Graph conversion network conversation recommendation method and system based on multi-behavior and item knowledge

Abstract

The present disclosure belongs to the field of artificial intelligence and information retrieval technology, and specifically relates to a graph conversion network conversation recommendation method and system based on multi-behavior and project knowledge, including: obtaining a user-project multi-behavior interaction graph and a knowledge-aware project-project relationship graph; fusing the obtained user-project multi-behavior interaction graph and the knowledge-aware project-project relationship graph based on multi-behavior interaction pattern embedding to obtain a conversation loss function; learning project knowledge through a knowledge graph embedding translation model based on paired relationship vectors, and assisting in learning user-project multi-behavior interaction relationships and project-project dependency relationships to obtain a project knowledge loss function; obtaining a comprehensive loss function based on multi-link aggregation and fusion of recommendation loss functions and project knowledge loss functions to achieve conversation recommendation.

Description

Technical Field

[0001] The present disclosure belongs to the field of artificial intelligence and information retrieval technology, and specifically relates to a graph conversion network conversation recommendation method and system based on multi-behavior and project knowledge. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Recommendation systems have played a significant role in the Internet field, such as e-commerce, online websites, website search, etc. Conversational recommendation is to predict the items that the target user will interact with next based on the user's interaction behavior in the recent period of time. In many websites, the characteristics of conversational data reflect the user's behavior patterns and current preferences; the current user intention is modeled based on the user's behavior in the most recent session to push better recommendation results to the target user.

[0004] Based on the existence of various types of behavioral relationships between users and items (such as browsing, collecting, adding to shopping carts, and purchasing), it is helpful to learn users' preferences for target behavior types. Learning the interdependencies between items is also beneficial to improving the performance of recommendations. For example, shopping websites can learn users' preferences based on their behaviors and recommend similar products to users. In addition, the dependencies between items are also a part of recommendations. For example, if two items have a lot in common (i.e., dependencies), they can be recommended to users. If there are no similarities, there is no need to recommend to users.

[0005] In the existing technology, multi-behavior interaction patterns can be applied to conversational recommendations. However, multi-type interactions are considered in a relatively independent and local manner, and user-item multi-type interactions and item-item dependencies are considered independently, which makes it difficult to capture high-order multi-behavior relationships between users and items. Knowledge-based recommendations have become an important part of recommendation systems, using time-delay neural networks (DNNs) to learn knowledge embedding and input it into downstream recommendation models.

[0006] According to the inventors, the graph conversion network session recommendation has the following disadvantages:

[0007] (1) Independently learn various types of user-item interaction behavior relationships and item-item interdependencies;

[0008] (2) The dependencies between items are not used as auxiliary signals to guide the learning of user-item multi-behavior interactions. Summary of the invention

[0009] In order to solve the above problems, the present disclosure proposes a graph conversion network conversation recommendation method and system based on multi-behavior and project knowledge, which uses a graph converter network to perform recursive embedding propagation to capture high-order behavioral heterogeneity between users and projects in an attention aggregation mode; the learned user-project multi-behavior embedding and project-project interdependency embedding are fused together through a multi-layer perceptron MLP and applied in the conversation; project embedding is learned by combining project knowledge with a knowledge graph embedding translation model PairRE based on pairwise relationship vectors; knowledge embedding is used as an auxiliary task and added to the multi-link aggregation MLT paradigm to predict the next interactive item and complete the recommendation.

[0010] According to some embodiments, a first solution of the present disclosure provides a graph conversion network session recommendation method based on multi-behavior and project knowledge, which adopts the following technical solution:

[0011] A graph transformation network conversation recommendation method based on multi-behavior and item knowledge, comprising:

[0012] Obtain user-item multi-behavior interaction graphs and knowledge-aware item-item relationship graphs;

[0013] Based on the multi-behavior interaction pattern embedding, the acquired user-item multi-behavior interaction graph and the knowledge-aware item-item relationship graph are fused to obtain the session loss function;

[0014] The project knowledge is learned through the knowledge graph embedding translation model based on the pairwise relationship vector, and the user-project multi-behavior interaction relationship and project-project dependency relationship are assisted in learning to obtain the project knowledge loss function;

[0015] Based on the multi-link aggregation fusion of the recommendation loss function and the item knowledge loss function, a comprehensive loss function is obtained to achieve conversational recommendation.

