identifiers may still have some duplicate identifier fields, and the purpose of removing the shared information between the identifiers cannot be achieved.

[0108] Therefore, after classifying the identification features of nodes in the embodiments of the present application, for the obtained object identification feature set and information identification feature set, clustering processing is performed on the identification features included in the object identification feature set to obtain object identification clusters. Through clustering, the identification features corresponding to object identifications that may have duplicate fields can be aggregated first to obtain object identification clusters, so as to uniformly perform re-encoding processing on the identification features corresponding to the identifications that may have duplicate fields within the clusters, avoiding the problem that separate re-encoding may still result in duplicate fields. Similarly, clustering processing is also performed on the identification features included in the information identification feature set here to obtain information identification clusters. Among them, the clustering processing can be implemented using clustering algorithms, for example, the K-means algorithm, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, etc.

[0109] In step 1023B, re-encoding processing is performed on the identification features of the nodes in each target cluster to obtain the re-encoded identifications of the nodes in the target cluster.

[0110] After clustering is completed, re-encoding processing can be performed on the identification features of the nodes in each target cluster to obtain the re-encoded identifications of the nodes in the target cluster, where the target cluster is an object identification cluster or an information identification cluster. Here, the object identification cluster and the information identification cluster are re-encoded separately, and they do not affect each other, and there is no order between the re-encoding processes. The following is a specific description.

[0111] In some embodiments, referring to Figure 3G , Figure 3F the step 1023B shown can be implemented by the following steps 10231B to 10233B. The following is a specific description.

[0112] In step 10231B, the identification feature of the node at the center of the target cluster is determined as the central identification feature, and re-encoding processing is performed on the central identification feature to obtain the re-encoded identification of the node at the center of the target cluster.

[0113] In some embodiments, since each target clustering cluster includes the identification features of multiple nodes, there is a sequence when performing re-encoding. Therefore, the following processing is performed for each target clustering cluster. The identification feature of the node at the center of the target clustering cluster is determined as the central identification feature, and the central identification feature is re-encoded to obtain the re-encoded identification of the node at the center of the target clustering cluster. Here, the central identification feature of the center of the target clustering cluster is determined first because the clustering cluster is formed based on this central clustering, and the central identification feature has similarity with other identification features in the target clustering cluster, that is, the identifications represented by the corresponding nodes may all have duplicate fields. Therefore, the central encoding feature of the center of the target clustering side is selected first and re-encoded to obtain the re-encoded identification of the node at the center of the target clustering cluster. Among them, other identification features are the identification features of the nodes other than the node at the center in the target clustering cluster. The process of re-encoding can be to call a fully connected neural network layer to perform mapping processing on the identification feature to predict the corresponding re-encoded identification. The re-encoded identification, like the object identification or information identification, is essentially also a field sequence, composed of multiple fields spliced together, but the length of the field sequence is the same as the original object identification or information identification.

[0114] In step 10232B, determine the distance between the other identification feature and the central identification feature.

[0115] Continuing with the above embodiments, after re-encoding the central identification feature in the target clustering cluster, considering that among the other identification features in the target clustering cluster, the closer the distance to the central identification feature, the more similar it is to the central identification feature, and the more likely the identification represented by the corresponding node has duplicate fields. Therefore, in the embodiments of the present application, first determine the distance between the other identification feature and the central identification feature, and then determine the re-encoding order of the other identification features in the target clustering cluster according to the distance. The method of calculating the distance can be to directly calculate the relative distance between each point in the cluster and the center point, that is, for each node's identification feature in the other identification features, the relative distance from it to the central identification feature can be calculated.

[0116] In step 10233B, sort the other identification features according to the distance, and sequentially re-encode the other identification features according to the sorted order to obtain the re-encoded identifications of the nodes other than the node at the center in the target clustering cluster.

[0117] After determining the distances between other identification features and the central identification feature for each target distance cluster, the other identification features are then sorted according to the distances, and the other identification features are sequentially re-encoded in the sorted order to obtain the re-encoded identifiers of the nodes other than the central node in the target clustering cluster. Here, the other identification features with shorter distances are sorted in the front and re-encoded first, while the other identification features with longer distances are sorted in the back and re-encoded later. That is, in the target distance cluster, starting from the central identification feature, the identification features of each node are sorted from the inside out, and the identification features of each node are sequentially re-encoded in the sorted order to obtain the re-encoded identifiers of the nodes other than the central node in the target clustering cluster.

[0118] Through step 102, for the object identifier of the object to be recommended and the information identifier of the information to be recommended represented by the nodes in the heterogeneous graph, the association relationship between the nodes is first mined through the node sequence, that is, the identification features of the nodes are extracted, and then the corresponding identification feature sets are obtained through the method of separate clustering. Then, the identification features of each node in the identification feature set are re-encoded, which can greatly reduce the possibility of duplicate fields between the obtained re-encoded identifiers, and eliminate the duplicate fields and shared information between the identifiers as much as possible. On the basis of retaining the association relationship between the identifiers, the influence of the noise of the shared information on the final recommendation effect is reduced.

[0119] Continue to refer to Figure 3A , in step 103, the re-encoded identifier of each node is split to obtain the sub-identifier of each node.

[0120] After obtaining the re-encoded identifier of each node, considering that the length of the field sequence of the re-encoded identifier is the same as the original object identifier and information identifier, and the sequence length is still very large, the number of word tables may reach the level of hundreds of millions. Therefore, in the embodiments of the present application, the re-encoded identifier of each node is split to obtain the sub-identifier of each node, where each node corresponds to multiple sub-identifiers. After splitting, the re-encoded identifiers at the level of hundreds of millions can be reduced to the level of ten thousands or even thousands, thereby reducing the number of identifiers to reduce the number of word tables. The specific process of splitting is described below.

[0121] In some embodiments, refer to Figure 3H , Figure 3A As shown in, step 103 can be implemented by the following steps 1031 to 1033, which are specifically described below.

[0122] In step 1031, an identifier sequence composed of the re-encoded identifiers of the nodes is obtained.

[0123] Here, first, the following processing is performed on the recoding identifier of each node to obtain an identifier sequence composed of the recoding identifiers of the nodes. Since the fields in the sequence have a sequential order, and the recoding identifier is a sequence of fields composed of multiple fields, it is necessary to ensure that the order of each field cannot be disrupted during the splitting process. Therefore, the recoding identifier of each node is regarded as an identifier sequence to ensure that the sequential order of the fields is not disrupted when splitting the identifier sequence corresponding to each node.

[0124] In step 1032, determine the splitting length of the identifier sequence.

