Recommendation processing method, server and medium based on behavior-aware graph neural network

Abstract

The present invention discloses a recommendation processing method based on a behavior perception graph neural network, a server and a medium, which specifically include: dividing each acquired heterogeneous behavior session into a target behavior sequence and an auxiliary behavior sequence. Based on the target behavior sequence, the target behavior sequence item representation is obtained by learning a pre-constructed isomorphic behavior transition graph, and the target behavior sequence representation is calculated by the target behavior sequence item representation; based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning a pre-constructed isomorphic behavior transition graph and a heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is further calculated; based on the target behavior sequence representation and the auxiliary behavior sequence representation, the session representation is fused, and the predicted score of the item is further obtained, and the item is recommended to the user according to the predicted score. By constructing a heterogeneous behavior transition graph, the contribution of the auxiliary behavior sequence to the target behavior sequence is made clearer, thereby improving the accuracy of the recommendation.

Description

Technical Field

[0001] The present invention relates to the field of data mining technology, and in particular to a recommendation processing method, server and medium based on a behavior-aware graph neural network. Background Art

[0002] With the continuous development of big data, a huge amount of data is generated on the Internet every day. In order to cope with the problem of information overload in the era of information explosion, people have proposed the concept of intelligent recommendation. Among the different research branches of recommendation systems, session-based recommendation aims to recommend the next possible interactive item to users based on the user's behavior sequence in the session.

[0003] Traditional session-based recommendation only uses a single type of user behavior sequence, which often faces the problem of data sparsity. For example, in e-commerce scenarios, the ultimate goal of e-commerce companies is to guide users to buy goods to make profits, and the data of purchase behavior is very sparse. Since users' interests are diverse and purchase data is sparse, using only traditional session-based recommendation algorithms to model purchase sequences will lead to poor performance and fail to accurately recommend items that users want to buy. Fortunately, the behavior sequences of sessions in real-world e-commerce platforms are usually heterogeneous, that is, they consist of more than one type of behavior, such as purchasing and browsing. Based on this, the method that uses the richer information in the heterogeneous behavior sequences of sessions to improve the recommendation performance of predicting target behaviors is called heterogeneous session-based recommendation (HSBR).

[0004] However, existing research on HSBR focuses on how to distinguish different types of behaviors in sequence modeling or utilize homogeneous behavior transitions within each behavior sequence, and none of them explicitly utilizes heterogeneous behavior transitions between different types of behavior sequences. For example, users usually browse some items for comparison before making a purchase decision, which means that the transitions between browsing behaviors represent similarities, and the transition from browsing to purchasing may represent a progressive relationship. Therefore, the existing HSBR may fail to capture the semantic relationship between the transitions between different types of behaviors, which may also lead to the problem of low recommendation accuracy.

[0005] As can be seen from the above, the session-based heterogeneous recommendation in the existing technology still cannot provide high-precision item recommendations, that is, target behavior prediction.

[0006] Therefore, the existing technology still needs to be improved and developed. Summary of the invention

[0007] In the existing technologies, session-based heterogeneous recommendations all focus on studying how to distinguish different types of behaviors in sequence modeling or utilize isomorphic behavior transitions within each behavior sequence, but do not explicitly utilize heterogeneous behavior transitions between different types of behavior sequences. As a result, the semantic relationship between transitions between different types of behaviors cannot be captured, resulting in unsatisfactory recommendation performance.

[0008] The present invention provides a recommendation processing method based on a behavior-aware graph neural network, specifically comprising: obtaining each heterogeneous behavior session, and dividing each heterogeneous behavior session into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing two or more users' different behaviors on items. Based on the target behavior sequence, the target behavior sequence item representation is obtained by learning a pre-constructed isomorphic behavior transition graph, and the target behavior sequence representation is calculated by the target behavior sequence item representation; based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning a pre-constructed isomorphic behavior transition graph and a heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is calculated by the auxiliary behavior sequence item representation; based on the target behavior sequence representation and the auxiliary behavior sequence representation, a session representation is obtained by fusion; based on the session representation, the predicted score of the item is calculated, and the item is recommended to the user according to the predicted score. It can be seen that the scheme of the present invention proposes a behavior-aware graph neural network based on a dual-channel learning strategy, wherein the dual channels include a target channel and an auxiliary channel. By pre-modeling heterogeneous behavior transitions in the form of graphs and applying them to auxiliary channels, the contribution of auxiliary behavior sequences to target behavior sequences is made clearer, thereby improving recommendation performance, that is, the recommendation accuracy.

[0009] In order to achieve the above technical effects, the first aspect of the present invention provides a recommendation processing method based on a behavior-aware graph neural network, wherein the method comprises:

[0010] Each heterogeneous behavior session is obtained, and each heterogeneous behavior session is divided into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing different behaviors of two or more users on an item.

[0011] Based on the target behavior sequence, obtaining the target behavior sequence item representation through learning of a pre-constructed isomorphic behavior transition graph, and obtaining the target behavior sequence representation through calculation of the target behavior sequence item representation;

[0012] Based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning through a pre-constructed isomorphic behavior transition graph and a heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is calculated through the auxiliary behavior sequence item representation;

[0013] Based on the target behavior sequence representation and the auxiliary behavior sequence representation, a conversation representation is obtained by fusing;

[0014] A prediction score of the item is calculated based on the session representation, and the item is recommended to the user according to the prediction score.

