Recommendation methods and apparatus, and training methods and apparatus for recommendation models.

Abstract

This disclosure relates to a recommendation method and apparatus, and a training method and apparatus for a recommendation model. The training method includes: obtaining the user account of the object to be recommended and the object identifiers of at least two objects to be recommended; inputting information from a hypergraph into a hypergraph convolutional network to obtain embedding representation information of the training user account, objects, and behaviors in the training data, wherein the training data includes multiple tuples, each tuple including a training user account, an object identifier of an object, and a behavior identifier of the behavior generated by the training user account on the object, and the hypergraph is a graph constructed based on the multiple tuples in the training data with the training user account, object identifier, and behavior identifier as nodes; inputting the embedding representation information, the user account, and the object identifiers of at least two objects to be recommended into the recommendation network to obtain recommendation information for each object to be recommended; and recommending at least two objects to be recommended to the user account based on the recommendation information.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence, and in particular to a recommendation method and apparatus, and a training method and apparatus for a recommendation model. Background Technology

[0002] Currently, recommender systems can learn users' interests based on their historical behavior data, providing personalized information access. Traditional collaborative filtering (CF) operates on only one behavior (such as a purchase or a click). However, modeling only one behavior has limitations in real-world recommender engines. For example, on e-commerce websites, users' purchase records are always much scarcer than their click records, leading to the problem of data scarcity.

[0003] To address this issue, traditional collaborative filtering is typically extended to multi-behavior recommendation. Specifically, multi-behavior recommendation utilizes multiple types of behaviors, including target behaviors and auxiliary behaviors, to improve the recommendation performance of the target behavior. Currently, the most advanced multi-behavior recommendation algorithms using graph neural networks rely on Multi-Behavior Graph Convolutional Networks (MBGCNs). Taking video recommendation as an example, MBGCN recommends videos to users by convolving a hypergraph containing user nodes and video nodes. This model allows various behaviors to be independently modeled through convolution, ignoring the strong correlation between behaviors. For instance, on short video platforms, a user watches a video, finds it interesting, and decides to like or save it. This "like-save" behavior represents a strong correlation between these two behaviors. However, MBGCN ignores this strong correlation, making it difficult to capture global information in semantically complex scenarios. Summary of the Invention

[0004] This disclosure provides a recommendation method and apparatus, and a training method and apparatus for a recommendation model, to at least solve the problem that related technologies cannot capture global information, resulting in recommended objects not meeting user needs well.

[0005] According to a first aspect of the present disclosure, a recommendation method is provided. The recommendation method is implemented based on a recommendation model, which includes a hypergraph convolutional network and a recommendation network. The recommendation method includes: obtaining a user account for an object to be recommended and object identifiers of at least two objects to be recommended; inputting information from the hypergraph into the hypergraph convolutional network to obtain embedding representation information of training user accounts, objects, and behaviors in training data, wherein the training data includes multiple tuples, each tuple including a training user account, an object identifier of an object, and a behavior identifier of a behavior generated by the training user account on the object, the hypergraph being a graph constructed based on multiple tuples in the training data with training user accounts, object identifiers, and behavior identifiers as nodes, and the edges of the hypergraph being lines connecting at least two nodes of the training user account, object identifier, and behavior identifier in a tuple; inputting the embedding representation information, the user account, and the object identifiers of at least two objects to be recommended into the recommendation network to obtain recommendation information for each object to be recommended, wherein the recommendation information represents the probability that the behavior generated by the user account on the corresponding object to be recommended is the target behavior, and the training data includes the target behavior; and recommending at least two objects to be recommended to the user account based on the recommendation information.

[0006] Optionally, the hypergraph includes at least one of the following: a first hypergraph and a second hypergraph, wherein the first and second hypergraphs are constructed as follows: a first hypergraph is constructed based on a target tuple, with user account, object identifier, and behavior identifier as nodes, wherein the target tuple is a tuple containing the behavior identifier of the target behavior among multiple tuples; and a second hypergraph is constructed based on non-target tuples, with user account, object identifier, and behavior identifier as nodes, wherein the non-target tuples are tuples other than the target tuple among multiple tuples.

[0007] Optionally, when the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The information of the hypergraph is input into the hypergraph convolutional network to obtain embedding representation information of training user accounts, objects, and behaviors in the training data. This includes: inputting the information of the first hypergraph into the first hypergraph convolutional network to obtain first embedding representation information of training user accounts, objects, and behaviors in the training data; inputting the information of the second hypergraph into the second hypergraph convolutional network to obtain second embedding representation information of training user accounts, objects, and behaviors in the training data; and obtaining embedding representation information based on the first embedding representation information, the second embedding representation information, and the weight coefficients.

[0008] Optionally, the hypergraph also includes a third hypergraph constructed as follows: a third hypergraph is constructed based on the target tuple, with user account and object identifier as nodes.

[0009] Optionally, the hypergraph convolutional network further includes a third hypergraph convolutional network. The hypergraph information is input into the third hypergraph convolutional network to obtain embedding representation information of training user accounts, objects, and behaviors in the training data. This includes: inputting the information of the third hypergraph into the third hypergraph convolutional network to obtain third embedding representation information of training user accounts and objects in the training data; and obtaining embedding representation information based on the first embedding representation information, the second embedding representation information, and the weight coefficients. This includes: obtaining embedding representation information based on the first embedding representation information, the second embedding representation information, the third embedding representation information, and the weight coefficients.

[0010] Optionally, the recommendation network includes a first recommendation network and a second recommendation network. Embedded representation information, user accounts, and object identifiers of at least two objects to be recommended are input into the recommendation network to obtain recommendation information for each object to be recommended. This includes: inputting the embedded representation information, user accounts, and object identifiers of at least two objects to be recommended into the first recommendation network to obtain first recommendation sub-information for each object to be recommended. The first recommendation sub-information is the probability that the user account's behavior towards the corresponding object to be recommended is the target behavior, based on information in the hypergraph from the user account and the object identifiers of at least two objects to be recommended. Then, the user account set of each object to be recommended is obtained; the embedded representation information, the object identifier set, and each user account set are input into the second recommendation network to obtain second recommendation sub-information for each object to be recommended. The second recommendation sub-information is the probability that the user account's behavior towards the corresponding object to be recommended is the target behavior, based on information in the hypergraph from the object identifier set and each user account set; and finally, the recommendation information for each object to be recommended is obtained based on the first and second recommendation sub-information.

[0011] According to a second aspect of the present disclosure, a method for training a recommendation model is provided. The recommendation model includes a hypergraph convolutional network and a recommendation network. The training method includes: acquiring training data, wherein the training data includes multiple tuples and behavior identifiers of target behaviors, each tuple including a training user account, an object identifier of an object, and behavior identifiers of behaviors between the training user account and the object; constructing a hypergraph based on the multiple tuples with user accounts, object identifiers, and behavior identifiers as nodes; inputting the information of the hypergraph into the hypergraph convolutional network to obtain estimated embedding representation information of training user accounts, objects, and behaviors in the training data; using a first training user account and a first object identifier in any target tuple as training positive samples, wherein the target tuple is a tuple containing behavior identifiers of target behaviors among multiple tuples; and using a second object identifier and a first object identifier in any non-target tuple as training positive samples. A training user account is used as a training negative sample, where non-target tuples are tuples other than the target tuples among multiple tuples. The training positive samples and the estimated embedding representations are input into the recommendation network to obtain the first estimated recommendation information for the object corresponding to the first object identifier, where the first estimated recommendation information represents the probability that the action performed by the first training user account on the object corresponding to the first object identifier is the target action. The training negative samples and the estimated embedding representations are input into the recommendation network to obtain the second estimated recommendation information for the object corresponding to the second object identifier, where the second estimated recommendation information represents the probability that the action performed by the first training user account on the object corresponding to the second object identifier is the target action. By maximizing the loss of the first and second estimated recommendation information, the parameters of the hypergraph convolutional network and the recommendation network are adjusted to train the recommendation model.

[0012] Optionally, a hypergraph is constructed based on multiple tuples, with user account, object identifier, and behavior identifier as nodes, including at least one of the following: a first hypergraph is constructed based on a target tuple, with user account, object identifier, and behavior identifier as nodes, wherein the target tuple is a tuple containing the behavior identifier of the target behavior among the multiple tuples; a second hypergraph is constructed based on non-target tuples, with user account, object identifier, and behavior identifier as nodes, wherein the non-target tuples are tuples other than the target tuple among the multiple tuples.

[0013] Optionally, when the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The hypergraph information is input into the hypergraph convolutional network to obtain estimated embedding representations of user accounts, objects, and behaviors in the training data. This includes: inputting the information of the first hypergraph into the first hypergraph convolutional network to obtain first estimated embedding representations of training user accounts, objects, and behaviors in the training data; inputting the information of the second hypergraph into the second hypergraph convolutional network to obtain second estimated embedding representations of training user accounts, objects, and behaviors in the training data; and obtaining estimated embedding representations based on the first estimated embedding representations, the second estimated embedding representations, and the weight coefficients.