[0125] Next, determine the splitting length of the identifier sequence, that is, the unit length of splitting. Here, it can be determined according to the total length of the identifier sequence. Since the object identifier and the information identifier are both assigned by the system or the database, there is generally a fixed length between object identifiers and between information identifiers. Therefore, the total length of the identifier sequence is also fixed. Here, the splitting length can be reasonably preset according to actual needs so that the lengths of the obtained sub-identifiers after splitting are as average as possible. Among them, when splitting, different splitting lengths can be preset for the identifier sequence according to different node representations, or the node representation can be ignored and a unified splitting length can be preset.

[0126] In step 1033, based on the splitting length, perform a splitting process on the identifier sequence to obtain the sub-identifiers of the node.

[0127] After determining the splitting length of the identifier sequence, the identifier sequence can be split based on the splitting length to obtain the sub-identifiers of the node. Among them, the method of the splitting process is to call certain word segmentation algorithms or word segmentation tools, such as the "jieba" word segmentation library, the "Sentence Piece" tokenizer, etc. According to the preset splitting length, then directly split the field sequence using the word segmentation algorithm or word segmentation tool to obtain multiple sub-sequences. Here, each sub-sequence is the sub-identifier of the node.

[0128] For example, the object identifier of the object to be recommended is composed of 16 characters (numbers or letters) concatenated together, and the total character length is 16. Therefore, the number of object identifiers can reach quadrillions. And the information identifier of the information to be recommended is composed of 12 characters (numbers or letters) concatenated together, and the total character length is 12. Therefore, the number of information identifiers also reaches hundreds of billions. When the node represents the object identifier, the length of the identifier sequence corresponding to the recoding identifier of the node is 16. When the node represents the information identifier, the length of the identifier sequence corresponding to the recoding identifier of the node is 12. So here, the splitting length of the identifier sequence can be uniformly determined as 4, or for the object identifier, the splitting length can be determined as 4, and for the information identifier, the splitting length can be determined as 3.

[0129] Next, an example will be given with the segmentation length of the identification sequence uniformly determined as 4. After segmentation, the re-encoded identification of the node corresponding to the object identification is split into 4 sub-identifications, with the segmentation length of each sub-identification being 4, reducing the number of object identifications from quadrillions to ten thousands. And the re-encoded identification of the node corresponding to the object identification is split into 3 sub-identifications, with the segmentation length of each sub-identification also being 4, reducing the number of object identifications from hundreds of billions to ten thousands. In this way, the number of the vocabulary of the identifications is greatly reduced, and only ten thousand words are needed to cover all object identifications and information identifications.

[0130] Through the above step 103, the re-encoded identification of each node is split to obtain multiple sub-identifications of each node. Thus, by splitting the identification field, the field length of the object identification or information identification represented by the node can be reduced to decrease the number of fields of the object identification or information identification, and further reduce the vocabulary of the object identification or information identification. This makes it easier for the pre-trained language model for information recommendation to understand and recognize the object identification and information identification, reducing the inference difficulty of the pre-trained language model.

[0131] In step 104, the sub-identifications of the object to be recommended and the sub-identifications of the information to be recommended are determined from the sub-identifications respectively corresponding to multiple nodes.

[0132] In some embodiments, after determining the sub-identifications of each node, the sub-identifications of the object to be recommended and the sub-identifications of the information to be recommended are determined from the sub-identifications respectively corresponding to multiple nodes. Since the nodes of the heterogeneous graph represent either the object identification of the object to be recommended or the information identification of the information to be recommended, according to the representation of the nodes, the sub-identifications corresponding to multiple nodes can be classified, and the sub-identifications of the object to be recommended and the sub-identifications of the information to be recommended can be determined therefrom.

[0133] Specifically, since each node in the heterogeneous graph is used to represent the object identifier of the object to be recommended or the information identifier of the information to be recommended, that is, some nodes in the heterogeneous graph are used to represent the object identifier of the object to be recommended, while some other nodes are used to represent the information identifier of the information to be recommended. Therefore, when a node is used to represent the object identifier of the object to be recommended, after the node representing the object identifier of the object to be recommended is processed through steps 102-103, the sub-identifier of the node is the sub-identifier of the object to be recommended; when a node is used to represent the information identifier of the information to be recommended, after the node representing the information identifier of the information to be recommended is processed through steps 102-103, the sub-identifier of the node is the sub-identifier of the information to be recommended, that is, the sub-identifiers corresponding to multiple nodes include the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended. Thus, for multiple nodes in the heterogeneous graph, classification can be performed according to the representation of each node, and the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended can be classified from the sub-identifiers corresponding to multiple nodes.

[0134] In step 105, based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, determine the recommendation index of the information to be recommended corresponding to the object to be recommended, and perform a recommendation operation based on the recommendation index.

[0135] After determining the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, the two can be used in the information recommendation process. Based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, determine the recommendation index of the information to be recommended corresponding to the object to be recommended, and perform a recommendation operation based on the recommendation index. Here, first, by combining the specific description information of the object to be recommended and the information to be recommended, construct an intelligent prompt message, and then call a language model to perform inference and prediction on the intelligent prompt message, calculate the recommendation index of the information to be recommended corresponding to the object to be recommended, and finally use the recommendation index to perform the recommendation operation.

[0136] In some embodiments, refer to Figure 3I , Figure 3A As shown in, the step of "based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, determine the recommendation index of the information to be recommended corresponding to the object to be recommended" in step 105 can be implemented through the following steps 1051 to 1053, and the following is a specific description.

[0137] In step 1051, obtain the description information of the object to be recommended and the description information of the information to be recommended.

[0138] Considering that the main process of information recommendation essentially still uses the relevant semantics of the object to be recommended and the information to be recommended for matching and recommendation. When constructing intelligent prompt information here, it is necessary to obtain the description information of the object to be recommended and the description information of the information to be recommended. Among them, the description information of the object to be recommended can be basic information such as age, gender, region, hobbies, etc., while the description information of the information to be recommended can be the classification to which the information belongs, content summary, comments, etc.

[0139] In step 1052, based on the description information of the object to be recommended, the description information of the information to be recommended, the sub-identifier of the object to be recommended, and the sub-identifier of the information to be recommended, at least one intelligent prompt information is constructed.

[0140] After obtaining the description information of the object to be recommended and the description information of the information to be recommended, then based on the description information of the object to be recommended, the description information of the information to be recommended, the sub-identifier of the object to be recommended, and the sub-identifier of the information to be recommended, at least one intelligent prompt information is constructed. Among them, the construction method can be implemented using some language generation models, such as the language models of the Generative Pre-Trained Transformer (GPT) series. These language models can generate various specified information text templates according to the input prompt information. By inputting the description information of the object to be recommended, the description information of the information to be recommended, the sub-identifier of the object to be recommended, and the sub-identifier of the information to be recommended into the generative pre-trained transformer, intelligent prompt information of various text templates can be generated.