[0015] Optionally, before the step of obtaining each heterogeneous behavior session and dividing each heterogeneous behavior session into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing different behaviors of two or more users on an item, the step includes:

[0016] Pre-acquire all heterogeneous behavior sessions, and divide all heterogeneous behavior sessions into global target behavior sequences and global auxiliary behavior sequences;

[0017] By capturing the isomorphic behavior conversion information in the global target behavior sequence and the global auxiliary behavior sequence, an isomorphic behavior conversion graph is constructed;

[0018] By capturing the heterogeneous behavior conversion information between the global target behavior sequence and the global auxiliary behavior sequence, a heterogeneous behavior conversion graph is constructed.

[0019] Optionally, the step of constructing an isomorphic behavior conversion graph by capturing isomorphic behavior conversion information in the global target behavior sequence and the global auxiliary behavior sequence includes:

[0020] Based on the global auxiliary behavior sequence and the global target behavior sequence, taking items in the heterogeneous behavior session as nodes, and constructing isomorphic behavior conversion edges between two consecutive items with isomorphic behavior conversion relations;

[0021] The isomorphic behavior conversion relationship includes a conversion relationship from auxiliary behavior to auxiliary behavior, and from target behavior to target behavior.

[0022] Optionally, the step of constructing a heterogeneous behavior conversion graph by capturing heterogeneous behavior conversion information between the global target behavior sequence and the global auxiliary behavior sequence includes:

[0023] Based on the global auxiliary behavior sequence and the global target behavior sequence, taking items in the heterogeneous behavior session as nodes, and constructing a heterogeneous behavior conversion edge between two consecutive items having a directed heterogeneous behavior conversion relationship;

[0024] The directional heterogeneous behavior conversion relationship includes a conversion relationship from an auxiliary behavior to a target behavior.

[0025] Optionally, the above step of learning the target behavior sequence object representation through a pre-constructed isomorphic behavior transition graph based on the target behavior sequence, and calculating the target behavior sequence representation through the target behavior sequence object representation includes:

[0026] By transferring each item of the target behavior sequence in the isomorphic behavior transformation graph, a representation of the target behavior sequence item is learned;

[0027] Based on the target behavior sequence item representation, a target behavior sequence representation is calculated.

[0028] Optionally, the above step of obtaining the auxiliary behavior sequence item representation based on the auxiliary behavior sequence by learning the pre-constructed isomorphic behavior transition graph and the heterogeneous behavior transition graph, and obtaining the auxiliary behavior sequence representation by calculating the auxiliary behavior sequence item representation includes:

[0029] By transferring each item of the auxiliary behavior sequence in the isomorphic behavior transformation graph, a first view item representation of the item representation of the auxiliary behavior sequence is learned;

[0030] Based on the target behavior sequence representation, each item of the auxiliary behavior sequence is transferred in a personalized manner in the heterogeneous behavior transition graph, thereby learning a second view item representation of the auxiliary behavior sequence item representation;

[0031] The auxiliary behavior sequence item representation is obtained by weighing and fusing the first view item representation and the second view item representation;

[0032] Based on the auxiliary behavior sequence item representation, the auxiliary behavior sequence representation is calculated.

[0033] Optionally, the step of learning and obtaining a second view item representation about the auxiliary behavior sequence item representation by performing personalized transfer of each item in the auxiliary behavior sequence in the heterogeneous behavior conversion graph based on the target behavior sequence representation includes:

[0034] Based on the heterogeneous behavior transition graph, two groups of adjacent nodes with directional heterogeneous behavior transitions are defined for each item node, wherein the directional heterogeneous behavior transitions include behavior transitions from auxiliary behaviors to target behaviors;

[0035] Based on the target behavior sequence representation, adaptively adjust the importance of each adjacent node;

[0036] Based on the importance, each item of the auxiliary behavior sequence is transferred in an individualized manner in the heterogeneous behavior transition graph, and a second view item representation related to the auxiliary behavior sequence representation is learned.

[0037] Optionally, the step of calculating a predicted score of an item based on the session representation and recommending the item to the user according to the predicted score further includes:

[0038] A prediction score of the item is calculated based on the session representation, and the prediction score is used to determine whether the model has converged.

[0039] A second aspect of the present invention provides a server, comprising a memory, a processor, and a recommendation processing method program based on a behavior-aware graph neural network stored in the memory and executable on the processor, wherein the recommendation processing method program based on a behavior-aware graph neural network implements the steps of the recommendation processing method based on a behavior-aware graph neural network as described in any one of the above items when executed by the processor.

[0040] A third aspect of the present invention provides a non-temporary computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a recommendation processing method based on a behavior-aware graph neural network as described in any one of the above items.

[0041] As can be seen from the above, the scheme of the present invention obtains each heterogeneous behavior session, and divides each heterogeneous behavior session into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing two or more users' different behaviors on items. Based on the target behavior sequence, the target behavior sequence item representation is obtained by learning the pre-constructed isomorphic behavior transition graph, and the target behavior sequence representation is calculated by the target behavior sequence item representation; based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning the pre-constructed isomorphic behavior transition graph and the heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is calculated by the auxiliary behavior sequence item representation; based on the target behavior sequence representation and the auxiliary behavior sequence representation, the session representation is fused; the predicted score of the item is calculated based on the session representation, and the item is recommended to the user according to the predicted score. Because the scheme of the present invention adopts a dual-channel learning strategy to distinguish different types of behavior sequence modeling, and pays more attention to the modeling of auxiliary behavior sequences. The auxiliary behavior is modeled more finely by using the pre-established heterogeneous behavior transition graph, aiming to clearly use the semantic relationship information guided by the target behavior in the heterogeneous behavior transition graph to achieve a high-accuracy recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1 It is a flowchart of a recommendation processing method based on a behavior perception graph neural network provided by an embodiment of the present invention;