[0016] As a further technical limitation, in the process of obtaining the user-item multi-behavior interaction graph, the different users of the session recommendation and the items corresponding to each user are determined, and according to the order in which the users click on the items, the user-item correspondence relationship and the heterogeneous relationship between the users and the session recommendation items are formed to construct the user-item multi-behavior interaction graph.

[0017] Furthermore, a project-project relationship graph perceived based on the acquired knowledge represents multiple dependencies between projects.

[0018] As a further technical limitation, based on time context coding, behavioral dynamic information for characterizing the interaction between different users and items in a time-aware scenario is obtained, and the obtained behavioral dynamic information is incorporated into the heterogeneous signal.

[0019] As a further technical limitation, a graph attention mechanism is used to propagate time-aware information between source nodes and target nodes in the obtained user-item multi-behavior interaction graph; in the process of time-aware information propagation, the global context information of different behavior types is merged, and multi-channel parameter learning is used to process the attention transfer matrix.

[0020] As a further technical limitation, in the process of embedding project knowledge, the knowledge graph embedding translation model based on pairwise relationship vectors is imported into the hyperplane to handle many-to-one and many-to-many project-project relationships.

[0021] As a further technical limitation, item knowledge embedding learning is imported into the multi-link aggregation paradigm as an auxiliary task, and the learning results of the task are used as prompts to guide the learning of another task, assisting in completing the conversational recommendation task.

[0022] According to some embodiments, the second solution of the present disclosure provides a graph conversion network conversation recommendation system based on multi-behavior and project knowledge, which adopts the following technical solution:

[0023] A graph transformation network conversational recommendation system based on multi-behavior and item knowledge, including:

[0024] an acquisition module configured to acquire a user-item multi-behavior interaction graph and a knowledge-aware item-item relationship graph;

[0025] A fusion module configured to fuse the acquired user-item multi-behavior interaction graph and the knowledge-aware item-item relationship graph based on multi-behavior interaction pattern embedding to obtain a session loss function;

[0026] An embedding module is configured to learn project knowledge through a knowledge graph embedding translation model based on pairwise relationship vectors, and to assist in learning user-project multi-behavior interaction relationships and project-project dependency relationships, thereby obtaining a project knowledge loss function;

[0027] The recommendation module is configured to obtain a comprehensive loss function based on multi-link aggregation and fusion of the recommendation loss function and the item knowledge loss function to achieve session recommendation.

[0028] According to some embodiments, a third solution of the present disclosure provides a computer-readable storage medium, which adopts the following technical solution:

[0029] A computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the graph conversion network session recommendation method based on multi-behavior and item knowledge as described in the first aspect of the present disclosure.

[0030] According to some embodiments, a fourth solution of the present disclosure provides an electronic device, which adopts the following technical solution:

[0031] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for recommending a graph conversion network session based on multi-behavior and item knowledge as described in the first aspect of the present disclosure are implemented.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention integrates multiple types of behavioral interactions and dependencies between items and applies them to conversational recommendations; takes item knowledge as an auxiliary task of the conversation in which user-item multi-behavioral interactions and item-item dependencies are integrated, and studies the hit rate HR and normalized discounted cumulative return NDCG of recommendations; integrates independent user-item multi-behavioral interactions and item-item dependencies through MLP and applies them to conversations; incorporates item knowledge into the model through the MLT paradigm, and takes item knowledge as an auxiliary task of the conversation in which user-item multi-behavioral interactions and item-item dependencies are integrated; captures different behaviors of users through multiple types of interactions, and provides users with high-precision recommendations; the dependencies between items in the recommended conversation are used to enhance the effectiveness of learning multi-behavioral recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0035] Figure 1 is a flowchart of a method for recommending a graph-converted network session based on multi-behavior and project knowledge in Embodiment 1 of the present disclosure;

[0036] Figure 2 is a schematic diagram of the structure of the recommendation model in the first embodiment of the present disclosure;

[0037] Figure 3 It is a structural block diagram of a graph conversion network conversation recommendation system based on multi-behavior and item knowledge in the second embodiment of the present disclosure. DETAILED DESCRIPTION

[0038] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0040] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0041] In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other.

[0042] Embodiment 1

[0043] Embodiment 1 of the present disclosure introduces a graph conversion network session recommendation method based on multi-behavior and project knowledge.