[0141] For example, the object identifier of a certain object to be recommended is 1025, and the description information is "male, 26 years old, city B, hobby is games", and the information identifier of a certain information to be recommended is 2309, and the description information is "games, action role-playing, good reviews". Inputting these information into the generative pre-trained transformer for generation, the obtained intelligent prompt information can be: Will the object to be recommended with the object identifier 1025 and the description of "male, 26 years old, city B, hobby is games" click on the game with the information identifier 2309 and the description of "games, action role-playing, good reviews"? In addition, the generated intelligent prompt information can also be: Will the object to be recommended with the object identifier 1025 and the description of "male, 26 years old, city B, hobby is games" like the game with the information identifier 2309 and the description of "games, action role-playing, good reviews"?

[0142] In step 1053, the intelligent prompt information is processed by reasoning to obtain the recommendation index of the information to be recommended corresponding to the object to be recommended.

[0143] After obtaining at least one intelligent prompt message, the pre-trained language model can be called to perform inference processing on the intelligent prompt message to obtain the recommendation metrics of the recommended object corresponding to the information to be recommended. This process is to match the object to be recommended and the information to be recommended in the intelligent prompt message through the pre-trained language model, and finally a prediction probability will be obtained. This prediction probability is the recommendation metric of the information to be recommended corresponding to the object to be recommended. The following is a specific description.

[0144] In some embodiments, referring to Figure 3J , Figure 3I As shown, step 1053 can be implemented by the following steps 10531 to 10534. The following is a specific description.

[0145] In step 10531, determine the word vector of the intelligent prompt message, the position encoding vector corresponding to the word vector, and the whole-word marking vector in the intelligent prompt message.

[0146] It should be noted that the inference processing process of the pre-trained language model is divided into an encoding part and a decoding part. The encoding part is used to encode the intelligent prompt message to obtain the semantic features of the intelligent prompt message, and the decoding part decodes the semantic features of the intelligent prompt message to obtain the final prediction probability, that is, the matching probability between the object to be recommended and the information to be recommended in the intelligent prompt message, as the recommendation metric.

[0147] Here, after inputting the intelligent recommendation message into the pre-trained language model, first perform encoding processing on the intelligent recommendation message, but before the encoding processing, first perform embedding processing on the intelligent prompt message through the embedding layer in the pre-trained model to determine the word vector of the intelligent prompt message, the position encoding vector corresponding to the word vector, and the whole-word marking vector in the intelligent prompt message. For the word vector, each word segment in the intelligent prompt message can be subjected to word embedding processing through the embedding layer to obtain the word vector of each word segment. For the position encoding vector corresponding to the word vector, by determining the position of each word segment in the intelligent prompt message and constructing it according to the position order, the position encoding vector corresponding to each word segment can be obtained.

[0148] Since the intelligent prompt message includes sub-identifiers of the object to be recommended or sub-identifiers of the information to be recommended, and these sub-identifiers are all split from the re-encoding identifier, in order not to cause confusion between the sub-identifiers belonging to the same re-encoding identifier and the sub-identifiers belonging to other re-encoding identifiers, in the embodiments of the present application, a whole-word mark is added to each word segment in the intelligent prompt message, and a whole-word marking vector is constructed based on the whole-word mark. The following is a specific description.

[0149] In some embodiments, referring to Figure 3K , Figure 3J"Determining the whole-word token vector in the intelligent prompt information" in step 10531 shown can be implemented through the following steps 105311 to 105314, which are specifically described below.

[0150] In step 105311, the intelligent prompt information is segmented to obtain multiple segments.

[0151] In some embodiments, first, the intelligent prompt information needs to be segmented to obtain multiple segments. Here, the segmentation standard can be based on splitting by each character, including punctuation marks which are also regarded as a character, and each split character can be used as a segment. For the sub-identifiers of the object to be recommended or the sub-identifiers of the information to be recommended included in the intelligent prompt information, since these sub-identifiers are obtained by splitting the re-encoded identifiers, there is no need to segment them here, and each sub-identifier can be directly used as a segment.

[0152] In step 105312, the segments belonging to the same re-encoded identifier are assigned the same whole-word token, where different re-encoded identifiers correspond to different whole-word tokens.

[0153] Continuing with the above embodiment, after obtaining multiple segments after segmenting the intelligent prompt information, considering that each segment in the intelligent prompt information is in a different position and may also have different semantics, for the multiple segments of the intelligent prompt information, different whole-word tokens are assigned to each segment.

[0154] When the segment is a sub-identifier of the object to be recommended or the information to be recommended, considering that these sub-identifiers are all split from the re-encoded identifier, in order not to cause confusion between the sub-identifiers belonging to the same re-encoded identifier and the sub-identifiers belonging to other re-encoded identifiers, the segments belonging to the same re-encoded identifier are assigned the same whole-word token here, where different re-encoded identifiers correspond to different whole-word tokens. That is, the whole-word tokens of the sub-identifiers belonging to the same re-encoded identifier are the same, and the whole-word tokens of different re-encoded identifiers are different from each other.

[0155] Exemplarily, a part of an intelligent prompt message is: Whether object 102510131089 clicks on message 23082309? First, perform word segmentation on the intelligent prompt message to obtain 14 word segments, namely: "to, object, 1025, 1013, 1089, yes, no, click, information, 2308, 2309,?". First, assign different integer word tags to each word segment. Among them, the three word segments "1025, 1013, 1089" are sub-identifiers of the object to be recommended and belong to the same re-encoding identifier "102510131089", so they are assigned the same integer word tag. The two word segments "2308, 2309" are also sub-identifiers of the information to be recommended and belong to the same re-encoding identifier "23082309", so they are also assigned the same integer word tag. Finally, the integer word tags corresponding to this part of the intelligent prompt message are divided into "W1, W2, W3, W3, W3, W4, W5, W6, W7, W8, W9, W 10 , W 10 , W 11 . That is, the three word segments "1025, 1013, 1089" are assigned the same integer word tag W3, and the two word segments "2308, 2309" are assigned the same integer word tag W 10 , while each of the other word segments is assigned a different integer word tag, and different re-encoding identifiers correspond to different integer word tags (i.e., W3 and W 10 are different).

[0156] In step 105313, perform embedding processing on the integer word tags assigned to each word segment to obtain the integer word tag vectors of the word segments.

[0157] After assigning integer word tags to each word segment in the intelligent prompt message, then construct corresponding integer word tag vectors based on the integer word tags of each word segment. Here, the embedding layer of the pre-trained language model is used to perform embedding processing on the integer word tags assigned to each word segment to obtain the integer word tag vectors of the word segments.

[0158] In step 105314, merge the integer word tag vectors of each word segment to obtain the integer word tag vector of the intelligent prompt message.