[0044] Figure 2 It is a schematic diagram of constructing a homogeneous behavior conversion graph and constructing a heterogeneous behavior conversion graph provided by an embodiment of the present invention;

[0045] Figure 3 It is a schematic diagram of a recommendation processing method BGNN based on a behavior-aware graph neural network provided in an embodiment of the present invention;

[0046] Figure 4 Embodiment of the present invention Figure 1 Specific flow diagram of step S200;

[0047] Figure 5 Embodiment of the present invention Figure 1 Specific flow diagram of step S300;

[0048] Figure 6 It is a schematic diagram of the comparison of the recommended performance of the BGNN algorithm provided by an embodiment of the present invention and other twelve baselines;

[0049] Figure 7 It is a visualization diagram of the representation of browsing and purchasing and the representation of the target item provided by the BGNN algorithm and the MGNN-SPred-W algorithm in the embodiment of the present invention;

[0050] Figure 8 The present invention provides a block diagram of the internal structure of a server. DETAILED DESCRIPTION

[0051] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0052] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0053] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0054] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0056] Intelligent recommendation is a technology that alleviates the problem of information overload in this era of information explosion. Among the different research branches of recommendation systems, session-based recommendation aims to recommend the next possible interactive item to users based on the user's behavior sequence in the session. In traditional session-based recommendation, a single type of user behavior sequence is often used for modeling, resulting in poor recommendation performance. People call this type of recommendation homogeneous recommendation. However, the behavior sequence of a session in a real-world e-commerce platform is usually heterogeneous, that is, it consists of more than one type of behavior, such as purchase and browsing. We call this type of research problem heterogeneous session-based recommendation (HSBR for short).

[0057] In the prior art, session-based heterogeneous recommendations can be divided into two categories, namely RNN-based methods and GNN-based methods. RNN-based methods such as RIB use behavior representations to distinguish different types of behaviors in heterogeneous sessions, and use the concatenation of item representations and behavior representations as the input layer of the gated recurrent unit. However, the recommendation effect of RNN-based methods when applied to this field is not as good as that of GNN-based methods. The GNN-based method MGNN-SPred, which is based on two types of isomorphic behavior sequences (i.e., browsing sequences and purchase sequences) in all sessions to construct a global multi-relational item graph, and uses the GNN method to enhance the learning of item representations. Although MGNN-SPred enhances the learning of item representations by utilizing isomorphic behavior conversion information, its modeling of different types of behavior sequences is the same, resulting in the same representation of browsing and purchasing items in the feature space, which may limit the expressive power of the model.

[0058] In order to solve the problems existing in the prior art, the present invention provides a recommendation processing method based on a behavior-aware graph neural network. In an embodiment of the present invention, each heterogeneous behavior session is obtained, and each heterogeneous behavior session is divided into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing two or more different behaviors of users on items. Based on the target behavior sequence, the target behavior sequence item representation is obtained by learning a pre-constructed isomorphic behavior transition graph, and the target behavior sequence representation is calculated by the target behavior sequence item representation; based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning a pre-constructed isomorphic behavior transition graph and a heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is calculated by the auxiliary behavior sequence item representation; based on the target behavior sequence representation and the auxiliary behavior sequence representation, the session representation is fused; the predicted score of the item is calculated based on the session representation, and the item is recommended to the user according to the predicted score. Therefore, compared with the solution in the prior art that only focuses on isomorphic behavior conversion, the present invention uses a pre-constructed heterogeneous behavior conversion graph to model the auxiliary behavior sequence in a more fine-grained manner, and pays more attention to the semantic relationship between the conversion of different types of behaviors, so as to improve the accuracy of the recommendation.

[0059] Exemplary Methods

[0060] like Figure 1 As shown, an embodiment of the present invention provides a recommendation processing method based on a behavior-aware graph neural network, and the method comprises the following steps:

[0061] Step S100: Obtain each heterogeneous behavior session, and divide each heterogeneous behavior session into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing different behaviors of two or more users on an item.

[0062] In this embodiment, each heterogeneous behavior session for training or prediction is first obtained, and each heterogeneous behavior session is divided into two parts: a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing different behaviors of two or more users. For example, when it is desired to predict the behavior of a user adding a song to a favorite list through a music software, the heterogeneous behavior session may contain data of the user adding a song to the favorite list, data of the user switching songs, and data of the user finishing listening to the song. In this step, the data of the user adding a song to the favorite list is considered to be the target behavior sequence, and the rest of the data are auxiliary behavior sequences.

[0063] Step S200: Based on the target behavior sequence, a target behavior sequence object representation is obtained by learning a pre-constructed isomorphic behavior transition graph, and a target behavior sequence representation is obtained by calculating the target behavior sequence object representation.

[0064] In this embodiment, according to the obtained target behavior sequence, each item in the target behavior sequence is input into a pre-constructed isomorphic behavior transformation graph, the target behavior sequence item representation is learned by transferring the isomorphic behavior transformation information in the isomorphic behavior transformation graph, and the target behavior sequence representation is calculated through the target behavior sequence item representation.

[0065] Step S300: Based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning the pre-constructed isomorphic behavior transition graph and the heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is calculated by the auxiliary behavior sequence item representation.