[0014] Optionally, constructing a hypergraph based on multiple tuples with user account, object identifier, and behavior identifier as nodes further includes: constructing a third hypergraph based on the target tuple with user account and object identifier as nodes.

[0015] Optionally, the hypergraph convolutional network further includes a third hypergraph convolutional network. The hypergraph information is input into the hypergraph convolutional network to obtain estimated embedding representations of user accounts, objects, and behaviors in the training data. This includes: inputting the information of the third hypergraph into the third hypergraph convolutional network to obtain third estimated embedding representations of training user accounts and objects in the training data; and obtaining estimated embedding representations based on the first estimated embedding representations, the second estimated embedding representations, and the weight coefficients. This includes: obtaining estimated embedding representations based on the first estimated embedding representations, the second estimated embedding representations, the third estimated embedding representations, and the weight coefficients.

[0016] Optionally, the training data also includes the actual conversion rate of each behavior in the training data, where the actual conversion rate is the probability that each behavior is actually converted into the target behavior; the recommendation model is trained by adjusting the parameters of the hypergraph convolutional network and the recommendation network by maximizing the loss of the first and second estimated recommendation information, including: determining a first loss based on the first and second estimated recommendation information; obtaining the estimated conversion rate of the intermediate layer output of the recommendation model, where the estimated conversion rate is the probability that each behavior in the training data is predicted to be converted into the target behavior; determining a second loss based on the estimated conversion rate and the actual conversion rate; and training the recommendation model by adjusting the parameters of the hypergraph convolutional network and the recommendation network by maximizing the first loss and minimizing the second loss.

[0017] According to a third aspect of the present disclosure, a recommendation apparatus is provided. The recommendation apparatus is implemented based on a recommendation model, which includes a hypergraph convolutional network and a recommendation network. The recommendation apparatus includes: an identifier acquisition unit configured to acquire user accounts of objects to be recommended and object identifiers of at least two objects to be recommended; and an embedding representation information acquisition unit configured to input hypergraph information into the hypergraph convolutional network to obtain embedding representation information of training user accounts, objects, and behaviors in training data. The training data includes multiple tuples, each tuple including a training user account, an object identifier of an object, and a behavior identifier of a behavior generated by the training user account on the object. The hypergraph... The hypergraph is constructed based on multiple tuples in the training data, with training user accounts, object identifiers, and behavior identifiers as nodes. The edges of the hypergraph are lines connecting at least two nodes in a tuple of training user accounts, object identifiers, and behavior identifiers. The recommendation information acquisition unit is configured to input embedded representation information, user accounts, and object identifiers of at least two objects to be recommended into the recommendation network to obtain recommendation information for each object to be recommended. The recommendation information represents the probability that the user account's behavior toward the corresponding object to be recommended is the target behavior, and the training data includes the target behavior. The recommendation unit is configured to recommend at least two objects to be recommended to the user account based on the recommendation information.

[0018] Optionally, the hypergraph includes at least one of the following: a first hypergraph and a second hypergraph, wherein the first and second hypergraphs are constructed as follows: a first hypergraph is constructed based on a target tuple, with user account, object identifier, and behavior identifier as nodes, wherein the target tuple is a tuple containing the behavior identifier of the target behavior among multiple tuples; and a second hypergraph is constructed based on non-target tuples, with user account, object identifier, and behavior identifier as nodes, wherein the non-target tuples are tuples other than the target tuple among multiple tuples.

[0019] Optionally, when the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The embedding representation information acquisition unit is further configured to input the information of the first hypergraph into the first hypergraph convolutional network to obtain first embedding representation information of training user accounts, objects, and behaviors in the training data; input the information of the second hypergraph into the second hypergraph convolutional network to obtain second embedding representation information of training user accounts, objects, and behaviors in the training data; and obtain embedding representation information based on the first embedding representation information, the second embedding representation information, and the weight coefficients.

[0020] Optionally, the hypergraph also includes a third hypergraph constructed as follows: a third hypergraph is constructed based on the target tuple, with user account and object identifier as nodes.

[0021] Optionally, the hypergraph convolutional network further includes a third hypergraph convolutional network. The embedding representation information acquisition unit is also configured to input the information of the third hypergraph into the third hypergraph convolutional network to obtain the third embedding representation information of the training user accounts and objects in the training data; and to obtain the embedding representation information based on the first embedding representation information, the second embedding representation information, the third embedding representation information and the weight coefficients.

[0022] Optionally, the recommender network includes a first recommender network and a second recommender network. The recommender information acquisition unit is configured to input embedded representation information, user accounts, and object identifiers of at least two objects to be recommended into the first recommender network to obtain first recommender sub-information for each object to be recommended. The first recommender sub-information is the probability that the user account's behavior towards the corresponding object to be recommended is the target behavior, based on the information of the user account and the object identifiers of the at least two objects to be recommended in the hypergraph. The unit also acquires a set of object identifiers of objects that the user account has interacted with through the target behavior and a set of user accounts of user accounts that have interacted with each object to be recommended through the target behavior. The unit further inputs embedded representation information, the set of object identifiers, and each user account set into the second recommender network to obtain second recommender sub-information for each object to be recommended. The second recommender sub-information is the probability that the user account's behavior towards the corresponding object to be recommended is the target behavior, based on the information of the set of object identifiers and each user account set in the hypergraph. Finally, the unit obtains recommender information for each object to be recommended based on the first and second recommender sub-information.

[0023] According to a fourth aspect of the present disclosure, a training apparatus for a recommendation model is provided. The recommendation model includes a hypergraph convolutional network and a recommendation network. The training apparatus includes: a training data acquisition unit configured to acquire training data, wherein the training data includes multiple tuples and behavior identifiers of target behaviors, each tuple including a training user account, an object identifier of an object, and behavior identifiers of behaviors between the training user account and the object; a hypergraph acquisition unit configured to construct a hypergraph with user accounts, object identifiers, and behavior identifiers as nodes based on the multiple tuples; a predicted embedding representation information acquisition unit configured to input the hypergraph information into the hypergraph convolutional network to obtain predicted embedding representation information of training user accounts, objects, and behaviors in the training data; a positive sample acquisition unit configured to use a first training user account and a first object identifier in any target tuple as training positive samples, wherein the target tuple is a tuple containing behavior identifiers of target behaviors among multiple tuples; and a negative sample acquisition unit configured to use any non-target tuple as training positive samples. The second object identifier and the first training user account in the group are used as training negative samples, wherein non-target tuples are tuples other than target tuples among multiple tuples; the first predicted recommendation information acquisition unit is configured to input the training positive samples and the predicted embedding representation information into the recommendation network to obtain the first predicted recommendation information of the object corresponding to the first object identifier, wherein the first predicted recommendation information represents the probability that the behavior generated by the first training user account on the object corresponding to the first object identifier is the target behavior; the second predicted recommendation information acquisition unit is configured to input the training negative samples and the predicted embedding representation information into the recommendation network to obtain the second predicted recommendation information of the object corresponding to the second object identifier, wherein the second predicted recommendation information represents the probability that the behavior generated by the first training user account on the object corresponding to the second object identifier is the target behavior; the training unit is configured to train the recommendation model by maximizing the loss of the first predicted recommendation information and the second predicted recommendation information and adjusting the parameters of the hypergraph convolutional network and the recommendation network.

[0024] Optionally, the hypergraph acquisition unit is further configured to: construct a first hypergraph based on a target tuple with user account, object identifier, and behavior identifier as nodes, wherein the target tuple is a tuple containing the behavior identifier of the target behavior among multiple tuples; and construct a second hypergraph based on non-target tuples with user account, object identifier, and behavior identifier as nodes, wherein the non-target tuples are tuples other than the target tuple among multiple tuples.

[0025] Optionally, when the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The estimated embedding representation information acquisition unit is further configured to input the information of the first hypergraph into the first hypergraph convolutional network to obtain first estimated embedding representation information of training user accounts, objects, and behaviors in the training data; input the information of the second hypergraph into the second hypergraph convolutional network to obtain second estimated embedding representation information of training user accounts, objects, and behaviors in the training data; and obtain estimated embedding representation information based on the first estimated embedding representation information, the second estimated embedding representation information, and the weight coefficients.

[0026] Optionally, the hypergraph acquisition unit is also configured to construct a third hypergraph based on the target tuple, with user accounts and object identifiers as nodes.

[0027] Optionally, the hypergraph convolutional network further includes a third hypergraph convolutional network. The estimated embedding representation information acquisition unit is also configured to input the information of the third hypergraph into the third hypergraph convolutional network to obtain the third estimated embedding representation information of the training user accounts and objects in the training data; and to obtain the estimated embedding representation information based on the first estimated embedding representation information, the second estimated embedding representation information, the third estimated embedding representation information and the weight coefficients.