[0159] After determining the integer word tag vectors of each word segment, here the integer word tag vectors of each word segment are also corresponding to the word vectors and position encoding vectors of the word segments. Then, merge the integer word tag vectors of each word segment to obtain the integer word tag vector of the intelligent prompt message. Among them, the merging method can be to directly perform physical splicing on the integer word tag vectors of each word segment to obtain the integer word tag vector of the intelligent prompt message.

[0160] Continue to refer to Figure 3J , in step 10532, merge the word vectors, position encoding vectors, and integer word tag vectors to obtain the fused feature vector of the intelligent prompt message.

[0161] After separately determining the word vector, position encoding vector, and whole-word marking vector of the intelligent prompt information, the word vector, position encoding vector, and whole-word marking vector are then combined to obtain the fused feature vector of the intelligent prompt information, which is used as the input to the pre-trained language model. Among them, the method of combination here is also to physically splice the corresponding multiple vectors.

[0162] In step 10533, the fused feature vector is encoded to obtain the semantic feature vector of the intelligent prompt information.

[0163] After obtaining the fused feature vector of the intelligent prompt information, the fused feature vector is then input into the encoder input layer of the encoding part of the pre-trained language model. The encoder encodes the fused feature vector to obtain the semantic feature vector of the intelligent prompt information. During the encoding process, the pre-trained language model is used to understand the semantics of the intelligent prompt information to determine the matching degree between the object to be recommended and the information to be recommended therein.

[0164] In step 10534, the semantic feature vector is decoded to obtain the recommendation index of the object to be recommended corresponding to the information to be recommended.

[0165] Obtaining the semantic feature vector of the intelligent prompt information through the encoder of the encoding part indicates that the pre-trained language model has understood the semantics of the intelligent prompt information. Then, the semantic feature vector is input into the decoder of the decoding part of the pre-trained language model. The decoder decodes the semantic feature vector to obtain the recommendation index of the object to be recommended corresponding to the information to be recommended. Here, the decoder can generate a prediction probability as the matching probability between the object to be recommended and the information to be recommended according to the semantics understood by the encoder, and finally outputs this matching probability through the activation function in the decoder as the recommendation index of the object to be recommended corresponding to the information to be recommended. In this way, for each intelligent prompt information, the pre-trained language model can be used for inference processing to obtain the recommendation index of the object to be recommended corresponding to the information to be recommended.

[0166] For example, a certain intelligent prompt message is: For a to-be-recommended object with an object identifier of 1025 and a description of "male, 26 years old, city B, hobby is games", will it click on a game with an information identifier of 2309 and a description of "games, action role-playing, good reviews"? Or, for a to-be-recommended object with an object identifier of 1025 and a description of "male, 26 years old, city B, hobby is games", will it like a game with an information identifier of 2309 and a description of "games, action role-playing, good reviews"? These two different intelligent recommendation messages are input into a pre-trained language model for inference processing, and finally the recommended probability (i.e., the recommendation metric) obtained by the decoder is 95%, that is, the recommendation metric of the to-be-recommended information 2309 corresponding to the to-be-recommended object 1025 is 95%, indicating that there is a 95% probability that the object 1025 will click on the information 2309, or there is a 95% probability that the object 1025 will like the information 2309.

[0167] In some embodiments, after determining the recommendation metric of the to-be-recommended information corresponding to the to-be-recommended object, for each to-be-recommended object, based on the recommendation metric of the to-be-recommended information corresponding to the to-be-recommended object, some to-be-recommended information can be screened out from the to-be-recommended information; and a recommendation operation is performed on the to-be-recommended object based on the partial to-be-recommended information.

[0168] Specifically, when there are multiple to-be-recommended objects in an information recommendation scenario, after determining the recommendation metric of the to-be-recommended information corresponding to the to-be-recommended object, for each to-be-recommended object, the multiple to-be-recommended information can be sorted in descending order by the recommendation metric, that is, the to-be-recommended information with the highest recommendation metric is sorted at the front of the sequence. Then, some to-be-recommended information at the head of the sequence is selected to perform a recommendation operation on the current to-be-recommended object. In this way, an information recommendation operation can be performed separately for each to-be-recommended object.

[0169] Through the embodiments of the present application, for the object identifier of the object to be recommended and the information identifier of the information to be recommended represented by the nodes in the heterogeneous graph, the association relationship between the nodes is mined through the node sequence, and the identifier features of the nodes are also extracted. Then, the corresponding identifier feature sets are obtained through the method of clustering respectively, and the identifier features of each node in the identifier feature set are re-encoded, which can greatly reduce the possibility of duplicate fields existing between the obtained re-encoded identifiers, and eliminate as much as possible the shared information with duplicate fields existing between the identifiers. On the basis of retaining the association relationship between the identifiers, the influence of the noise of the shared information on the final recommendation effect is reduced. Next, the re-encoded identifier of each node is split to obtain multiple sub-identifiers of each node. By splitting the identifier field in this way, the field length of the object identifier or information identifier represented by the node can be shortened, so as to reduce the number of fields of the object identifier or information identifier, and further reduce the vocabulary of the object identifier or information identifier. Finally, based on multiple sub-identifiers, intelligent prompt information is constructed, so that the pre-trained language model for information recommendation can more easily understand and recognize the object identifier and information identifier in the intelligent prompt information, and infer the recommendation metrics for information recommendation, reducing the inference difficulty of the pre-trained language model and improving the effect of information recommendation.

[0170] Next, an exemplary application of the embodiments of the present application in an actual application scenario will be described.

[0171] As Figure 4 shown, in some advertising placement tasks, the similarity score between the user description and the game description is calculated as the matching score between the user and the game, and then the multiple games are sorted in descending order according to the similarity score in the user dimension, so as to select the most suitable game for the user to recommend. In actual applications, both users and games have a large amount of natural language description information, such as the statistical portrait description of users, the historical game behavior description, the basic gameplay description of games, the game introduction, etc. Using a language model to deeply understand the user description information and text description information and mine the potential semantic associations is crucial for improving the quality of game recommendations and user satisfaction.

[0172] In the related art, a multi-layer perceptron (MLP) or a pre-trained language model (PLM) is used to estimate the "user-game" matching score. First, the semantic representations of the user description information and the game description information are extracted respectively, and then the similarity measure between the representations is calculated to estimate the "user-game" matching score. Among them, the pre-trained language model can be a bidirectional encoder representation from transformers (BERT) model, a chat generative language model (ChatGLM), etc. In addition, a generative model in the PLM is also used to incorporate user ID (identity) information and game ID information during the recommendation process. Here, the user ID is the unique identifier of the user, and the game ID is the unique identifier of the game. During the recommendation process, first, a prompt template is used to construct the user ID information, user description information, game ID information, and game description information into a natural language example. Then, the natural language example is input into the PLM, and the PLM will directly generate whether the user clicks on or registers for the game.