[0066] In this embodiment, according to the obtained auxiliary behavior sequence, each item in the auxiliary behavior sequence is input into the isomorphic behavior conversion diagram, and the first view item representation about the auxiliary behavior sequence item representation is learned by transferring the isomorphic behavior conversion information in the isomorphic behavior conversion diagram. Then, by inputting each item in the auxiliary behavior sequence into the heterogeneous behavior conversion diagram for transfer, the second view item representation for improving the modeling accuracy of the auxiliary behavior sequence item representation is learned, and the first view item representation and the second view item representation are fused to obtain a higher accuracy auxiliary behavior sequence item representation, and further the corresponding auxiliary behavior sequence representation is obtained by calculation. Through this step, the accuracy of auxiliary behavior sequence modeling is improved through the heterogeneous behavior conversion diagram.

[0067] Step S400: Based on the target behavior sequence representation and the auxiliary behavior sequence representation, a session representation is obtained by fusing them.

[0068] In this embodiment, the final session representation is obtained by fusion calculation through a gating module based on the acquired target behavior sequence representation and the auxiliary behavior sequence representation.

[0069] Step S500: Calculate a predicted score of an item based on the session representation, and recommend the item to the user according to the predicted score.

[0070] In this embodiment, based on the session representation, the prediction score of each item is further calculated. In the prediction stage of the model, the prediction score can be used to push products that meet the user's purchase intention to the user. In the training stage of the model, the prediction score can also be used to judge whether the model has converged.

[0071] As can be seen from the above, the recommendation processing method based on the behavior perception graph neural network provided by the embodiment of the present invention obtains each heterogeneous behavior session, and divides each heterogeneous behavior session into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing two or more users' different behaviors on items. Based on the target behavior sequence, the target behavior sequence item representation is obtained by learning the pre-constructed isomorphic behavior transition graph, and the target behavior sequence representation is calculated by the target behavior sequence item representation; based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning the pre-constructed isomorphic behavior transition graph and the heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is calculated by the auxiliary behavior sequence item representation; based on the target behavior sequence representation and the auxiliary behavior sequence representation, the session representation is fused; the predicted score of the item is calculated based on the session representation, and the item is recommended to the user according to the predicted score. Compared with the prior art that focuses on how to distinguish different types of behaviors in sequence modeling or use isomorphic behavior transitions within each behavior sequence for recommendation, the solution of the present invention can achieve more accurate intelligent recommendation by capturing the semantic relationship between the transitions of different types.

[0072] In a further embodiment, taking the application scenario of an e-commerce platform as an example, a behavior-aware graph neural network (BGNN) is proposed. In this embodiment, the target behavior is the user's purchase behavior, and the auxiliary behavior is the user's browsing behavior.

[0073] In an application scenario, each heterogeneous behavior session is obtained, and each heterogeneous behavior session is divided into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing different behaviors of two or more users on an item.

[0074] The step of obtaining each heterogeneous behavior session and dividing each heterogeneous behavior session into a target behavior sequence and an auxiliary behavior sequence includes:

[0075] Pre-acquire all heterogeneous behavior sessions, and divide all heterogeneous behavior sessions into global target behavior sequences and global auxiliary behavior sequences;

[0076] By capturing the isomorphic behavior conversion information in the global target behavior sequence and the global auxiliary behavior sequence, an isomorphic behavior conversion graph is constructed;

[0077] By capturing the heterogeneous behavior conversion information between the global target behavior sequence and the global auxiliary behavior sequence, a heterogeneous behavior conversion graph is constructed.

[0078] Specifically, the step of constructing an isomorphic behavior conversion graph by capturing isomorphic behavior conversion information in the global target behavior sequence and the global auxiliary behavior sequence includes:

[0079] Based on the global auxiliary behavior sequence and the global target behavior sequence, taking items in the heterogeneous behavior session as nodes, and constructing isomorphic behavior conversion edges between two consecutive items with isomorphic behavior conversion relations;

[0080] The isomorphic behavior conversion relationship includes a conversion relationship from auxiliary behavior to auxiliary behavior, and from target behavior to target behavior.

[0081] The step of constructing a heterogeneous behavior conversion graph by capturing heterogeneous behavior conversion information between the global target behavior sequence and the global auxiliary behavior sequence comprises:

[0082] Based on the global auxiliary behavior sequence and the global target behavior sequence, taking items in the heterogeneous behavior session as nodes, and constructing a heterogeneous behavior conversion edge between two consecutive items having a directed heterogeneous behavior conversion relationship;

[0083] The directional heterogeneous behavior conversion relationship includes a conversion relationship from an auxiliary behavior to a target behavior.

[0084] For example, all heterogeneous behavior sessions are pre-acquired, and the control divides all the acquired heterogeneous behavior sessions into two parts, including a browsing sequence and a purchase sequence Based on the above browsing sequence E s and purchase sequence P sTwo different graphs are constructed, including a homogeneous behavior transition graph (HoBTG) and a heterogeneous behavior transition graph (HeBTG). The HoBTG captures the homogeneous behavior transition information of "from browsing to browsing" in the browsing sequence and "from buying to buying" in the buying sequence. The HeBTG captures the heterogeneous behavior transition information that occurs between the browsing sequence and the buying sequence, specifically the behavior transition information of "from browsing to buying" in the same session in this embodiment.

[0085] Further, Figure 2 Let HoBTG be ,in, is the node set, and ε is the edge set. Specifically, we take all items in all sessions as the node set of HoBTG , and construct edges with isomorphic behavior transitions between two consecutive items, i.e., from browse to browse (e2e) or from purchase to purchase (p2p). For example, in the purchase sequence of session 1, we can see that there is a p2p transition from v2 to v4, so we add a p2p edge from node v2 to node v4 in the HoBTG. Finally, the HoBTG consists of two types of edges, i.e., (v, e2e, v′) and (v, p2p, v′). Specifically, the edge (v, e2e, v′) indicates that the user browsed item v′ after browsing item v, and (v, p2p, v′) indicates that the user purchased item v′ after purchasing item v.