[0028] Optionally, the training data also includes the actual conversion rate of each behavior in the training data, where the actual conversion rate is the probability that each behavior is actually converted into the target behavior; the training unit is further configured to determine a first loss based on a first estimated recommendation information and a second estimated recommendation information; obtain the estimated conversion rate of the intermediate layer output of the recommendation model, where the estimated conversion rate is the probability that each behavior in the training data is estimated to be converted into the target behavior; determine a second loss based on the estimated conversion rate and the actual conversion rate; and train the recommendation model by adjusting the parameters of the hypergraph convolutional network and the recommendation network based on maximizing the first loss and minimizing the second loss.

[0029] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement a training method for a recommended method and / or a recommended model according to the present disclosure.

[0030] According to a sixth aspect of the present disclosure, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by at least one processor, causes at least one processor to perform a training method of the recommended method and / or recommended model as described above according to the present disclosure.

[0031] According to a seventh aspect of the present disclosure, a computer program product is provided, including computer instructions that, when executed by a processor, implement a training method for a recommended method and / or a recommended model according to the present disclosure.

[0032] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0033] According to the recommendation method and apparatus, and the training method and apparatus for the recommendation model disclosed herein, not only are users and objects treated as nodes in the graph, but each behavior is also considered as a node in the graph. That is, behavior nodes are embedded in the graph. By utilizing the hypergraph framework, the multi-behavior recommendation problem is redefined. The representation of behavior can perceive global information across the entire training data, thus possessing complex semantics. Based on this, the recommendation information for the object to be recommended to the user is determined, improving the accuracy of the recommended object and ensuring that the recommended object received by the user better meets the user's needs. Therefore, this disclosure solves the problem that related technologies cannot capture global information, resulting in recommended objects not well meeting the user's needs.

[0034] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0035] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0036] Figure 1 This is a schematic diagram illustrating a training method for a recommendation model and an implementation scenario of the recommendation method according to exemplary embodiments of the present disclosure;

[0037] Figure 2 This is a flowchart illustrating a recommendation method according to an exemplary embodiment;

[0038] Figure 3 This is a flowchart illustrating a training method for a recommendation model according to an exemplary embodiment;

[0039] Figure 4 This is a schematic diagram illustrating the structure of a recommendation model according to an exemplary embodiment;

[0040] Figure 5 This is a block diagram illustrating a recommended apparatus according to an exemplary embodiment;

[0041] Figure 6 This is a block diagram illustrating a training apparatus for a recommendation model according to an exemplary embodiment;

[0042] Figure 7This is a block diagram of an electronic device 700 according to an embodiment of the present disclosure. Detailed Implementation

[0043] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0044] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following examples do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0045] It should be noted that the phrase "at least one of several items" in this disclosure refers to three parallel cases: "any one of the several items", "a combination of any number of the several items", and "all of the several items". For example, "including at least one of A and B" includes the following three parallel cases: (1) including A; (2) including B; (3) including A and B. As another example, "performing at least one of step one and step two" indicates the following three parallel cases: (1) performing step one; (2) performing step two; (3) performing both step one and step two.

[0046] This disclosure provides a training method and a recommendation method for a recommendation model, which can help the recommendation model learn user preference information and object features, and recommend more suitable objects to users. The following example illustrates this scenario of recommending videos of interest to users.

[0047] Figure 1 This is a schematic diagram illustrating the training method of the recommendation model and the implementation scenario of the recommendation method according to exemplary embodiments of the present disclosure, such as... Figure 1 The implementation scenario includes server 100, user terminal 110, and user terminal 120. The number of user terminals is not limited to two, and includes, but is not limited to, devices such as mobile phones and personal computers. The user terminals can install video browsing applications. The server can be a single server, a server cluster composed of several servers, or a cloud computing platform or virtualization center.

[0048] After receiving a request from user terminal 110 or 120 to train a recommendation model, server 100 collects statistics on users watching videos, videos watched, and behaviors related to videos. Based on this information, it obtains training data, which includes multiple tuples and behavior identifiers for target behaviors. Each tuple includes a training account, a video identifier for a video, and a behavior identifier for the behavior between the training user account and the video. Server 100 constructs a hypergraph with user accounts, video identifiers, and behavior identifiers as nodes based on the multiple tuples, and inputs the hypergraph information into a hypergraph convolutional network to obtain the predicted embedding representation information of training users, videos, and behaviors in the training data. Server 100 then uses the first training tuple from any target tuple... The training user account and the first video identifier are used as positive training samples, where the target tuple is a tuple containing the behavior identifier of the target behavior. The second video identifier and the first training user account in any non-target tuple are used as negative training samples, where the non-target tuple is a tuple other than the target tuple. Then, the estimated embedding representation information, the positive training samples, and the negative training samples are respectively input into the recommendation network to obtain the first estimated recommendation information of the video corresponding to the first video identifier and the second estimated recommendation information of the video corresponding to the second video identifier. By maximizing the loss of the first estimated recommendation information and the second estimated recommendation information, the parameters of the hypergraph convolutional network and the recommendation network are adjusted to train the recommendation model.

[0049] After training the recommendation model, when recommending videos to users, server 100 can pre-screen a portion of the videos, i.e., screen out at least two videos to be recommended, obtain the user's account and the video identifiers of at least two videos to be recommended, and input the information of the constructed hypergraph into the hypergraph convolutional network to obtain the embedding representation information of the training user, video, and behavior in the training data; input the embedding representation information, user account, and video identifiers of at least two videos to be recommended into the recommendation network to obtain recommendation information for at least two videos to be recommended, wherein the recommendation information represents the probability that the user account's behavior towards the corresponding recommended object is the target behavior; based on the recommendation information, recommend at least two videos to be recommended to the user on user terminal 110 or 120, that is, more suitable videos can be recommended to the user.

[0050] The recommendation method and apparatus, and the recommendation model training method and apparatus according to exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings.

[0051] Figure 2 This is a flowchart illustrating a recommended method according to an exemplary embodiment, such as... Figure 2 As shown, the recommendation method is based on a recommendation model, which includes a hypergraph convolutional network and a recommendation network. The recommendation method includes the following steps:

[0052] In step S201, the user accounts of the objects to be recommended and the object identifiers of at least two objects to be recommended are obtained. It should be noted that this disclosure applies to the fine-tuning stage; therefore, a certain number of objects are pre-selected from the users of the objects to be recommended, namely the aforementioned at least two objects to be recommended, such as 100 objects pre-selected. These objects can be products, such as videos, music, etc.

[0053] In step S202, the hypergraph information is input into the hypergraph convolutional network to obtain the embedding representation information of training user accounts, objects, and behaviors in the training data. The training data includes multiple tuples, each containing a training user account, an object identifier for an object, and a behavior identifier for the behavior generated by the training user account on the object. The hypergraph is a graph constructed based on the multiple tuples in the training data, with training user accounts, object identifiers, and behavior identifiers as nodes. Edges of the hypergraph are lines connecting at least two nodes in a tuple. The tuples in the training data can be variable-length tuples, which can be represented as "user-object-behavior 1-behavior 2-...". Each variable-length tuple represents an interaction between the user and the object involving behavior 1, behavior 2, ... . The specific representation of the tuples can be aligned and recorded using multi-hot encoding. It should be noted that the above tuples only require the identifiers of the user, object, and behavior; no entity content of the user and object is needed. In this step, after obtaining the training data, each user, object, and behavior in the tuples of the training data is treated as a node in the graph, and each tuple in the training data is treated as a hyperedge. The training data is reconstructed into a complete hypergraph. The advantage of using this hypergraph for modeling is that it can effectively model complex semantic scenarios where multiple behaviors occur simultaneously. For example, in online systems, the behaviors of liking and forwarding are often associated. In this hypergraph, there will be multiple hyperedges connecting the nodes of both liking and forwarding behaviors. During the subsequent convolution process, there will be significant information propagation between the two nodes, resulting in strong similarity between the two nodes in the embedding space.

[0054] According to an exemplary embodiment of this disclosure, the hypergraph includes at least one of the following: a first hypergraph and a second hypergraph. The first and second hypergraphs are constructed as follows: a first hypergraph is constructed based on target tuples, with user account, object identifier, and behavior identifier as nodes, wherein the target tuple is a tuple containing the behavior identifier of the target behavior among multiple tuples; a second hypergraph is constructed based on non-target tuples, with user account, object identifier, and behavior identifier as nodes, wherein the non-target tuples are tuples other than the target tuple among multiple tuples. According to this embodiment, two hypergraphs are divided based on the target behavior, namely, a hypergraph containing the target behavior and a hypergraph not containing the target behavior. The hypergraphs are divided into a more relevant first hypergraph and a less relevant second hypergraph. Thus, when obtaining embedded representation information, the classification of relevance and irrelevance facilitates better subsequent learning of the embedded representation information of users, products, and behaviors, and helps the recommendation model learn key information for predicting target behaviors.

[0055] For example, the training data can be divided into three channels: Conversion, Auxiliary, and Convolution. These correspond to the first and second hypergraphs in the above embodiment. Then, convolution is performed on each hypergraph. Specifically, tuples (hyperedges) containing the target behavior are selected from the training data to construct a hypergraph—this is the Conversion channel. Tuples (hyperedges) not containing the target behavior are selected from the training data to construct a hypergraph—this is the Auxiliary channel. This channel division helps the model learn key information for predicting the target behavior. Inputting the information from both channels into the convolutional network yields their respective embedding representations. Combining these embedding representations provides a richer overall embedding representation.