[0173] Although the pre-trained language model can solve the semantic understanding problem, it directly abandons information such as user ID and item ID during the recommendation process. This is because the pre-trained language models cannot understand the user ID and item ID in the "user-game matching" recommendation task, and these IDs are usually not in the vocabulary of the pre-trained language model. In addition, in the recommendation scenario, on the one hand, it is necessary to process user IDs at the hundred-million level and game IDs at the ten-thousand level. The sudden increase in the ID vocabulary will lead to insufficient learning of the PLM, making it limited in understanding user ID and game ID information. On the other hand, the IDs are randomly assigned by the system and may have duplicate fields, that is, there is shared information between the ID information, which will affect the estimation results of the model. However, the ID information is very important for the personalized recommendation of the recommendation system. Abandoning information such as user ID and item ID results in a sub-optimal matching score calculated finally.

[0174] Based on the above scenario, the embodiment of the present application provides an information recommendation method, which improves the game delivery effect by integrating the pre-trained language model of heterogeneous graph coordination information. On the one hand, when processing the user ID (i.e., the object identifier of the above-mentioned object to be recommended) and the item ID (i.e., the information identifier of the above-mentioned information to be recommended) at the billion level, considering that the ID vocabulary added to the PLM is very large, it brings great difficulties to model training and reasoning, so a new encoding method is proposed to reduce the vocabulary size. Since the user ID is generally composed of 16-dimensional letters or numbers, the user ID and the item ID can be segmented, and the user ID can be segmented into sub-word granularity by using the Sentence Piece word segmenter, for example, a user ID "U2592597530894676" is segmented into "U-2592-5975-3089-4676". In this way, the ID tokens of the billion level are segmented and compressed into sub-words of the ten thousand level, and then the segmented sub-words are added to the PLM vocabulary, which can effectively compress the vocabulary size and improve the performance and recommendation effect of the model. On the other hand, since the user ID is randomly assigned by the game system, consecutive user ID segments will share a lot of information in the sub-word dimension, but these sub-words do not have explicit correlation in the recommendation dimension. For example, the user ID "0001-0002-0003-0000" and the user ID "0001-0002-0003-00004" have three sub-words "0001, 0002, 0003" shared information at the sub-word granularity (accounting for 75% of the total number of ID segments). In order to resolve the contradiction between ID random encoding and ID segmentation, the embodiment of the present application uses meta-paths to perform node representation learning on the "user-game" interaction heterogeneous graph, re-encodes the collaborative information through ID representation, and then replaces the original ID random encoding to help PLM understand the semantics of the ID.

[0175] In addition, the embodiment of the present application also uses the recommended data to retrain the pre-trained language model. Considering that the pre-trained language model has a rich understanding of natural language descriptions, the prompt engineering can be used to expand the prompt sample data of "user-game" one by one, and different prompt templates can be used to construct different recommended sample data to retrain the pre-trained language model. This allows the pre-trained language model to learn the same sample from different perspectives, thereby improving the pre-trained language model's understanding and adaptability to different contexts and expressions.

[0176] See also Figure 5 , Figure 5 is a schematic diagram of a multi-game launch recommendation provided by an embodiment of the present application, such as Figure 5As shown, in the scenario of certain social software, when the user browses the game official account or clicks on the red dot in the scenario, a game recommendation service will be requested from the terminal. The method provided by the embodiments of this application can use the ID information of the current user (abbreviated as user ID), user description information (abbreviated as user description), game ID information of multiple games in the game library (abbreviated as game ID), and game description information (abbreviated as game description), construct a prompt sample using prompt engineering, and then input it into the encoder and decoder of the pre-trained language model (i.e., PLM) for inference. Among them, the inference process can be to use a prompt template (i.e., user ID, user description, game ID, game description) in prompt engineering to construct a prompt sample, then input the prompt sample into the encoder part of the PLM for encoding, and then input the encoded result into the decoder part of the PLM for decoding, and finally return the click probability (Click Through Rate, CTR) and conversion probability (Conversion Rate, CVR) of the user (ID) relative to each game (ID). Next, the terminal sorts the multiple games in descending order according to the click probability and conversion probability of each game (ID), selects the top K (TOP-K) games at the head of the sequence, and returns the first (TOP-1) game among the K games to recommend to the user.

[0177] The following combines Figure 6 to specifically illustrate the recommendation method provided by the embodiments of this application. In the offline training stage, since the original ID (i.e., user ID and game ID) segments are randomly assigned by the game system, if the ID is directly tokenized, there is shared information among the obtained multiple subwords, which will bring noise to the subsequent model prediction. Therefore, before inputting the user ID and game ID into the pre-trained language model PLM, the heterogeneous graph primitive path node representation algorithm is used to mine the collaborative information between the IDs, and the original user ID and game ID are re-encoded. The specific process of re-encoding is as follows:

[0178] 1. First, collect the "user-game" interaction samples within the most recent T time window and construct a "user-game" interaction bipartite graph (heterogeneous graph);

[0179] 2. Define the heterogeneous graph primitive path collection rule, and perform primitive path collection on the heterogeneous graph by adopting the collection rule of "user-game-user", that is, define the primitive path, and then perform primitive path random walk on the "user-game" interaction bipartite graph to collect sequence data;

[0180] 3. According to the sequence data collected by the primitive path, use the Word2Vector algorithm to perform representation learning on the IDs in the heterogeneous graph vertices to obtain the ID vector representation of each node in the sequence data;

[0181] 4. Perform K-Means clustering on the user ID node vector representation and the game ID node vector representation respectively to obtain multiple clustering clusters. For the multiple clustering clusters, perform intra-class encoding first and then inter-class encoding. That is, preferentially encode the node vector representations within the clustering clusters, and then encode the other clustering clusters. Among them, the encoding order of the node vector representations of each clustering cluster is determined according to the distance from the cluster center. That is, first encode the center point of the clustering cluster, and then encode the remaining node vector representations in order of increasing distance from the center point. After re-encoding, adjacent user ID segments share collaborative information and are correlated.

[0182] After re-encoding the original user ID and game ID through the above steps, use prompt tuning to construct a natural language template, and organize the user ID, user description, game ID, and game description into a natural language representation to generate Natural Language Processing (NLP) samples. For example: "Will the user with ID {user ID} and description {user description} click on the game with ID {game ID} and description {game description}?"