[0086] Please continue reading Figure 2 Although HoBTG collects rich homogeneous behavior conversion information, it ignores the heterogeneous behavior conversion information between browsing sequences and purchase sequences of the same session, which is crucial for discovering complex semantic relationships between different types of behaviors. To solve this problem, HeBTG is pre-constructed. Specifically, an edge is constructed between two consecutive items that have a directional behavior transition relationship from "browsing to purchasing". For example, in session 2, we can see that there is an e2p behavior transition, so we connect node v1 to node v4 to construct an edge (v, e2p, v'). Then an edge in the HeBTG can be represented as (v, e2p, v'), indicating that the user purchased item v' after browsing item v.

[0087] It should be noted that in this embodiment, we focus on the directional relationship from browsing to purchasing rather than the opposite direction because we want to model the conversion from auxiliary behavior to target behavior, which is more in line with our goal of predicting purchase behavior, that is, recommending some items that users may buy, and can make the contribution of auxiliary behavior sequences to target behavior sequences clearer. HeBTG can be regarded as a behavior perception graph, which builds a bridge for direct conversion from auxiliary behavior to target behavior.

[0088] Furthermore, for each edge in HoBTG and HeBTG, the frequency of its occurrence in all sessions is used as its edge weight. In addition, for each edge type of each node in HoBTG or HeBTG, only a certain number of edges with the highest edge weight are retained, which can greatly reduce the computational overhead and avoid the noise problem caused by insufficient information of adjacent nodes, thereby further improving the recommendation accuracy.

[0089] Based on the above-mentioned pre-built HoBTG and HeBTG, in order to simultaneously utilize isomorphic and heterogeneous behavior conversion information in session-based heterogeneous recommendation scenarios, a recommendation processing method called behavior-aware graph neural network-based BGNN is proposed in this embodiment.

[0090] like Figure 3 As shown, the BGNN first obtains each heterogeneous behavior session s and divides each heterogeneous behavior session s into a browsing sequence E s and purchase sequence P s . Then, the browsing sequence E obtained from a heterogeneous behavior session s is s and purchase sequence P s Input into the auxiliary channel and the target channel to learn the item representation of each sequence.

[0091] In an application scenario, the BGNN learns the target behavior sequence item representation based on the target behavior sequence through a pre-constructed isomorphic behavior transition graph, and calculates the target behavior sequence representation through the target behavior sequence item representation.

[0092] Specifically, Figure 4 As shown, the above step S200 includes:

[0093] Step S201, learning and obtaining the target behavior sequence item representation by transferring each item in the target behavior sequence in the isomorphic behavior transformation graph;

[0094] Step S202: Calculate and obtain a target behavior sequence representation based on the target behavior sequence item representation.

[0095] For an example, please see Figure 3 . In the target channel of BGNN, our goal is to transfer the information in HoBTG to learn high-quality item representations of the purchase sequence, so in this embodiment, the BGNN uses the GNN-based graphSAGE method and mean aggregator to learn item representations because it is simple and effective in the setting of this embodiment. Among them, the graphSAGE method and mean aggregator can be specifically referenced in the literature: William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive Representation Learning on Large Graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS'17). 1025-1035.

[0096] Each node v has four groups of isomorphic adjacent nodes, namely, out-degree / in-degree adjacent nodes of e2e / p2p isomorphic behavior conversion, including: and

[0097] First, the four groups of neighboring nodes of node v are aggregated separately, and the representations of its four k-order neighbor groups are obtained by mean pooling as follows:

[0098]

[0099]

[0100]

[0101]

[0102] in, represents the (k-1)th order representation of node v′, represents a set of real number matrices with d rows and 1 column. The k-order representation of the target node v It will be updated recursively by the following formula:

[0103]

[0104] Among them, k is the depth of GNN. That is, the initial representation (layer 0) of node v is the item representation m v .

[0105] From the above, it can be seen that for the purchase sequence P = {p1, p2, p3, ..., p |P|} (superscript s is omitted here for brevity), we can get its item representation in

[0106] Further, based on the purchase sequence item representation The purchase sequence representation of session s is calculated by the following formula:

[0107]

[0108]

[0109] in, Yes H p The mean vector of . is a trainable projection matrix and σ(·) is the sigmoid function.

[0110] In an application scenario, the BGNN obtains the auxiliary behavior sequence item representation based on the auxiliary behavior sequence through pre-constructed isomorphic behavior transition graphs and heterogeneous behavior transition graphs, and calculates the auxiliary behavior sequence representation through the auxiliary behavior sequence item representation.

[0111] Specifically, Figure 5 As shown, the above step S300 includes:

[0112] Step S301, learning and obtaining a first view item representation of the item representation of the auxiliary behavior sequence by transferring each item of the auxiliary behavior sequence in the isomorphic behavior transformation graph;

[0113] Step S302: Based on the target behavior sequence representation, each item of the auxiliary behavior sequence is transferred in a personalized manner in the heterogeneous behavior conversion graph to learn and obtain a second view item representation of the auxiliary behavior sequence item representation;

[0114] Step S303: Obtain auxiliary behavior sequence item representation by weighing and fusing the first view item representation and the second view item representation;

[0115] Step S304: Based on the auxiliary behavior sequence object representation, an auxiliary behavior sequence representation is calculated.