[0056] According to an exemplary embodiment of this disclosure, when the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. Information from the hypergraph is input into the hypergraph convolutional network to obtain embedding representation information of training user accounts, objects, and behaviors in the training data. This includes: inputting information from the first hypergraph into the first hypergraph convolutional network to obtain first embedding representation information of training user accounts, objects, and behaviors in the training data; inputting information from the second hypergraph into the second hypergraph convolutional network to obtain second embedding representation information of training user accounts, objects, and behaviors in the training data; and obtaining embedding representation information based on the first embedding representation information, the second embedding representation information, and the weight coefficients. According to this embodiment, hypergraph convolution is performed on the information of the two hypergraphs respectively to obtain their respective embedding representation information. Combining these respective embedding representation information can yield richer overall embedding representation information, i.e., better embedding representation information can be learned. Furthermore, introducing learnable weight information can also assist in obtaining better embedding representation information.

[0057] For example, after constructing a hypergraph with two channels, a lightweight graph convolutional network (LightGCN) can be used, but is not limited to, to propagate information across the two channels. Information propagation for each channel can be performed using the following convolution formula:

[0058]

[0059] in, Let P represent the embedding representation matrix of the l-th layer. l Q l ,K l These are the embedding representation matrices for users, objects, and behaviors (i.e., the embedding representation information mentioned above), where t represents the channel and H... t It is the incidence matrix of the channel, D t The diagonalized node degree matrix, B t It is a diagonalized edge degree matrix.

[0060] Each channel is obtained after undergoing L layers of convolution. Weighted fusion is performed using learnable weight coefficients.

[0061]

[0062] Where, α t These are learnable parameters.

[0063] According to an exemplary embodiment of this disclosure, the hypergraph further includes a third hypergraph constructed as follows: a third hypergraph is constructed based on the target tuple, with user accounts and object identifiers as nodes. According to this embodiment, the third hypergraph is further partitioned based on the target behavior, adding information about the user and object corresponding to the target behavior, which helps to subsequently obtain better embedded representation information of users, products, and behaviors.

[0064] For example, in addition to dividing the training data into two channels, Conversion and Auxiliary, an additional Collaborative Filtering (CF) channel can be created, which is the third hypergraph in the above embodiment. Then, convolution is performed on each of the three channels. Specifically, tuples (hyperedges) containing the target behavior are selected from the training data to construct a hypergraph, which is the Conversion channel; tuples (hyperedges) not containing the target behavior are selected from the training data to construct a hypergraph, which is the Auxiliary channel; and user-item pairs of tuples that have exhibited the target behavior are selected from the training data, connected with a separate edge to construct a graph, which is the Collaborative Filtering channel. This channel division helps the model learn key information for predicting target behavior. Inputting the information from the three channels into the convolutional network yields their respective embedding representations. Combining these individual embedding representations results in a richer overall embedding representation.

[0065] According to an exemplary embodiment of this disclosure, the hypergraph convolutional network further includes a third hypergraph convolutional network. Information from the hypergraph is input into the third hypergraph convolutional network to obtain embedding representation information of training user accounts, objects, and behaviors in the training data. This includes: inputting information from the third hypergraph into the third hypergraph convolutional network to obtain third embedding representation information of training user accounts and objects in the training data; and obtaining embedding representation information based on the first embedding representation information, the second embedding representation information, the third embedding representation information, and weight coefficients. This includes: obtaining embedding representation information based on the first embedding representation information, the second embedding representation information, the third embedding representation information, and weight coefficients. According to this embodiment, hypergraph convolution is performed on the information of the three hypergraphs respectively to obtain their respective embedding representation information. Combining these respective embedding representation information can yield richer overall embedding representation information, i.e., better embedding representation information can be learned. Furthermore, introducing learnable weight information can also assist in obtaining better embedding representation information.

[0066] For example, after constructing a hypergraph with three channels, a hypergraph convolutional network in the form of LightGCN, but not limited to this form, can be used to propagate information in each of the three channels. Each channel can propagate information using the convolution formula (1) above. After obtaining their respective embedding representation information, they can be combined using weights. That is, each channel is obtained after undergoing L layers of convolution. Weighted fusion is performed using learnable weight coefficients.

[0067] return Figure 2In step S203, the embedded representation information, the user account, and the object identifiers of at least two objects to be recommended are input into the recommendation network to obtain recommendation information for each object to be recommended. The recommendation information represents the probability that the user account's behavior towards the corresponding object to be recommended is the target behavior, and the training data includes the target behavior. In this step, the 100 objects can be paired with users to obtain 100 user-object pairs. These 100 user-object pairs and the embedded representation information are then input into the recommendation model to obtain recommendation information for each object. For example, a scoring system can be used to determine the recommendation information for each object; that is, the recommendation information can be a score. Of course, any other form can also be used, and this disclosure does not limit this.

[0068] According to an exemplary embodiment of this disclosure, the recommendation network includes a first recommendation network and a second recommendation network. Embedded representation information, user accounts, and object identifiers of at least two objects to be recommended are input into the recommendation network to obtain recommendation information for each object to be recommended. This process may include: inputting the embedded representation information, user accounts, and object identifiers of at least two objects to be recommended into the first recommendation network to obtain first recommendation sub-information for each object to be recommended. The first recommendation sub-information is the probability that the user account's behavior towards the corresponding object to be recommended is a target behavior, based on information in the hypergraph from the user account and the object identifiers of at least two objects to be recommended; obtaining a set of object identifiers of objects that the user account has interacted with through the target behavior, and a set of user accounts of user accounts that each object to be recommended has interacted with through the target behavior; inputting the embedded representation information, the set of object identifiers, and each set of user accounts into the second recommendation network to obtain second recommendation sub-information for each object to be recommended. The second recommendation sub-information is the probability that the user account's behavior towards the corresponding object to be recommended is a target behavior, based on information in the hypergraph from the set of object identifiers and each set of user accounts; and obtaining recommendation information for each object to be recommended based on the first and second recommendation sub-information. According to this embodiment, on the one hand, global information is obtained through hypergraph convolution to determine the first estimated recommendation sub-information and the second estimated recommendation sub-information. On the other hand, the second estimated recommendation sub-information is determined through local collaborative filtering information (i.e., information that has been interacted with). Then, recommendation information is determined based on the two parts of recommendation sub-information, which takes into account both global and local information, thus increasing the accuracy of recommendation information.

[0069] Specifically, taking the recommendation information as the score as an example, the final scoring module (i.e., the recommendation network) can be divided into two parts:

[0070] The first part is multi-row global scoring, which is the process of obtaining the first estimated recommendation sub-information. The specific first recommendation network can be as follows:

[0071]

[0072] in, Let P represent the dot product, μ represent the user, ν represent the object, and P represent the dot product. L Q L They are E L The embedding representation matrices of users and objects are represented by the dot product of the embedding representation matrices of user μ and object ν. The goal of this part is to obtain global information by utilizing the previous hypergraph convolution.

[0073] The second part involves scoring using local collaborative filtering information, which is the process of obtaining the second predicted recommendation sub-information. The specific second recommendation network can be as follows:

[0074]

[0075] in, This indicates that user account u achieves target behavior b. T A collection of objects that have been interacted with. Represents object v through target behavior b T The set of user accounts that have interacted with each other, P and Q are the embedding representation matrices of user accounts and objects in E3 (i.e., the third embedding matrix information), respectively. This formula means that the information in the embedding representation matrix obtained from the third channel above is subjected to the corresponding dot product processing.

[0076] Ultimately, for user u, the target action b occurs on object v. T The predicted probability, that is, the degree to which user μ likes object ν, is a combination of the two parts mentioned above:

[0077]

[0078] in, It is a score of global structural information captured through a hypergraph. It is a scoring system based on local collaborative filtering information. By combining these two scores, efficient information recommendation can be achieved.

[0079] In step S204, at least two objects are recommended to the user's account based on the recommendation information. Assuming the recommendation information is a score, in this step, the 100 objects can be recommended to the user in order of their scores, or the top 50 highest-scoring objects can be selected from all objects and recommended to the user.

[0080] Figure 3 This is a flowchart illustrating a training method for a recommendation model according to an exemplary embodiment, such as... Figure 3 As shown, the recommendation model includes a hypergraph convolutional network and a recommendation network. The training method for the recommendation model includes the following steps:

[0081] In step S301, training data is acquired. This training data includes multiple tuples and behavior identifiers for target behaviors. Each tuple includes a training user account, an object identifier for an object, and a behavior identifier for the interaction between the training user account and the object. The tuples in the training data can be variable-length tuples, represented as "user-object-behavior 1-behavior 2-...". Each variable-length tuple represents an interaction between the user and the object involving behavior 1, behavior 2, ... . The specific representation of the tuples can be aligned and recorded using multi-hot encoding. The object can be a product, such as a video or music. It should be noted that the tuples only require the identifiers of the user, object, and behavior; no specific user or object entity content is needed.