[0183] At the same time, considering that the pre-trained language model (PLM) has a large number of parameters and requires rich recommended sample data during training, the single "user-game" interaction sample can be extended with multiple prompt templates through the template extension method to generate multiple NLP samples, as shown in Table 1 below:

[0184] Table 1

[0185]

[0186] After generating multiple NLP samples through prompt engineering, these samples can be directly input into a pre-trained language model for model re-training and inference. However, existing PLMs cannot understand ID information in recommendations, such as user IDs and game IDs, because these IDs are not in the PLM's vocabulary. Since the number of user IDs is in the hundreds of millions, directly expanding the vocabulary will result in a very large PLM, greatly increasing the computational pressure. Therefore, IDs are not directly added to the vocabulary of the existing PLM. Based on this, the embodiments of this application use Sentence Piece to tokenize user IDs and game IDs, with a continuous four-digit number as a sub-word to split user IDs and game IDs. For example, if a user ID is 4372489274, after tokenization, it becomes three sub-words: 0043-7248-9274. To enable the PLM model to recognize these sub-words during training, ten thousand sub-words in the range of 0000-9999 are added to the PLM vocabulary. By tokenizing IDs, it better balances the problems of "PLMs cannot recognize IDs" and "directly expanding the vocabulary with IDs causes the PLM to run out of memory".

[0187] In addition, during the model training process, the embodiments of this application also introduce whole-word markers to identify each user ID and game ID, that is, the same marker is assigned to all sub-words corresponding to the whole word (user ID or game ID) for identification, and different markers are assigned to the sub-words of different whole words, enabling the model to understand user IDs and game IDs as a whole during the learning process.

[0188] After tokenizing the user IDs and game IDs in the NLP samples and marking the resulting sub-words, the obtained NLP samples can be directly input into the PLM for re-training or inference. As Figure 6 shown, taking the pre-trained language model as an example of the Text to Text Transfer Transformer (T5) model, when an NLP sample "Does user 029874459020 click on game 183974368456?" is input into the T5 model, the start marker of the sample is <s>, the end mark is < / s>. First, we can embed each word in the NLP sample to obtain the corresponding word vector and position encoding vector. At the same time, we can construct the whole word tag vector of each word according to the whole word tag of each word. Among them, "0298, 7445, 9020" belong to the three sub-words after the user ID "029874459020" is segmented, so the whole word tag vectors corresponding to the three sub-words "0298, 7445, 9020" are the same, and the same whole word tag w3 is assigned. Similarly, the whole word tag vectors of the three sub-words "1839, 7436, 8456" after the game ID "183974368456" is assigned to w10. Next, the word vector, position encoding vector and whole word tag vector of the NLP sample are input into the encoder of PLM for encoding processing to obtain the semantic feature vector of the NLP sample. Then the semantic feature vector is input into the autoregressive decoder (referred to as decoder) of PLM for decoding processing. When the T5 model is trained, the cross entropy loss function of the T5 model is constructed according to the decoding results, and the parameters of the T5 model are updated through back propagation of the loss function to retrain the T5 model.

[0189] In the online prediction stage, the user ID that initiates the recommendation service request is first re-encoded using the re-encoding method. The game ID is stored in the terminal and can be processed in batches, that is, one user ID can recommend only one game ID, or one user ID can recommend multiple game IDs, and multiple game IDs can be recommended for each user ID. Then, for each user ID and the corresponding game ID, "user ID, user description information, game ID, game description information" is mapped to a standard NLP sample according to the prompt template in the prompt project. Finally, the NLP sample is input into the re-trained PLM model (such as the T5 model), first encoded by the encoder and then decoded by the decoder, and the "user-sample" matching score can be predicted. For each user ID, the click probability can be used as the matching score to sort all the game IDs in descending order, and the game sequence obtained by descending order can be recommended to the corresponding user.

[0190] For example, a certain NLP sample "Does user 029874459020 click on game 183974368456?" is input into the retrained T5 model for inference, and the final output "Does user 029874459020 click on game 183974368456?" has a click probability of 1.0, which means that user 029874459020 will definitely click on game 183974368456. Thus, the click probability of each game ID in the game library can be inferred for user 029874459020. Similarly, the click probability of each game ID in the game library can be obtained for other user IDs.

[0191] To verify the effectiveness of the recommendation method, in the embodiments of the present application, evaluation is carried out from an offline dimension. In the offline evaluation session, three prediction algorithms based on MLP (Algorithm 1), based on PLM (Algorithm 2), and based on generative PLM (Algorithm 3) are respectively compared. The evaluation metrics are the area under the curve (AUC), loss value (LOSS), and accuracy (Accuracy).

[0192] The results of the offline experiment are shown in Table 2 below:

[0193] Table 2

[0194]

[0195] As can be seen from Table 2, in the experimental stage of offline training, compared with the three prediction algorithms based on MLP (Algorithm 1), based on PLM (Algorithm 2), and based on generative PLM (Algorithm 3), the algorithm of the present application can obtain a lower loss function on the test set and a higher prediction accuracy. And through other experiments, it is proved that compared with other prediction algorithms, the algorithm of the present application has improved in terms of the click-through rate and conversion rate of the game.

[0196] Next, the implementation of the information recommendation device 453 provided in the embodiments of the present application as an exemplary structure of software modules will be continued. In some embodiments, as Figure 2 shown, the software modules in the information recommendation device 453 stored in the memory 450 may include: an acquisition module 4531, configured to acquire a heterogeneous graph including a plurality of nodes, where each node is used to represent any one of the object identifier of the object to be recommended and the information identifier of the information to be recommended; an encoding module 4532, configured to determine the identification feature of each node, and perform re-encoding processing on the identification feature of each node to obtain the re-encoded identifier of each node; a splitting module 4533, configured to perform splitting processing on the re-encoded identifier of each node to obtain the sub-identifier of each node, where each node corresponds to a plurality of sub-identifiers; a determination module 4534, configured to determine the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended from the sub-identifiers respectively corresponding to the plurality of nodes; a recommendation module 4535, configured to determine the recommendation metric of the information to be recommended corresponding to the object to be recommended based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, and perform a recommendation operation based on the recommendation metric.

[0197] In some embodiments, the encoding module 4532 is further configured to collect paths of the heterogeneous graph to obtain a node sequence of the heterogeneous graph; and determine the identification feature of each node based on the node sequence of the heterogeneous graph.

[0198] In some embodiments, the encoding module 4532 is further configured to arbitrarily select a node from the heterogeneous graph as the starting random walk node; start from the starting random walk node and perform random walk processing along the edges in the heterogeneous graph to obtain a plurality of random walk nodes; sort the starting random walk node and the plurality of random walk nodes in the order of the random walk to obtain a node sequence.

[0199] In some embodiments, the encoding module 4532 is further configured to obtain a target node sequence including a target node from the node sequence of the heterogeneous graph, where the target node is any node in the heterogeneous graph; update the initial feature of the target node based on the adjacent nodes of the target node in the target node sequence to obtain the identification feature of the target node.