[0116] The step of learning and obtaining a second view item representation of the auxiliary behavior sequence item representation based on the target behavior sequence representation by personalized transmission of each item of the auxiliary behavior sequence in the heterogeneous behavior conversion graph comprises:

[0117] Based on the heterogeneous behavior transition graph, two groups of adjacent nodes with directional heterogeneous behavior transitions are defined for each item node, wherein the directional heterogeneous behavior transitions include behavior transitions from auxiliary behaviors to target behaviors;

[0118] Based on the target behavior sequence representation, adaptively adjust the importance of each adjacent node;

[0119] Based on the importance, each item of the auxiliary behavior sequence is transferred in an individualized manner in the heterogeneous behavior transition graph, and a second view item representation related to the auxiliary behavior sequence representation is learned.

[0120] For an example, please see Figure 3 In the auxiliary channel of BGNN, this embodiment is provided with three parts, including representation learning of browsing sequence in HoBTG, representation learning in HeBTG and representation gating module, which are used to learn and calculate the final item representation of browsing sequence.

[0121] First, as done in the target channel, for the browsing sequence E = {e1, e2, e3, ..., e |E| By leveraging the isomorphic behavior transformations in HoBTG that are common to the target channel, a generic view item representation of the browsing sequence can be obtained. That is, the first view item representation of the auxiliary behavior sequence item representation mentioned above, where

[0122] Second, although the general view item representation H of the browsing sequence e It already contains rich isomorphic behavior transition information from HoBTG, but we think it is not sufficient. On the one hand, the representation learned in HoBTG is behavior agnostic. For example, given two identical item sequences, one is a click sequence and the other is a purchase sequence, HoBTG will pass exactly the same information to the item node and then return exactly the same item representation. This is unreasonable, which also shows that HoBTG cannot distinguish different types of behaviors. On the other hand, the representation learned in HoBTG ignores the semantic connection with the target behavior, so it may not capture the user's intention well. Therefore, we propose to learn the purchase-oriented view item representation of the browsing sequence in HeBTG, that is, the second view item representation mentioned above about the auxiliary behavior sequence item representation.

[0123] First, according to HeBTG, we define two sets of behavior-aware neighboring nodes for each item node v, namely, out-degree / in-degree nodes transformed by e2p heterogeneous behaviors and in is the set of adjacent nodes when node v is in browsing behavior, It is the set of adjacent nodes when node v is a purchase behavior. It can be seen that in the HeBTG of this embodiment, the nodes of the browsing behavior are only connected to the nodes of the purchase behavior, and vice versa.

[0124] Next, in order to personalize the information transfer in HeBTG, a personalized item neighbor aggregator (PING) is proposed in BGNN to capture the user purchase preferences in the purchase sequence, that is, the purchase sequence representation Adaptively adjust the importance of different neighbor nodes. For example, given two different sessions with the same browsing sequence and completely different purchase sequences, PING in BGNN will adjust the contribution of different neighbor nodes to generate different item representations for the two same browsing sequences, which helps learn the purchase-oriented view item representation of the browsing sequence. Specifically, PING recursively aggregates and updates the representation of the browsing behavior node v through the following formula:

[0125]

[0126]

[0127] in, k-th order representation of node v representing browsing behavior, The (k-1)-order representation of node v′ representing the purchase behavior, is an attention value, indicating the importance of the neighboring node v′ in the (k-1) layer to the target node v. represents the purchase sequence representation of session s, is the projection matrix, and σ(·) represents the sigmoid function.

[0128] Make the attention value The calculation of p represents the purchase sequence of the current session s s Taking this into account, the aggregation of neighboring nodes in HeBTG can be personalized. Similarly, the representation of the purchase behavior node v is aggregated and updated as follows:

[0129]

[0130]

[0131] in, is the attention value, and It should be noted that when messages are transmitted in the HeBTG composed of e2p edges, nodes with purchasing behavior will collect information from nodes with browsing behavior, and vice versa. After extending the PING aggregation operation in HeBTG to k layers, for a browsing sequence E = {e1, e2, e3, ..., e |E|}, we can get the purchase-oriented view item representation of the item That is, the second view item representation, wherein

[0132] Furthermore, based on the above-obtained general view item representation of the browsing sequence and purchase-oriented view item representation The BGNN weighs and fuses these two item representations about the browsing sequence through a representation gating module (RNG), and the RNG specifically includes:

[0133]

[0134]

[0135] Among them, [·, ·] represents the concatenation operation, σ(·) is the sigmoid function, and γ e represents the trade-off between two representation matrices of a browsing sequence, is a trainable parameter, and Represents the fused representation matrix of a browsing sequence. It can be seen that the RNG does not directly operate the two representation matrices like average pooling or maximum pooling, but uses the mean vector of the two matrices to calculate the weights of the two representation matrices, and takes the corresponding weighted sum of the two representation matrices as the final representation of the item.

[0136] Further, based on the item representation Z of the browsing sequence e , the browsing sequence representation of heterogeneous session s can be calculated by the same method as formulas (6) and (7):

[0137] In an application scenario, the BGNN obtains a session representation based on the target behavior sequence representation and the auxiliary behavior sequence representation.

[0138] Specifically, the above step S400 includes:

[0139] For example, based on the learned purchase sequence representation and browse sequence representations The BGNN obtains the final representation of session s through the gating function:

[0140] r s =g·e s +(1-g)·p s (14)

[0141]

[0142] in, is a trainable parameter.

[0143] In an application scenario, the BGNN calculates a prediction score of an item based on the session representation, and recommends the item to the user based on the prediction score.