[0082] In step S302, a hypergraph is constructed based on multiple tuples, with user accounts, object identifiers, and behavior identifiers as nodes. For example, after obtaining the training data, each user, object, and behavior in the tuples of the training data is regarded as a node in the graph, and each tuple in the training data is regarded as a hyperedge, thus reconstructing the training data into a complete hypergraph. The advantage of using this hypergraph for modeling is that it can effectively model complex semantic scenarios where multiple behaviors occur simultaneously. For example, in online systems, the behaviors of liking and forwarding are often associated. In this hypergraph, multiple hyperedges will connect the liking and forwarding behavior nodes simultaneously. During the subsequent convolution process, there will be significant information propagation between the two nodes, resulting in strong similarity between the two nodes in the embedding space.

[0083] According to an exemplary embodiment of this disclosure, constructing a hypergraph with user account, object identifier, and behavior identifier as nodes based on multiple tuples includes at least one of the following: constructing a first hypergraph with user account, object identifier, and behavior identifier as nodes based on target tuples, wherein the target tuple is a tuple containing the behavior identifier of the target behavior among the multiple tuples; and constructing a second hypergraph with user account, object identifier, and behavior identifier as nodes based on non-target tuples, wherein the non-target tuples are tuples other than the target tuple among the multiple tuples. According to this embodiment, two hypergraphs are divided based on the target behavior: a hypergraph containing the target behavior and a hypergraph not containing the target behavior. The hypergraphs are further divided into a more relevant first hypergraph and a less relevant second hypergraph. Thus, when obtaining embedded representation information, the classification of relevance and irrelevance facilitates better subsequent learning of the embedded representation information of users, products, and behaviors, and helps the recommendation model learn key information for predicting target behaviors.

[0084] According to an exemplary embodiment of this disclosure, a hypergraph is constructed based on multiple tuples, with user accounts, object identifiers, and behavior identifiers as nodes. The method further includes constructing a third hypergraph based on target tuples, with user accounts and object identifiers as nodes. According to this embodiment, the third hypergraph is further partitioned based on the target behavior, adding information about the user and object corresponding to the target behavior, thus assisting in obtaining better embedded representation information of users, products, and behaviors in subsequent learning.

[0085] In step S303, the information of the hypergraph is input into the hypergraph convolutional network to obtain the predicted embedding representation information of training user accounts, objects and behaviors in the training data.

[0086] According to an exemplary embodiment of this disclosure, when the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The hypergraph information is input into the hypergraph convolutional network to obtain estimated embedding representation information of user accounts, objects, and behaviors in the training data. This includes: inputting the information of the first hypergraph into the first hypergraph convolutional network to obtain first estimated embedding representation information of training user accounts, objects, and behaviors in the training data; inputting the information of the second hypergraph into the second hypergraph convolutional network to obtain second estimated embedding representation information of training user accounts, objects, and behaviors in the training data; and obtaining estimated embedding representation information based on the first estimated embedding representation information, the second estimated embedding representation information, and the weight coefficients. According to this embodiment, hypergraph convolution is performed on the information of the two hypergraphs respectively to obtain their respective embedding representation information. Combining these respective embedding representation information can yield richer overall embedding representation information, i.e., better embedding representation information can be learned. Furthermore, introducing learnable weight information can also assist in obtaining better embedding representation information.

[0087] According to an exemplary embodiment of this disclosure, the hypergraph convolutional network further includes a third hypergraph convolutional network. Information from the hypergraph is input into the third hypergraph convolutional network to obtain estimated embedding representation information of user accounts, objects, and behaviors in the training data. This includes: inputting information from the third hypergraph into the third hypergraph convolutional network to obtain third estimated embedding representation information of training user accounts and objects in the training data; and obtaining estimated embedding representation information based on the first estimated embedding representation information, the second estimated embedding representation information, and weight coefficients. This includes: obtaining estimated embedding representation information based on the first estimated embedding representation information, the second embedded representation information, the third embedded representation information, and weight coefficients. According to this embodiment, hypergraph convolution is performed on the information of the three hypergraphs respectively to obtain their respective embedding representation information. Combining these respective embedding representation information can yield richer overall embedding representation information, i.e., better embedding representation information can be learned. Furthermore, introducing learnable weight information can also assist in obtaining better embedding representation information.

[0088] return Figure 3 In step S304, the first training user account and the first object identifier in any target tuple are used as positive training samples. Here, a target tuple is a tuple containing the behavior identifier of the target behavior among multiple tuples. For example, "user-object" pairs can be selected from any tuple containing the target behavior in the training data and used as positive training samples.

[0089] In step S305, the second object identifier and the first training user account in any non-target tuple are used as training negative samples, where the non-target tuple is any tuple other than the target tuple. For example, for each training positive sample, a negative sample can be performed simultaneously, that is, an object that has not performed the target behavior is randomly selected for the user to form a "user-object" pair as a training negative sample.

[0090] In step S306, the training positive samples and the estimated embedding representation information are input into the recommendation network to obtain the first estimated recommendation information of the object corresponding to the first object identifier. The first estimated recommendation information represents the probability that the action taken by the first training user account towards the object corresponding to the first object identifier is the target action. For example, it can be in the form of a scoring system, i.e., the first estimated recommendation information can be a score. Of course, it can also be in any other form, and this disclosure does not limit it in this regard.

[0091] In step S307, the training negative samples and the estimated embedding representation information are input into the recommendation network to obtain the second estimated recommendation information for the object corresponding to the second object identifier. The second estimated recommendation information represents the probability that the action taken by the first training user account towards the object corresponding to the second object identifier is the target action. For example, it can be in the form of a scoring system, i.e., the second estimated recommendation information can be a score. Of course, it can also take any other form, and this disclosure does not limit it in this regard.

[0092] According to an exemplary embodiment of this disclosure, the recommendation network includes a first recommendation network and a second recommendation network. Predicted embedding representation information, training positive samples, and training negative samples are input into the recommendation network to obtain first predicted recommendation information for objects corresponding to a first object identifier and second predicted recommendation information for objects corresponding to a second object identifier. This includes: inputting predicted embedding representation information and training positive samples into the first recommendation network to obtain first predicted recommendation sub-information; inputting predicted embedding representation information and training negative samples into the first recommendation network to obtain second predicted recommendation sub-information; and obtaining the set of object identifiers of objects that the user corresponding to the first user identifier has interacted with through target behavior, and the set of object identifiers of objects that the first object identifier has interacted with through target behavior. The objects corresponding to the first user identifier set and the second object identifier of the users who have interacted with the target behavior are obtained through the second user identifier set of the users who have interacted with the target behavior. The third predicted embedding representation information, the object identifier set, and the first user identifier set are input into the second recommendation network to obtain the third predicted recommendation sub-information for training positive samples. The third predicted embedding representation information, the object identifier set, and the second user identifier set are input into the second recommendation network to obtain the fourth predicted recommendation sub-information for training negative samples. Based on the first and third predicted recommendation sub-information, the first predicted recommendation information is obtained. Based on the second and fourth predicted recommendation sub-information, the second predicted recommendation information is obtained. According to this embodiment, on the one hand, global information is obtained through hypergraph convolution to determine the first and second predicted recommendation sub-information; on the other hand, the third and fourth predicted recommendation sub-information are determined through local collaborative filtering information (i.e., information about interactions). Then, the first and second recommendation information are determined based on these two parts of recommendation sub-information, thus considering both global and local information, increasing the accuracy of the recommendation information.

[0093] return Figure 3 In step S308, the parameters of the hypergraph convolutional network and the recommendation network are adjusted by maximizing the loss of the first and second estimated recommendation information, and the recommendation model is trained.

[0094] According to an exemplary embodiment of this disclosure, the training data further includes the actual conversion rate of each behavior in the training data, where the actual conversion rate is the probability that each behavior is actually converted into the target behavior; the recommendation model is trained by adjusting the parameters of the hypergraph convolutional network and the recommendation network by maximizing the loss of the first estimated recommendation information and the second estimated recommendation information, including: determining a first loss based on the first estimated recommendation information and the second estimated recommendation information; obtaining the estimated conversion rate of the intermediate layer output of the recommendation model, wherein the estimated conversion rate is the probability that each behavior in the training data is estimated to be converted into the target behavior; determining a second loss based on the estimated conversion rate and the actual conversion rate; and adjusting the parameters of the hypergraph convolutional network and the recommendation network by maximizing the first loss and minimizing the second loss, thereby training the recommendation model. According to this embodiment, the probability of a behavior being converted into a target behavior in the training data is used to assist in the training of the recommendation model. Since the target behavior can be set as needed, such as the behavior performed by a user when watching a video they are interested in, using the probability that the behavior is the target behavior to assist in training the recommendation model can make the recommended objects obtained based on the trained recommendation model more in line with the user's needs.