[0200] In some embodiments, the encoding module 4532 is further configured to perform adjacent node prediction processing on the target node based on the initial feature of the target node in the target node sequence to obtain the predicted distribution probability of the adjacent nodes of the target node; determine the error between the predicted distribution probability and the distribution probability label, where the distribution probability label is the label of the distribution probability of the adjacent nodes of the target node; update the initial feature of the target node until the error converges, and use the updated initial feature of the target node when the error converges as the identification feature of the target node.

[0201] In some embodiments, the encoding module 4532 is further configured to perform the following processing for each node: classify the identification feature of the node to obtain an object identification feature set and an information identification feature set, where the object identification feature set includes the identification features of at least one object to be recommended, and the information identification feature set includes the identification features of at least one piece of information to be recommended; perform clustering processing on the identification features included in the object identification feature set to obtain an object identification clustering cluster, and perform clustering processing on the identification features included in the information identification feature set to obtain an information identification clustering cluster; perform re-encoding processing on the identification features of the nodes in each target clustering cluster to obtain the re-encoded identification of the nodes in the target clustering cluster, where the target clustering cluster is an object identification clustering cluster or an information identification clustering cluster.

[0202] In some embodiments, the encoding module 4532 is further configured to perform the following processing for each target clustering cluster: determine the identification feature of the node at the center of the target clustering cluster as the center identification feature, and perform re-encoding processing on the center identification feature to obtain the re-encoded identification of the node at the center of the target clustering cluster; determine the distance between other identification features and the center identification feature, where the other identification features are the identification features of the nodes other than the node at the center in the target clustering cluster; sort the other identification features according to the distance, and perform re-encoding processing on the other identification features in the sorted order to obtain the re-encoded identification of the nodes other than the node at the center in the target clustering cluster.

[0203] In some embodiments, the splitting module 4533 is further configured to obtain an identification sequence composed of the recoding identifications of nodes; determine the splitting length of the identification sequence; and perform splitting processing on the identification sequence based on the splitting length to obtain sub-identifications of the nodes.

[0204] In some embodiments, the recommendation module 4535 is further configured to obtain the description information of the object to be recommended and the description information of the information to be recommended; and construct at least one intelligent prompt message based on the description information of the object to be recommended, the description information of the information to be recommended, the sub-identifications of the object to be recommended, and the sub-identifications of the information to be recommended.

[0205] Perform inference processing on the intelligent prompt message to obtain the recommendation index of the information to be recommended corresponding to the object to be recommended.

[0206] In some embodiments, the recommendation module 4535 is further configured to determine the word vectors of the intelligent prompt message and the position encoding vectors corresponding to the word vectors, and determine the whole-word marking vectors in the intelligent prompt message; merge the word vectors, the position encoding vectors, and the whole-word marking vectors to obtain the fused feature vectors of the intelligent prompt message; perform encoding processing on the fused feature vectors to obtain the semantic feature vectors of the intelligent prompt message; and perform decoding processing on the semantic feature vectors to obtain the recommendation index of the information to be recommended corresponding to the object to be recommended.

[0207] In some embodiments, the recommendation module 4535 is further configured to perform word segmentation processing on the intelligent prompt message to obtain a plurality of segmented words; assign the same whole-word marking to the segmented words belonging to the same recoding identification, where different recoding identifications correspond to different whole-word markings; perform embedding processing on the whole-word markings assigned to each segmented word to obtain the whole-word marking vectors of the segmented words; and merge the whole-word marking vectors of each segmented word to obtain the whole-word marking vectors of the intelligent prompt message.

[0208] An embodiment of the present application provides a computer program product, which includes a computer program or computer-executable instructions, and the computer program or computer-executable instructions are stored in a computer-readable storage medium. The processor of the electronic device reads the computer-executable instructions from the computer-readable storage medium, and the processor executes the computer-executable instructions, so that the electronic device executes the information recommendation method described above in the embodiments of the present application.

[0209] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, where computer-executable instructions or a computer program are stored, and when the computer-executable instructions or the computer program are executed by a processor, the processor will be caused to execute the information recommendation method provided in the embodiments of the present application, for example, Figures 3A to 3K the information recommendation method shown.

[0210] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disc, or CD-ROM; it may also be various devices including one or any combination of the above memories.

[0211] In some embodiments, the computer-executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including being deployed as a stand-alone program or being deployed as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0212] As an example, the computer-executable instructions may or may not correspond to a file in the file system, may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, stored in a single file dedicated to the program being discussed, or, stored in multiple cooperating files (e.g., files that store one or more modules, subroutines, or portions of code).

[0213] As an example, the computer-executable instructions may be deployed to execute on one electronic device, or on multiple electronic devices located at one location, or, on multiple electronic devices distributed at multiple locations and interconnected by a communication network.

[0214] In summary, through the embodiments of the present application, for the object identifier of the object to be recommended represented by the nodes in the heterogeneous graph and the information identifier of the information to be recommended, the association relationship between the nodes is mined through the node sequence, and the identifier features of the nodes are also extracted. Then, the corresponding identifier feature sets are obtained through the method of clustering respectively, and the identifier features of each node in the identifier feature set are re-encoded, which can greatly reduce the possibility of duplicate fields existing between the obtained re-encoded identifiers, and eliminate as much as possible the shared information of the duplicate fields existing between the identifiers. On the basis of retaining the association relationship between the identifiers, the influence of the noise of the shared information on the final recommendation effect is reduced. Next, the re-encoded identifier of each node is split to obtain multiple sub-identifiers of each node. By splitting the identifier field in this way, the field length of the object identifier or information identifier represented by the node can be shortened, so as to reduce the number of fields of the object identifier or information identifier, and further reduce the vocabulary of the object identifier or information identifier. Finally, based on multiple sub-identifiers, intelligent prompt information is constructed, so that the pre-trained language model for information recommendation can more easily understand and recognize the object identifier and information identifier in the intelligent prompt information, and infer the recommendation index for information recommendation, reducing the inference difficulty of the pre-trained language model and improving the effect of information recommendation.

[0215] It can be understood that in the embodiments of the present application, relevant data such as user information is involved. When the embodiments of the present application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0216] The above is only the embodiments of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. An information recommendation method, characterized in that, The method includes: Obtaining a heterogeneous graph including multiple nodes, where each of the nodes is used to represent either an object identifier of an object to be recommended or an information identifier of the information to be recommended; Determining the identification feature of each of the nodes, and performing a re-encoding process on the identification feature of each of the nodes to obtain a re-encoded identifier of each of the nodes; Performing a splitting process on the re-encoded identifier of each of the nodes to obtain a sub-identifier of each of the nodes, where each of the nodes corresponds to multiple sub-identifiers; Determining the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended from the sub-identifiers respectively corresponding to the multiple nodes; Based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, determining a recommendation metric for the information to be recommended corresponding to the object to be recommended, and performing a recommendation operation based on the recommendation metric.