[0144] Specifically, the above step S500 includes:

[0145] For example, based on the session representation r s , the BGNN calculates the prediction score for each item:

[0146]

[0147] in, is a trainable parameter, m i is the representation of item i. When the BGNN model is applied to the prediction stage, the prediction score The items that the user wants to buy can be recommended to the user more accurately.

[0148] Wherein, the step of calculating the predicted score of the item based on the session representation and recommending the item to the user according to the predicted score also includes:

[0149] A prediction score of the item is calculated based on the session representation, and the prediction score is used to determine whether the model has converged.

[0150] For example, during the training phase, you can also use cross entropy as the loss function of the model to train the entire model:

[0151]

[0152] where y i ∈{0, 1}, represents the true label of item i. i =1, it means that item i is the item that the user actually interacts with in the next step, otherwise y i =0.

[0153] Based on the recommendation processing method based on the behavior-aware graph neural network proposed in the above embodiment, the recommendation effect is further verified, specifically including:

[0154] In order to evaluate the recommendation effect of the BGNN model, BGNN is compared with the following baseline algorithms, including some traditional, RNN-based, attention-based, Transformer-based and GNN-based methods, specifically two traditional methods POP and ItemKNN, five isomorphic session-based methods GRU4Rec, NARM, STAMP, SR-GNN and GCE-GNN, and five heterogeneous session-based methods RIB, M-SR, DMT-trans, MGNN-SPred and MGNN SPred-W.

[0155] like Figure 6 As shown, the recommendation performance of BGNN and the above twelve baselines on three datasets, Tmall, Yoochoose and UB, is shown, where the best result is marked in bold and the second best result is marked with an underline. It can be seen that the experimental results show that the BGNN model proposed in the embodiment of the present invention is superior to other related recommendation algorithms, and is 21.87%, 18.49 and 37.16% higher than the best baseline on the three datasets, respectively.

[0156] In addition, if Figure 7 As shown, we also visualize the representations of browsing and purchase to show whether our BGNN has learned semantic information of different types of behaviors. Among them, the five-pointed star represents the representation of the target item, the circle represents the representation of browsing, and the triangle represents the representation of purchase. Specifically, we randomly select a target item (the purchase item to be predicted) in each dataset and retrieve the sessions using this item as the label, and finally visualize the representations of browsing and purchase in these sessions as well as the representation of the target item. Note that the dimension of the representation is 64, so we use the T-SNE tool for visualization. To compare with BGNN, we also show the visualization of MGNN-SPred-W, which is the best baseline model in our previous experiment. In the visualization of MGNN-SPred-W, we do not observe obvious patterns, and we can see that there are some overlaps between the representations of purchase and browsing. And we can find that for our BGNN: 1) the representation of browsing gradually converges to the target item, which is reasonable because the browsing of users before purchasing items is more relevant; 2) compared with the relatively tight distribution of the learned browsing representation, the purchase representation presents a more scattered state, which reflects the semantic difference between browsing and purchasing. It can be seen that the BGNN has learned the semantic information of different types of behaviors.

[0157] Exemplary Devices

[0158] Based on the above embodiments, the present invention provides a server, whose principle block diagram can be shown as follows: Figure 8 As shown. The above-mentioned server includes a processor, a memory, a network interface and a display screen connected through a system bus. Among them, the processor of the server is used to provide computing and control capabilities. The memory of the server includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the server is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the steps of any one of the above-mentioned recommendation processing methods based on the behavior-aware graph neural network are implemented. The display screen of the server can be a liquid crystal display or an electronic ink display.

[0159] Those skilled in the art will understand that Figure 8 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the solution of the present invention, and does not constitute a limitation on the server to which the solution of the present invention is applied. The specific server may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0160] In one embodiment, a server is provided, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein the program performs the following operation instructions when executed by the processor:

[0161] Each heterogeneous behavior session is obtained, and each heterogeneous behavior session is divided into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing different behaviors of two or more users on an item.

[0162] Based on the target behavior sequence, obtaining the target behavior sequence item representation through learning of a pre-constructed isomorphic behavior transition graph, and obtaining the target behavior sequence representation through calculation of the target behavior sequence item representation;

[0163] Based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning through a pre-constructed isomorphic behavior transition graph and a heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is calculated through the auxiliary behavior sequence item representation;

[0164] Based on the target behavior sequence representation and the auxiliary behavior sequence representation, a conversation representation is obtained by fusing;

[0165] The predicted score of the item is calculated based on the session representation, and the item is recommended to the user according to the predicted score. It should be understood that the sequence number of each step in the above embodiment does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0166] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the above-mentioned device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0167] In the above embodiments, the description of each embodiment has its own emphasis. For the part that is not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments. For quantities with the same expression form in various formulas, unless otherwise specified, they represent the same quantity, and various formulas can refer to each other.