[0044] like Figure 1 A graph conversion network session recommendation method based on multi-behavior and item knowledge is shown, comprising:

[0045] Obtain user-item multi-behavior interaction graphs and knowledge-aware item-item relationship graphs;

[0046] Based on the multi-behavior interaction pattern embedding, the acquired user-item multi-behavior interaction graph and the knowledge-aware item-item relationship graph are fused to obtain the session loss function;

[0047] The project knowledge is learned through the knowledge graph embedding translation model based on the pairwise relationship vector, and the user-project multi-behavior interaction relationship and project-project dependency relationship are assisted in learning to obtain the project knowledge loss function;

[0048] Based on the multi-link aggregation fusion of the recommendation loss function and the item knowledge loss function, a comprehensive loss function is obtained to achieve conversational recommendation.

[0049] The goal of this example is to capture the multi-behavior user-item interaction patterns and item-item dependencies in a graph transformer network. The two are integrated through MLP and applied to conversational recommendation. The embedding representation learned from the item knowledge is used as an auxiliary task to guide the main task of conversational recommendation SR through the MLT paradigm. The overall structural diagram is shown in the figure below. Figure 2 shown.

[0050] In this embodiment, I is used to represent the user U of the session recommendation = {u1, u2, ..., u i ,....u I}, J represents the items recommended by the session V = {v1,v2,....,v j ,....,vJ By combining different types of user-items into an interaction graph: Gu = (U, V, Eu), where the edge set Eu represents K heterogeneous relationships between users and items (such as browsing, adding to favorites, and shopping cart). In Eu, each edge Indicates that user u in the scenario of behavior type k i and project v j In order to merge the side information of the project, a graph Gv = (V, Ev) is defined to describe the multiple dependencies between projects. Connect in project v j and v' j The meta-relationship between them is expressed as {(v j ,r,v' j )|v j ,v' j ∈V,r∈R}. R represents a set of relations that can be generated from different aspects, such as vj and v' j Belong to the same category, come from the same location, or interact with the same user under the same k behavior types. Take the user-item multi-behavior interaction graph Gu and the knowledge-aware item-item relationship graph Gv as input to generate the matching score of the session and the item. use A top-k list is generated for a given session, and the item with the highest score will be predicted as the next interaction item.

[0051] The multi-behavior interaction pattern embedding module aims to aggregate heterogeneous signals of multi-behavior patterns between a user and her interaction items for different treatments of multiple behaviors. To this end, an adaptive multi-behavior self-attention network is utilized in a global context enhancement mode. The multi-behavior interaction pattern embedding module consists of three submodules: temporal context encoding scheme, information propagation, and information aggregation.

[0052] (1) Temporal context coding scheme

[0053] In this embodiment, a temporal context encoding scheme is used to capture the impact of different types of user-item interactions in time-aware scenarios, so as to incorporate behavioral dynamics into the heterogeneous message aggregate structure. First, given user u under k behaviors i and v j Connections between Their corresponding interaction timestamps Mapped to time slots And use the sine function to generate the embedding vector This is driven by the Transformer's position embedding framework, that is, the time embedding vector is

[0054]

[0055] (2) Information dissemination stage

[0056] On the multi-behavior user-item interaction graph Gu, the graph attention mechanism is used to implement time-aware information propagation between source nodes and target nodes, which can be expressed as:

[0057]

[0058]

[0059] in, and Respectively represent the items from v j To user u i and from user u i To project v j Information dissemination. j Is the initialized project embedded in e j and the element-wise addition of the corresponding temporal context representation. Similarly, p i It is element-wise addition on the user side. Represents the h-head projection matrix of the k-th behavior type. Respectively represent the information p constructed j and p i The attention propagation weights learned above are calculated as follows:

[0060]

[0061] in, It is the head-specific query and value transformation of the K-th behavior type.

[0062] In order to incorporate global contextual information of different behavior types during message passing, a multi-channel parameter learning framework is used to learn an attention-based transformation matrix Design a basic transformation paradigm, which includes M parameter channels, namely They correspond to M latent projection subspaces, reflecting different aspects of the common behavioral context of different types. A gating mechanism is used to transform specific types:

[0063]

[0064] in, is the quantitative gating weight of the mth channel transition with respect to the kth behavior type. Generally speaking, the number of channels M should be smaller than the number of behavior types K, which is conducive to parameter-efficient heterogeneous message aggregation.

[0065] (3) Information aggregation stage

[0066] Based on the constructed propagation information, the neighbor information is aggregated using the sum operation:

[0067]

[0068] Among them, N i and N j They represent user u in the interaction graph Gu respectively. i and project v j Neighbors. and is u i and v j Aggregate information of . f(·) is an activation function.