[0095] For example, training a recommendation model can be divided into two parts for collaborative training:

[0096] The first part is the main recommendation task. This part can be trained using the traditional positive and negative sampling and Bayesian Personalized Ranking (BPR) loss function. That is, the recommendation model is optimized by maximizing the score difference between training positive samples and training negative samples. At the same time, the L2 norm of all parameters of the recommendation model can be constrained, as follows:

[0097]

[0098] Among them, y ui It is the score of the training positive samples, y uj σ is the score of the training negative samples, λ and σ are hyperparameters, and Θ represents all parameters of the recommendation model.

[0099] The second part is the auxiliary conversion rate prediction task, which utilizes the behavior embedding representation matrix K. L The predictor forecasts the probability that each action will transform into the target action. It is a single-layer Multilayer Perceptron (MLP), and the loss function is... Mean Square Error (MSE) can be used.

[0100] The final loss function is a weighted sum of the two loss functions:

[0101]

[0102] Here, β is a hyperparameter that controls the scale of the auxiliary task.

[0103] It should be noted that, Figure 4 This is a schematic diagram illustrating the structure of a recommendation model according to an exemplary embodiment. The overall structure of the recommendation model can be as follows: Figure 4 As shown, μ1, μ2, and μ3 represent user 1, user 2, and user 3 in the training data. The lines representing objects 1, 2, 3, and 4 in the training data, with different shades of gray representing the interactions between them, and ε... c ε a ε f This represents the conversion channel, auxiliary channel, and collaborative filtering channel; b1, b2, and b3 represent behavior 1, behavior 2, and behavior 3, respectively. Let E represent the first embedding representation matrix, the second embedding representation matrix, and the third embedding representation matrix. L The embedding representation matrix is ​​used to represent the matrix. R represents a predictor, which predicts the probability that an action in the training data is the target action. b P represents the actual probability that an action in the training data is the target action. L Q L E L Embedding representation matrix of users and objects in the middle, represents the embedding representation matrices of users and objects obtained by convolution on the collaborative filtering channels, respectively.

[0104] To verify the above embodiments, this disclosure was tested on two real-world online datasets. Compared to state-of-the-art baseline models, this disclosure achieves a performance improvement of 1.96%–23.19%. On cold-start users (those with fewer than 5 target behavior interactions), the performance of this disclosure can reach up to 28.19%, meaning that the recommendation model of this disclosure significantly improves the recommendation effect based on target behavior under complex semantics. Specifically, compared to other multi-behavior recommendation models, the recommendation model of this disclosure embeds behavior nodes in the graph. The representation of behavior can perceive the global structural information of the entire dataset and possesses complex semantics. Through constraints from downstream tasks or the design of self-supervised auxiliary tasks, recommendation performance can also be improved to assist downstream tasks.

[0105] In summary, this disclosure treats each behavior as a node in a graph, redefines the multi-behavior recommendation problem using a hypergraph framework, and completes the encoding of multiple behaviors. Furthermore, by dividing the graph into three channels—conversion, auxiliary, and collaborative filtering—and performing convolutions, it efficiently learns the embedded representations of user-item-behavior information. Finally, a self-supervised learning module is introduced to assist the recommendation model in predicting the conversion rate of the target behavior through behavior representations, thus helping it better learn user preferences and object features. Therefore, this disclosure efficiently encodes multi-behavior interaction data and learns the embedded representation of each behavior in multi-behavior recommendation scenarios with complex semantics. Using this embedded representation, target behavior recommendations can be efficiently performed, improving the recommendation performance of target behaviors in multi-behavior recommendation scenarios and thereby enhancing the accuracy of object recommendations.

[0106] Figure 5 This is a block diagram illustrating a recommended apparatus according to an exemplary embodiment, such as Figure 5 As shown, the recommendation device is implemented based on a recommendation model, which includes a hypergraph convolutional network and a recommendation network. The recommendation device includes:

[0107] The identifier acquisition unit 50 is configured to acquire the user account of the object to be recommended and the object identifiers of at least two objects to be recommended; the embedding representation information acquisition unit 52 is configured to input the information of the hypergraph into the hypergraph convolutional network to obtain the embedding representation information of the training user account, object and behavior in the training data, wherein the training data includes multiple tuples, each tuple including a training user account, an object identifier of an object, and a behavior identifier of the behavior generated by the training user account on the object, the hypergraph is a graph constructed based on multiple tuples in the training data with the training user account, object identifier and behavior identifier as nodes, and the edges of the hypergraph are lines connecting at least two nodes of the training user account, object identifier and behavior identifier in a tuple; the recommendation information acquisition unit 54 is configured to input the embedding representation information, user account and the object identifiers of at least two objects to be recommended into the recommendation network to obtain the recommendation information of each object to be recommended, wherein the recommendation information represents the probability that the behavior generated by the user account on the corresponding object to be recommended is the target behavior, and the training data includes the target behavior; the recommendation unit 56 is configured to recommend at least two objects to be recommended to the user account based on the recommendation information.

[0108] According to an exemplary embodiment of this disclosure, the hypergraph includes at least one of the following: a first hypergraph and a second hypergraph, wherein the first hypergraph and the second hypergraph are constructed in the following manner: a first hypergraph is constructed based on a target tuple with user account, object identifier, and behavior identifier as nodes, wherein the target tuple is a tuple containing a behavior identifier of a target behavior among multiple tuples; and a second hypergraph is constructed based on non-target tuples with user account, object identifier, and behavior identifier as nodes, wherein the non-target tuples are tuples other than the target tuple among multiple tuples.

[0109] According to an exemplary embodiment of this disclosure, when the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The embedding representation information acquisition unit 52 is further configured to input information of the first hypergraph into the first hypergraph convolutional network to obtain first embedding representation information of training user accounts, objects, and behaviors in the training data; input information of the second hypergraph into the second hypergraph convolutional network to obtain second embedding representation information of training user accounts, objects, and behaviors in the training data; and obtain embedding representation information based on the first embedding representation information, the second embedding representation information, and the weight coefficients.

[0110] According to an exemplary embodiment of this disclosure, the hypergraph also includes a third hypergraph constructed in such a way as to construct a third hypergraph with user account and object identifier as nodes based on the target tuple.

[0111] According to an exemplary embodiment of the present disclosure, the hypergraph convolutional network further includes a third hypergraph convolutional network. The embedding representation information acquisition unit 52 is further configured to input the information of the third hypergraph into the third hypergraph convolutional network to obtain the third embedding representation information of the training user accounts and objects in the training data; and to obtain the embedding representation information based on the first embedding representation information, the second embedding representation information, the third embedding representation information and the weight coefficients.

[0112] According to an exemplary embodiment of this disclosure, the recommendation network includes a first recommendation network and a second recommendation network. The recommendation information acquisition unit 54 is configured to input embedded representation information, user accounts, and object identifiers of at least two objects to be recommended into the first recommendation network to obtain first recommendation sub-information for each object to be recommended. The first recommendation sub-information is the probability that the behavior of a user account towards the corresponding object to be recommended is the target behavior, based on information in the hypergraph of the user account and the object identifiers of at least two objects to be recommended. The unit is also configured to acquire a set of object identifiers of objects that the user account has interacted with through the target behavior and a set of user accounts of user accounts that have interacted with each object to be recommended through the target behavior. The unit further configures to input embedded representation information, the set of object identifiers, and the set of user accounts into the second recommendation network to obtain second recommendation sub-information for each object to be recommended. The second recommendation sub-information is the probability that the behavior of a user account towards the corresponding object to be recommended is the target behavior, based on information in the hypergraph of the set of object identifiers and the set of user accounts. Finally, the unit is configured to obtain recommendation information for each object to be recommended.

[0113] Figure 6 This is a block diagram illustrating a training apparatus for a recommendation model according to an exemplary embodiment, such as... Figure 6 As shown, the recommendation model includes a hypergraph convolutional network and a recommendation network. The training device for the recommendation model includes:

[0114] Training data acquisition unit 60 is configured to acquire training data, which includes multiple tuples and behavior identifiers of target behaviors. Each tuple includes a training user account, an object identifier of an object, and a behavior identifier of the behavior between the training user account and the object. Hypergraph acquisition unit 62 is configured to construct a hypergraph with user accounts, object identifiers, and behavior identifiers as nodes based on multiple tuples. Estimated embedding representation information acquisition unit 64 is configured to input the hypergraph information into a hypergraph convolutional network to obtain the estimated embedding representation information of training user accounts, objects, and behaviors in the training data. Positive sample acquisition unit 66 is configured to use the first training user account and the first object identifier in any target tuple as training positive samples, where the target tuple is a tuple containing the behavior identifier of the target behavior. Negative sample acquisition unit 68 is configured to use the second object identifier and the first training user account in any non-target tuple as training negative samples. Among them, non-target tuples are tuples other than target tuples among multiple tuples; the first predicted recommendation information acquisition unit 610 is configured to input training positive samples and predicted embedding representation information into the recommendation network to obtain the first predicted recommendation information of the object corresponding to the first object identifier, wherein the first predicted recommendation information represents the probability that the behavior generated by the first training user account on the object corresponding to the first object identifier is the target behavior; the second predicted recommendation information acquisition unit 612 is configured to input training negative samples and predicted embedding representation information into the recommendation network to obtain the second predicted recommendation information of the object corresponding to the second object identifier, wherein the second predicted recommendation information represents the probability that the behavior generated by the first training user account on the object corresponding to the second object identifier is the target behavior; the training unit 614 is configured to train the recommendation model by maximizing the loss of the first predicted recommendation information and the second predicted recommendation information and adjusting the parameters of the hypergraph convolutional network and the recommendation network.