2. The method according to claim 1, wherein The determining the identification feature of each of the nodes includes: Performing path collection on the heterogeneous graph to obtain a node sequence of the heterogeneous graph; Based on the node sequence of the heterogeneous graph, determining the identification feature of each of the nodes.

3. The method according to claim 2, wherein The heterogeneous graph includes edges for connecting two of the nodes. The performing path collection on the heterogeneous graph to obtain a node sequence of the heterogeneous graph includes: Arbitrarily selecting a node from the heterogeneous graph as a starting wandering node; Starting from the starting wandering node, performing a random walk process along the edges in the heterogeneous graph to obtain multiple wandering nodes; Sorting the starting wandering node and the multiple wandering nodes in the order of wandering to obtain the node sequence.

4. The method according to claim 2, wherein The based on the node sequence of the heterogeneous graph, determining the identification feature of each of the nodes includes: Obtaining a target node sequence including a target node from the node sequence of the heterogeneous graph, where the target node is any node in the heterogeneous graph; Based on the adjacent nodes of the target node in the target node sequence, performing an update process on the initial feature of the target node to obtain the identification feature of the target node.

5. The method according to claim 4, wherein The based on the adjacent nodes of the target node in the target node sequence, performing an update process on the initial feature of the target node to obtain the identification feature of the target node includes: Based on the initial feature of the target node in the target node sequence, performing an adjacent node prediction process on the target node to obtain a predicted distribution probability of the adjacent nodes of the target node; Determining an error between the predicted distribution probability and a distribution probability label, where the distribution probability label is a label of the distribution probability of the adjacent nodes of the target node; Updating the initial feature of the target node until the error converges, and taking the updated initial feature of the target node when the error converges as the identification feature of the target node.

6. The method according to claim 1, characterized in that The performing a re-encoding process on the identification feature of each of the nodes to obtain a re-encoded identifier of each of the nodes includes: Performing the following process for each of the nodes: Classify the identification features of the nodes to obtain an object identification feature set and an information identification feature set, where the object identification feature set includes at least one identification feature of the object to be recommended, and the information identification feature set includes at least one identification feature of the information to be recommended; Perform clustering processing on the identification features included in the object identification feature set to obtain an object identification clustering cluster, and perform clustering processing on the identification features included in the information identification feature set to obtain an information identification clustering cluster; Perform re-encoding processing on the identification features of the nodes in each target clustering cluster to obtain the re-encoded identification of the nodes in the target clustering cluster, where the target clustering cluster is the object identification clustering cluster or the information identification clustering cluster.

7. The method according to claim 6, wherein The performing re-encoding processing on the identification features of the nodes in each target clustering cluster to obtain the re-encoded identification of the nodes in the target clustering cluster includes: Perform the following processing for each of the target clustering clusters: Determine the identification feature of the node at the center of the target clustering cluster as the central identification feature, and perform re-encoding processing on the central identification feature to obtain the re-encoded identification of the node at the center of the target clustering cluster; Determine the distance between other identification features and the central identification feature, where the other identification features are the identification features of the nodes other than the node at the center in the target clustering cluster; Sort the other identification features according to the distance, and perform re-encoding processing on the other identification features in the order of the sorting to obtain the re-encoded identification of the nodes other than the node at the center in the target clustering cluster.

8. The method according to claim 1, wherein The splitting the re-encoded identification of each node to obtain the sub-identifications of each node includes: Perform the following processing for the re-encoded identification of each node: Obtain an identification sequence composed of the re-encoded identification of the node; Determine the splitting length of the identification sequence; Based on the splitting length, perform splitting processing on the identification sequence to obtain the sub-identifications of the node.

9. The method according to claim 1, wherein The determining the recommendation index of the information to be recommended corresponding to the object to be recommended based on the sub-identifications of the object to be recommended and the sub-identifications of the information to be recommended includes: Obtain the description information of the object to be recommended and the description information of the information to be recommended; Based on the description information of the object to be recommended, the description information of the information to be recommended, the sub-identifications of the object to be recommended, and the sub-identifications of the information to be recommended, construct at least one intelligent prompt message; Perform inference processing on the intelligent prompt message to obtain the recommendation index of the information to be recommended corresponding to the object to be recommended.

10. The method according to claim 9, characterized in that, The performing inference processing on the intelligent prompt message to obtain the recommendation index of the information to be recommended corresponding to the object to be recommended includes: Determine the word vector of the intelligent prompt message, the position encoding vector corresponding to the word vector, and the whole word marking vector in the intelligent prompt message; Combine the word vector, the position encoding vector, and the whole word marking vector to obtain the fused feature vector of the intelligent prompt message; Encode the fused feature vector to obtain the semantic feature vector of the intelligent prompt information; Decode the semantic feature vector to obtain the recommendation metrics of the information to be recommended corresponding to the object to be recommended.

11. The method according to claim 10, characterized in that, The determination of the whole-word token vector of the intelligent prompt information includes: Perform word segmentation on the intelligent prompt information to obtain a plurality of segmented words; Assign the same whole-word token to the segmented words belonging to the same re-encoding identifier, where different re-encoding identifiers correspond to different whole-word tokens; Perform embedding processing on the whole-word token assigned to each segmented word to obtain the whole-word token vector of the segmented word; Merge the whole-word token vectors of each segmented word to obtain the whole-word token vector of the intelligent prompt information.

12. An information recommendation device, characterized in that, The device includes: An acquisition module, configured to acquire a heterogeneous graph including a plurality of nodes, where each node is used to represent any one of the object identifier of the object to be recommended and the information identifier of the information to be recommended; An encoding module, configured to determine the identification feature of each node, and perform re-encoding processing on the identification feature of each node to obtain the re-encoding identifier of each node; A splitting module, configured to perform splitting processing on the re-encoding identifier of each node to obtain the sub-identifier of each node, where each node corresponds to a plurality of sub-identifiers; A determination module, configured to determine the sub-identifier of the object to be recommended or the sub-identifier of the information to be recommended from the sub-identifiers respectively corresponding to the plurality of nodes; A recommendation module, configured to determine the recommendation metrics of the information to be recommended corresponding to the object to be recommended based on the sub-identifier of the object to be recommended and the sub-identifier of the information to be recommended, and perform a recommendation operation based on the recommendation metrics.

13. An electronic device, characterized in that, The electronic device includes: A memory, configured to store computer-executable instructions or computer programs; A processor, configured to implement the information recommendation method according to any one of claims 1 to 11 when executing the computer-executable instructions or computer programs stored in the memory.

14. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that, The computer-executable instructions or computer programs, when executed by the processor, implement the information recommendation method according to any one of claims 1 to 11.

15. A computer program product, comprising computer-executable instructions or a computer program, characterized in that, The computer-executable instructions or computer programs, when executed by the processor, implement the information recommendation method according to any one of claims 1 to 11.

Images