[0069] The heterogeneous signals of item-item interdependencies are fused with attention aggregation and expressed as follows:

[0070]

[0071] Where r is the index of R different project relations, N j Represents v on the graph Gv j The adjacent nodes of .

[0072] Finally, based on MLP, the final representation of the session is obtained:

[0073]

[0074] After obtaining session s, we input session s and the embedding i of the candidate j , and then use the softmax function to calculate the final score

[0075]

[0076] The binary cross entropy is used as the loss function of the SR task, which is expressed as follows:

[0077]

[0078] Where S and I are the session set and item set of the training sample. If the next interaction item of the user after session s is j, then y s,j =1, otherwise y s,j =0.

[0079] like Figure 1 As shown in Figure 1, in a KG project, there are often many-to-one and many-to-many relationships. PairRE is imported into the hyperplane to handle many-to-one and many-to-many relationships. Therefore, the training loss of PairRE is imported into the model to learn knowledge embedding. Specifically, for each attribute r, first, on the relationship-specific hyperplane wr Locate a relation-specific translation vector d in r . Given a triple<j,r,k> , the embedding vector j of item j and the embedding vector k of attribute k are firstly transformed into w r is the projection of the normal vector onto the hyperplane, and the projection is denoted by j ⊥ and k ⊥ .if<j,r,a> is correct, think j ⊥ and k ⊥ It can be achieved by a translation vector d on the hyperplane r Therefore, the score function is used to measure the plausibility of the triple being incorrect. r and j to represent j ⊥ , the formula is as follows:

[0080]

[0081] The loss function for knowledge embedding learning is:

[0082]

[0083] Among them, κ is the set of all knowledge triples.

[0084] Generally speaking, if two learning tasks are related to each other, merging the two learning tasks into the MLT paradigm is more effective than achieving their respective goals separately. In the MLT paradigm, the learning results of one task can be used as a hint to guide the learning of another task. Therefore, learning knowledge embedding can be regarded as an auxiliary task of the SR prediction task. Therefore, in MKGTN, knowledge embedding learning is imported into the MTL paradigm as an auxiliary task to assist the SR task.

[0085] The goal of MLT is to maximize the posterior probability of the model parameter Φ given the knowledge triple set κ and the SR training set y. According to the Bayesian rule, it is written as:

[0086]

[0087] where p(Φ) is the prior probability of Φ, which is set to follow a Gaussian distribution with zero mean and 0.1 standard deviation.

[0088] The comprehensive loss function of the MLT objective is:

[0089]

[0090] in, It is a regularization term to prevent overfitting. λ1 and λ2 are control parameters, and the values ​​of λ1 and λ2 are obtained through adjustment experiments.

[0091] This embodiment proposes to fuse multiple types of behavioral interactions and dependencies between items for the conversational recommendation algorithm, and apply them to the conversational recommendation. At the same time, project knowledge is used as an auxiliary task for the conversation that fuses user-item multi-behavioral interactions and item-item dependencies. This method studies the recommended HR and NDCG; fuses independent user-item multi-behavioral interactions and item-item dependencies through MLP and applies them to the conversation; incorporates project knowledge into the model through the MLT paradigm, which uses project knowledge as an auxiliary task for the conversation that fuses user-item multi-behavioral interactions and item-item dependencies; and applies the model to three datasets: Yelp, MovieLens, and Online Retail. Experiments show that the model has better recommendation performance than several other baseline methods. In addition, through ablation experiments, the influence of several sub-modules of the model on the model is determined.

[0092] Embodiment 2

[0093] The second embodiment of the present disclosure introduces a graph conversion network conversation recommendation system based on multi-behavior and project knowledge.

[0094] like Figure 3 A graph transformation network conversation recommendation system based on multi-behavior and item knowledge is shown, comprising:

[0095] an acquisition module configured to acquire a user-item multi-behavior interaction graph and a knowledge-aware item-item relationship graph;

[0096] A fusion module configured to fuse the acquired user-item multi-behavior interaction graph and the knowledge-aware item-item relationship graph based on multi-behavior interaction pattern embedding to obtain a session loss function;

[0097] An embedding module is configured to learn project knowledge through a knowledge graph embedding translation model based on pairwise relationship vectors, and to assist in learning user-project multi-behavior interaction relationships and project-project dependency relationships, thereby obtaining a project knowledge loss function;

[0098] The recommendation module is configured to obtain a comprehensive loss function based on multi-link aggregation and fusion of the recommendation loss function and the item knowledge loss function to achieve session recommendation.