[0168] Those skilled in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0169] In the embodiments provided by the present invention, it should be understood that the disclosed apparatus / terminal equipment and method can be implemented in other ways. For example, the apparatus / terminal equipment embodiments described above are only illustrative, for example, the division of the above modules or units is only a logical function division, and in actual implementation, other division methods can be used, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0170] If the above-mentioned integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The above-mentioned computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the above-mentioned computer program includes computer program code, and the above-mentioned computer program code can be in source code form, object code form, executable file or some intermediate form. The above-mentioned computer-readable medium may include: any entity or device capable of carrying the above-mentioned computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the above-mentioned computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0171] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features therein may be replaced by equivalents. However, these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A recommendation processing method based on behavior-aware graph neural network, characterized in that: The steps include: Pre-acquire all heterogeneous behavior sessions, and divide all heterogeneous behavior sessions into global target behavior sequences and global auxiliary behavior sequences; By capturing the isomorphic behavior conversion information in the global target behavior sequence and the global auxiliary behavior sequence, an isomorphic behavior conversion graph is constructed; By capturing heterogeneous behavior conversion information between the global target behavior sequence and the global auxiliary behavior sequence, a heterogeneous behavior conversion graph is constructed; Obtain each heterogeneous behavior session, and divide each heterogeneous behavior session into a target behavior sequence and an auxiliary behavior sequence, wherein the heterogeneous behavior session is data containing different behaviors of two or more users on an item; Based on the target behavior sequence, obtaining the target behavior sequence item representation through learning of a pre-constructed isomorphic behavior transition graph, and obtaining the target behavior sequence representation through calculation of the target behavior sequence item representation; Based on the auxiliary behavior sequence, the auxiliary behavior sequence item representation is obtained by learning through a pre-constructed isomorphic behavior transition graph and a heterogeneous behavior transition graph, and the auxiliary behavior sequence representation is calculated through the auxiliary behavior sequence item representation; The steps of obtaining auxiliary behavior sequence item representation based on the auxiliary behavior sequence by learning through a pre-constructed isomorphic behavior transition graph and a heterogeneous behavior transition graph, and obtaining the auxiliary behavior sequence representation by calculating through the auxiliary behavior sequence item representation include: By transferring each item of the auxiliary behavior sequence in the isomorphic behavior transformation graph, a first view item representation of the item representation of the auxiliary behavior sequence is learned; Based on the target behavior sequence representation, each item of the auxiliary behavior sequence is transferred in a personalized manner in the heterogeneous behavior transition graph, thereby learning a second view item representation of the auxiliary behavior sequence item representation; The auxiliary behavior sequence item representation is obtained by weighing and fusing the first view item representation and the second view item representation; Based on the auxiliary behavior sequence object representation, calculating the auxiliary behavior sequence representation; Based on the target behavior sequence representation and the auxiliary behavior sequence representation, a conversation representation is obtained by fusing; A prediction score of the item is calculated based on the session representation, and the item is recommended to the user according to the prediction score.

2. The recommendation processing method based on behavior-aware graph neural network according to claim 1, characterized in that: The step of constructing an isomorphic behavior conversion graph by capturing isomorphic behavior conversion information in the global target behavior sequence and the global auxiliary behavior sequence comprises: Based on the global auxiliary behavior sequence and the global target behavior sequence, taking items in the heterogeneous behavior session as nodes, and constructing isomorphic behavior conversion edges between two consecutive items with isomorphic behavior conversion relations; The isomorphic behavior conversion relationship includes a conversion relationship from auxiliary behavior to auxiliary behavior, and from target behavior to target behavior.

3. The recommendation processing method based on behavior-aware graph neural network according to claim 1, characterized in that: The step of constructing a heterogeneous behavior conversion graph by capturing heterogeneous behavior conversion information between the global target behavior sequence and the global auxiliary behavior sequence comprises: Based on the global auxiliary behavior sequence and the global target behavior sequence, taking items in the heterogeneous behavior session as nodes, and constructing a heterogeneous behavior conversion edge between two consecutive items having a directed heterogeneous behavior conversion relationship; The directional heterogeneous behavior conversion relationship includes a conversion relationship from an auxiliary behavior to a target behavior.

4. The recommendation processing method based on behavior-aware graph neural network according to claim 1, characterized in that: The steps of obtaining the target behavior sequence object representation through learning a pre-constructed isomorphic behavior transition graph based on the target behavior sequence, and obtaining the target behavior sequence representation through calculation of the target behavior sequence object representation include: By transferring each item of the target behavior sequence in the isomorphic behavior transformation graph, a representation of the target behavior sequence item is learned; Based on the target behavior sequence item representation, a target behavior sequence representation is calculated.

5. The recommendation processing method based on behavior-aware graph neural network according to claim 1, characterized in that: The step of learning and obtaining a second view item representation of the auxiliary behavior sequence item representation by performing personalized transfer of each item of the auxiliary behavior sequence in the heterogeneous behavior conversion graph based on the target behavior sequence representation comprises: Based on the heterogeneous behavior transition graph, two groups of adjacent nodes with directional heterogeneous behavior transitions are defined for each item node, wherein the directional heterogeneous behavior transitions include behavior transitions from auxiliary behaviors to target behaviors; Based on the target behavior sequence representation, adaptively adjust the importance of each adjacent node; Based on the importance, each item of the auxiliary behavior sequence is transferred in an individualized manner in the heterogeneous behavior transition graph, and a second view item representation related to the auxiliary behavior sequence representation is learned.

6. The recommendation processing method based on behavior-aware graph neural network according to claim 1, characterized in that: The step of calculating a predicted score of an item based on the session representation and recommending an item to a user according to the predicted score also includes: A prediction score of the item is calculated based on the session representation, and the prediction score is used to determine whether the model has converged.

7. A server, characterized in that: The server includes a memory, a processor, and a recommendation processing method program based on a behavior-aware graph neural network stored in the memory and executable on the processor. When the recommendation processing method program based on a behavior-aware graph neural network is executed by the processor, the steps of the recommendation processing method based on a behavior-aware graph neural network as described in any one of claims 1 to 6 are implemented.

8. A non-temporary computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the recommendation processing method based on a behavior-aware graph neural network as described in any one of claims 1-6.

Images