[0115] According to an exemplary embodiment of this disclosure, the hypergraph acquisition unit 62 is further configured to: construct a first hypergraph based on a target tuple with user account, object identifier, and behavior identifier as nodes, wherein the target tuple is a tuple containing a behavior identifier of a target behavior among multiple tuples; and construct a second hypergraph based on non-target tuples with user account, object identifier, and behavior identifier as nodes, wherein the non-target tuples are tuples other than the target tuple among multiple tuples.

[0116] According to an exemplary embodiment of this disclosure, when the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The estimated embedding representation information acquisition unit 64 is further configured to input information of the first hypergraph into the first hypergraph convolutional network to obtain first estimated embedding representation information of training user accounts, objects, and behaviors in the training data; input information of the second hypergraph into the second hypergraph convolutional network to obtain second estimated embedding representation information of training user accounts, objects, and behaviors in the training data; and obtain estimated embedding representation information based on the first estimated embedding representation information, the second estimated embedding representation information, and the weight coefficients.

[0117] According to an exemplary embodiment of this disclosure, the hypergraph acquisition unit 62 is further configured to construct a third hypergraph based on the target tuple, with user accounts and object identifiers as nodes.

[0118] According to an exemplary embodiment of the present disclosure, the hypergraph convolutional network further includes a third hypergraph convolutional network. The estimated embedding representation information acquisition unit 64 is further configured to input the information of the third hypergraph into the third hypergraph convolutional network to obtain the third estimated embedding representation information of the training user accounts and objects in the training data; and to obtain the estimated embedding representation information based on the first estimated embedding representation information, the second estimated embedding representation information, the third estimated embedding representation information and the weight coefficients.

[0119] According to an exemplary embodiment of this disclosure, the training data further includes the actual conversion rate of each behavior in the training data, the actual conversion rate being the probability that each behavior is actually converted into the target behavior; the training unit 614 is further configured to determine a first loss based on a first estimated recommendation information and a second estimated recommendation information; obtain the estimated conversion rate of the intermediate layer output of the recommendation model, wherein the estimated conversion rate is the probability that each behavior in the training data is estimated to be converted into the target behavior; determine a second loss based on the estimated conversion rate and the actual conversion rate; and train the recommendation model by adjusting the parameters of the hypergraph convolutional network and the recommendation network based on maximizing the first loss and minimizing the second loss.

[0120] According to embodiments of this disclosure, an electronic device may be provided. Figure 7 This is a block diagram of an electronic device 700 according to an embodiment of the present disclosure. The electronic device includes at least one memory 701 and at least one processor 702. The at least one memory stores a set of computer-executable instructions. When the set of computer-executable instructions is executed by the at least one processor, a training method and / or a recommendation method for a recommended model according to an embodiment of the present disclosure are executed.

[0121] As an example, electronic device 700 may be a PC, tablet, personal digital assistant, smartphone, or other device capable of executing the aforementioned set of instructions. Here, electronic device 1000 is not necessarily a single electronic device, but may be a collection of any devices or circuits capable of executing the aforementioned instructions (or instruction sets) individually or in combination. Electronic device 700 may also be part of an integrated control system or system manager, or may be configured to interconnect with a portable electronic device locally or remotely (e.g., via wireless transmission) through an interface.

[0122] In electronic device 700, processor 702 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, processor 702 may also include analog processors, digital processors, microprocessors, multi-core processors, processor arrays, network processors, etc.

[0123] The processor 702 can execute instructions or code stored in memory, wherein memory 701 can also store data. Instructions and data can also be sent and received via a network through a network interface device, wherein the network interface device can employ any known transmission protocol.

[0124] The memory 701 can be integrated with the processor 702, for example, by placing RAM or flash memory within an integrated circuit microprocessor. Alternatively, the memory 701 can include a separate device, such as an external disk drive, a storage array, or other storage device usable by any database system. The memory 701 and the processor 702 can be operatively coupled, or can communicate with each other, for example, via I / O ports, network connections, etc., enabling the processor 702 to read files stored in the memory 701.

[0125] In addition, the electronic device 700 may also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, mouse, touch input device, etc.). All components of the electronic device can be interconnected via a bus and / or network.

[0126] According to embodiments of this disclosure, a computer-readable storage medium may also be provided, wherein when instructions in the computer-readable storage medium are executed by at least one processor, the at least one processor causes to execute a training method and / or a recommendation method for a recommended model according to embodiments of this disclosure. Examples of computer-readable storage media include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disc storage, hard disk drive (HDD), solid-state drive (SSD), card storage (such as multimedia cards, secure digital (SD) cards, or ultra-fast digital (XD) cards), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, and any other device configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and to provide the computer program and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the aforementioned computer-readable storage medium can run in an environment deployed in computer devices such as clients, hosts, agent devices, servers, etc. Furthermore, in one example, the computer program and any associated data, data files, and data structures are distributed across a networked computer system, such that the computer program and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner through one or more processors or computers.

[0127] According to embodiments of this disclosure, a computer program product is provided, including computer instructions, which, when executed by a processor, implement a training method and / or a recommendation method for a recommendation model according to embodiments of this disclosure.

[0128] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0129] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A recommendation method, characterized in that, The recommendation method is implemented based on a recommendation model, which includes a hypergraph convolutional network and a recommendation network. The recommendation method includes: Obtain the user account of the object to be recommended and the object identifiers of at least two objects to be recommended; The information of the hypergraph is input into the hypergraph convolutional network to obtain the embedding representation information of training user accounts, objects and behaviors in the training data. The training data includes multiple tuples, each tuple including a training user account, an object identifier of an object, and a behavior identifier of the behavior generated by the training user account on the object. The hypergraph is a graph constructed based on multiple tuples in the training data, with training user accounts, object identifiers and behavior identifiers as nodes. The edges of the hypergraph are lines connecting at least two nodes of the training user account, object identifier and behavior identifier in a tuple. The embedded representation information, the user account, and the object identifiers of the at least two objects to be recommended are input into the recommendation network to obtain recommendation information for each object to be recommended. The recommendation information represents the probability that the user account's behavior toward the corresponding object to be recommended is the target behavior. The training data includes the target behavior. Based on the recommendation information, the user account is recommended at least two objects to be recommended.

2. The recommended method as described in claim 1, characterized in that, The hypergraph includes at least one of the following: a first hypergraph and a second hypergraph, wherein the first hypergraph and the second hypergraph are constructed in the following manner: A first hypergraph is constructed based on the target tuple, with user account, object identifier and behavior identifier as nodes, wherein the target tuple is a tuple containing the behavior identifier of the target behavior among the plurality of tuples; A second hypergraph is constructed based on non-target tuples, with user accounts, object identifiers, and behavior identifiers as nodes, wherein the non-target tuples are tuples other than the target tuples among the plurality of tuples.

3. The recommended method as described in claim 2, characterized in that, In the case where the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The step of inputting the hypergraph information into the hypergraph convolutional network to obtain the embedding representation information of training user accounts, objects, and behaviors in the training data includes: The information of the first hypergraph is input into the first hypergraph convolutional network to obtain the first embedding representation information of the training user accounts, objects and behaviors in the training data; The information of the second hypergraph is input into the second hypergraph convolutional network to obtain the second embedding representation information of the training user accounts, objects and behaviors in the training data; The embedding representation information is obtained based on the first embedding representation information, the second embedding representation information, and the weight coefficients.

4. The recommended method as described in claim 3, characterized in that, The hypergraph also includes a third hypergraph constructed in the following manner: Construct a third hypergraph based on the target tuple, with user accounts and object identifiers as nodes.

5. The recommended method as described in claim 4, characterized in that, The hypergraph convolutional network also includes a third hypergraph convolutional network. The step of inputting the hypergraph information into the hypergraph convolutional network to obtain the embedding representation information of training user accounts, objects, and behaviors in the training data includes: The information of the third hypergraph is input into the third hypergraph convolutional network to obtain the third embedding representation information of training user accounts and objects in the training data; The step of obtaining the embedding representation information based on the first embedding representation information, the second embedding representation information, and the weight coefficients includes: The embedding representation information is obtained based on the first embedding representation information, the second embedding representation information, the third embedding representation information, and the weight coefficients.