[0099] The detailed steps are the same as those of the graph conversion network session recommendation method based on multi-behavior and project knowledge provided in Example 1, and will not be repeated here.

[0100] Embodiment 3

[0101] A third embodiment of the present disclosure provides a computer-readable storage medium.

[0102] A computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the method for recommending a graph-converted network session based on multi-behavior and item knowledge as described in the first embodiment of the present disclosure.

[0103] The detailed steps are the same as those of the graph conversion network session recommendation method based on multi-behavior and project knowledge provided in Example 1, and will not be repeated here.

[0104] Embodiment 4

[0105] A fourth embodiment of the present disclosure provides an electronic device.

[0106] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for recommending a graph-converted network session based on multi-behavior and project knowledge as described in the first embodiment of the present disclosure are implemented.

[0107] The detailed steps are the same as those of the graph conversion network session recommendation method based on multi-behavior and project knowledge provided in Example 1, and will not be repeated here.

[0108] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A graph conversion network session recommendation method based on multi-behavior and item knowledge, characterized in that: include: Obtain user-item multi-behavior interaction graphs and knowledge-aware item-item relationship graphs; Based on the multi-behavior interaction pattern embedding, the acquired user-item multi-behavior interaction graph and the knowledge-aware item-item relationship graph are fused by MLP to obtain the session loss function; The project knowledge is learned through the knowledge graph embedding translation model based on the pairwise relationship vector, and the user-project multi-behavior interaction relationship and project-project dependency relationship are assisted in learning to obtain the project knowledge loss function; Based on the multi-link aggregation fusion of the session loss function and the item knowledge loss function, a comprehensive loss function is obtained to achieve session recommendation; Among them, based on time context encoding, the behavioral dynamic information used to characterize the interaction between different users and items in the time perception scenario is obtained, and the obtained behavioral dynamic information is incorporated into the heterogeneous signal; The graph attention mechanism is used to propagate time-aware information between source nodes and target nodes in the obtained user-item multi-behavior interaction graph. In the process of time-aware information propagation, the global context information of different behavior types is merged, and the attention transfer matrix is ​​processed using multi-channel parameter learning. In the process of embedding project knowledge, the knowledge graph embedding translation model based on pairwise relationship vectors is imported into the hyperplane to handle many-to-one and many-to-many project-project relationships; The item knowledge embedding learning is imported into the multi-link aggregation paradigm as an auxiliary task, and the learning results of the task are used as prompts to guide the learning of another task, assisting in completing the conversational recommendation task.

2. A method for recommending network conversations based on graph transformation based on multiple behaviors and item knowledge as claimed in claim 1, characterized in that: In the process of obtaining the user-item multi-behavior interaction graph, the different users of the session recommendation and the items corresponding to each user are determined. According to the order in which the users click on the items, the user-item correspondence relationship and the heterogeneous relationship between the users and the session recommendation items are formed to construct the user-item multi-behavior interaction graph.

3. A method for recommending network conversations based on graph transformation based on multiple behaviors and item knowledge as claimed in claim 2, characterized in that: The project-project relationship diagram perceived based on the acquired knowledge represents the multiple dependencies between projects.

4. A graph conversion network conversation recommendation system based on multiple behaviors and project knowledge, adopting the graph conversion network conversation recommendation method based on multiple behaviors and project knowledge as described in any one of claims 1-3, characterized in that: include: an acquisition module configured to acquire a user-item multi-behavior interaction graph and a knowledge-aware item-item relationship graph; A fusion module configured to fuse the obtained user-item multi-behavior interaction graph and the knowledge-aware item-item relationship graph based on multi-behavior interaction pattern embedding to obtain a recommendation loss function; An embedding module is configured to learn project knowledge through a knowledge graph embedding translation model based on pairwise relationship vectors, and to assist in learning user-project multi-behavior interaction relationships and project-project dependency relationships, thereby obtaining a project knowledge loss function; The recommendation module is configured to obtain a comprehensive loss function based on multi-link aggregation and fusion of the recommendation loss function and the item knowledge loss function to achieve session recommendation.

5. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the graph conversion network session recommendation method based on multi-behavior and item knowledge as described in any one of claims 1 to 3 are implemented.

6. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the graph conversion network session recommendation method based on multi-behavior and item knowledge as described in any one of claims 1 to 3 are implemented.

Images