6. The recommended method as described in claim 2, characterized in that, The recommendation network includes a first recommendation network and a second recommendation network. The step of inputting the embedded representation information, the user account, and the object identifiers of at least two objects to be recommended into the recommendation network to obtain recommendation information for each object to be recommended includes: The embedded representation information, the user account, and the object identifiers of the at least two objects to be recommended are input into the first recommendation network to obtain the first recommendation sub-information for each object to be recommended. The first recommendation sub-information is the probability that the user account’s behavior toward the corresponding object to be recommended is the target behavior, based on the information of the user account and the object identifiers of the at least two objects to be recommended in the hypergraph. Obtain the set of object identifiers of objects that the user account has interacted with through the target behavior, and the set of user accounts of each user account that the recommended object has interacted with through the target behavior; The embedded representation information, the set of object identifiers, and each set of user accounts are input into the second recommendation network to obtain the second recommendation sub-information for each object to be recommended. The second recommendation sub-information is the probability that the behavior of the user account on the corresponding object to be recommended is the target behavior, based on the information of the set of object identifiers and each set of user accounts in the hypergraph. Based on the first recommendation sub-information and the second recommendation sub-information, recommendation information for each object to be recommended is obtained.

7. A method for training a recommendation model, characterized in that, The recommendation model includes a hypergraph convolutional network and a recommendation network, and the training method includes: Acquire training data, wherein the training data includes multiple tuples and behavior identifiers of target behaviors, each tuple including a training user account, an object identifier of an object, and a behavior identifier of the behavior between the training user account and the object; A hypergraph is constructed based on the aforementioned tuples, with user accounts, object identifiers, and behavior identifiers as nodes; The information of the hypergraph is input into the hypergraph convolutional network to obtain the predicted embedding representation information of training user accounts, objects and behaviors in the training data; The first training user account and the first object identifier in any target tuple are used as training positive samples, wherein the target tuple is a tuple that contains the behavior identifier of the target behavior in the plurality of tuples; The second object identifier and the first training user account in any non-target tuple are used as training negative samples, wherein the non-target tuple is a tuple other than the target tuple among the plurality of tuples; The training positive samples and the estimated embedding representation information are input into the recommendation network to obtain the first estimated recommendation information of the object corresponding to the first object identifier, wherein the first estimated recommendation information represents the probability that the behavior generated by the first training user account on the object corresponding to the first object identifier is the target behavior; The training negative samples and the estimated embedding representation information are respectively input into the recommendation network to obtain the second estimated recommendation information of the object corresponding to the second object identifier, wherein the second estimated recommendation information represents the probability that the behavior generated by the first training user account on the object corresponding to the second object identifier is the target behavior; The recommendation model is trained by adjusting the parameters of the hypergraph convolutional network and the recommendation network by maximizing the loss of the first and second estimated recommendation information.

8. The training method as described in claim 7, characterized in that, The construction of a hypergraph based on the multiple tuples, with user accounts, object identifiers, and behavior identifiers as nodes, includes at least one of the following: A first hypergraph is constructed based on the target tuple, with user account, object identifier, and behavior identifier as nodes, wherein the target tuple is a tuple containing the behavior identifier of the target behavior among the plurality of tuples; A second hypergraph is constructed based on the non-target tuples, with user account, object identifier, and behavior identifier as nodes, wherein the non-target tuples are tuples other than the target tuple among the plurality of tuples.

9. The training method as described in claim 8, characterized in that, In the case where the hypergraph includes a first hypergraph and a second hypergraph, the hypergraph convolutional network includes a first hypergraph convolutional network, a second hypergraph convolutional network, and weight coefficients. The step of inputting the hypergraph information into the hypergraph convolutional network to obtain the estimated embedding representation information of user accounts, objects, and behaviors in the training data includes: The information of the first hypergraph is input into the first hypergraph convolutional network to obtain the first estimated embedding representation information of the training user accounts, objects and behaviors in the training data; The information of the second hypergraph is input into the second hypergraph convolutional network to obtain the second estimated embedding representation information of the training user accounts, objects and behaviors in the training data; The estimated embedding representation information is obtained based on the first estimated embedding representation information, the second estimated embedding representation information, and the weight coefficients.

10. The training method as described in claim 9, characterized in that, The construction of a hypergraph based on the multiple tuples, with user accounts, object identifiers, and behavior identifiers as nodes, further includes: Construct a third hypergraph based on the target tuple, with user accounts and object identifiers as nodes.

11. The training method as described in claim 10, characterized in that, The hypergraph convolutional network also includes a third hypergraph convolutional network. The step of inputting the hypergraph information into the hypergraph convolutional network to obtain the estimated embedding representation information of user accounts, objects, and behaviors in the training data includes: The information of the third hypergraph is input into the third hypergraph convolutional network to obtain the third estimated embedding representation information of training user accounts and objects in the training data; The step of obtaining the estimated embedding representation information based on the first estimated embedding representation information, the second estimated embedding representation information, and the weight coefficients includes: The estimated embedding representation information is obtained based on the first estimated embedding representation information, the second estimated embedding representation information, the third estimated embedding representation information, and the weight coefficients.

12. The training method as described in claim 7, characterized in that, The training data also includes the actual conversion rate of each behavior in the training data, where the actual conversion rate is the probability that each behavior is actually converted into the target behavior; The step of training the recommendation model by maximizing the loss of the first estimated recommendation information and the second estimated recommendation information, and adjusting the parameters of the hypergraph convolutional network and the recommendation network, includes: The first loss is determined based on the first estimated recommendation information and the second estimated recommendation information; Obtain the estimated conversion rate of the intermediate layer output of the recommendation model, wherein the estimated conversion rate is the probability that each behavior in the training data is estimated to be converted into the target behavior; The second loss is determined based on the estimated conversion rate and the actual conversion rate; The parameters of the hypergraph convolutional network and the recommendation network are adjusted based on maximizing the first loss and minimizing the second loss, and the recommendation model is trained.

13. A recommendation device, characterized in that, The recommendation device is implemented based on a recommendation model, which includes a hypergraph convolutional network and a recommendation network. The recommendation device includes: The identifier acquisition unit is configured to acquire the user account of the object to be recommended and at least two object identifiers of the object to be recommended. An embedding representation information acquisition unit is configured to input information of the hypergraph into the hypergraph convolutional network to obtain embedding representation information of training user accounts, objects, and behaviors in the training data. The training data includes multiple tuples, each tuple including a training user account, an object identifier of an object, and a behavior identifier of the behavior generated by the training user account on the object. The hypergraph is a graph constructed based on multiple tuples in the training data, with training user accounts, object identifiers, and behavior identifiers as nodes. The edges of the hypergraph are lines connecting at least two nodes of the training user account, object identifier, and behavior identifier in a tuple. The recommendation information acquisition unit is configured to input the embedded representation information, the user account, and the object identifiers of at least two objects to be recommended into the recommendation network to obtain recommendation information for each object to be recommended, wherein the recommendation information represents the probability that the user account's behavior toward the corresponding object to be recommended is the target behavior, and the training data includes the target behavior; The recommendation unit is configured to recommend the at least two objects to be recommended to the user account based on the recommendation information.

14. A training device for a recommendation model, characterized in that, The recommendation model includes a hypergraph convolutional network and a recommendation network, and the training device includes: The training data acquisition unit is configured to acquire training data, wherein the training data includes multiple tuples and behavior identifiers of target behaviors, and each tuple includes a training user account, an object identifier of an object, and a behavior identifier of the behavior between the training user account and the object; The hypergraph acquisition unit is configured to construct a hypergraph with user account, object identifier, and behavior identifier as nodes based on the multiple tuples; The estimated embedding representation information acquisition unit is configured to input the information of the hypergraph into the hypergraph convolutional network to obtain the estimated embedding representation information of training user accounts, objects and behaviors in the training data; The positive sample acquisition unit is configured to use the first training user account and the first object identifier in any target tuple as training positive samples, wherein the target tuple is a tuple containing the behavior identifier of the target behavior in the plurality of tuples; The negative sample acquisition unit is configured to use the second object identifier and the first training user account in any non-target tuple as training negative samples, wherein the non-target tuple is a tuple other than the target tuple among the plurality of tuples. The first estimated recommendation information acquisition unit is configured to input the training positive samples and the estimated embedding representation information into the recommendation network to obtain the first estimated recommendation information of the object corresponding to the first object identifier, wherein the first estimated recommendation information represents the probability that the behavior generated by the first training user account for the object corresponding to the first object identifier is the target behavior; The second estimated recommendation information acquisition unit is configured to input the training negative samples and the estimated embedding representation information into the recommendation network to obtain the second estimated recommendation information of the object corresponding to the second object identifier, wherein the second estimated recommendation information represents the probability that the behavior generated by the first training user account for the object corresponding to the second object identifier is the target behavior; The training unit is configured to train the recommendation model by adjusting the parameters of the hypergraph convolutional network and the recommendation network by maximizing the loss of the first estimated recommendation information and the second estimated recommendation information.

15. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the recommendation method as described in any one of claims 1 to 6 and / or the training method of the recommendation model as described in any one of claims 7 to 12.

16. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by at least one processor, the at least one processor causes the processor to perform the recommended method as described in any one of claims 1 to 6 and / or the training method of the recommended model as described in any one of claims 7 to 12.

17. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the recommendation method as described in any one of claims 1 to 6 and / or the training method of the recommendation model as described in any one of claims 7 to 12.

Images