A multi-behavior recommendation method based on knowledge interaction graph

Abstract

The application belongs to the technical field of deep learning and recommendation system, and particularly relates to a multi-behavior recommendation method based on a knowledge interaction graph, comprising: acquiring data of users and items, constructing a knowledge interaction graph according to the data of users and items, inputting the knowledge interaction graph into a trained user multi-behavior recommendation model to obtain an interaction probability of users and items; the user multi-behavior recommendation model comprises a knowledge interaction graph embedding module, a behavior knowledge perception graph attention module and a prediction module; the application constructs a unified knowledge interaction graph by combining a multi-behavior interaction graph and an item knowledge graph, models complex dependency relationships between multiple behaviors which are not affected by knowledge relationships, and updates and processes the knowledge interaction graph through the behavior knowledge perception graph attention module, thereby revealing potential dependency relationships between behaviors and knowledge, overcoming the limitations of traditional multi-behavior recommendation methods in behavior dependency relationship modeling, and improving the recommendation effect.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and recommendation system technology, specifically relating to a multi-behavior recommendation method based on knowledge interaction graphs. Background Technology

[0002] Recommender systems have become an important tool for alleviating information overload in various fields, such as e-commerce. The key to personalized recommendations is accurately modeling user preferences. A traditional paradigm for recommender systems is collaborative filtering, which extracts user preferences from historical interactions between users and items. In recent years, many deep neural network-related recommender models have been proposed, but most are limited to research on single behaviors.

[0003] To address the data sparsity problem associated with single-action recommendations, multi-action recommendation algorithms have been proposed. In many real-world recommendation scenarios, there are various types of behavioral interactions between users and items. For example, on e-commerce platforms, auxiliary behaviors such as "click," "favorite," and "add to cart" can provide useful information for recommending the target behavior of "purchase," which directly generates revenue for the platform.

[0004] Given its significant practical value, multi-behavior recommendation has been extensively and deeply studied. Graph neural networks, contrastive learning, and meta-learning have been successively introduced into the field of multi-behavior recommendation algorithms, achieving some success. On the other hand, some studies have introduced knowledge graphs into the field of recommendation algorithms, utilizing the rich semantic information in knowledge graphs to enhance the embedding representation of items. Although numerous experiments have confirmed that the above methods improve recommendation performance and effectiveness to some extent, some problems remain unresolved.

[0005] Existing knowledge graph-based multi-behavior recommendation methods typically involve explicitly modeling user multi-behavior dependencies in user item interactions to enrich user representations, and then using item knowledge graphs to enrich item representations. However, user behavior is influenced by item knowledge, and there are complex dependencies between behaviors influenced by item knowledge. The aforementioned existing technologies do not take into account the potential dependencies between user item interactions and item-item relationships in the knowledge graph. Summary of the Invention

[0006] To address the aforementioned problems in the prior art, this invention employs a multi-behavior recommendation method based on a knowledge interaction graph, comprising: acquiring user and item data; constructing a knowledge interaction graph based on the user and item data; inputting the knowledge interaction graph into a trained user multi-behavior recommendation model to obtain the interaction probabilities of users and items; the user multi-behavior recommendation model comprises: a knowledge interaction graph embedding module, a behavior knowledge perception graph attention module, and a prediction module;

[0007] The training process of a user multi-behavior recommendation model includes:

[0008] S1: Obtain the user project dataset, construct a user multi-behavior interaction graph and a project knowledge graph based on the user project dataset, and construct a knowledge interaction graph based on the user multi-behavior interaction graph and the project knowledge graph. ,in, For a collection of entities, For a set of relations, triples Represents the head entity With tail body There is a relationship ;

[0009] S2: Input the knowledge interaction graph into the knowledge interaction graph embedding module to obtain the embedded knowledge interaction graph;

[0010] S3: Divide the embedded knowledge interaction graph into multiple knowledge interaction subgraphs based on the types of user behaviors in the user project dataset. Where k is the index of the user behavior type;

[0011] S4: Connect each knowledge interaction subgraph The input behavior knowledge-aware graph attention module yields the final embedded representations of users and items.

[0012] S5: Input the final embedded representations of users and items into the prediction module to obtain the interaction probability between users and items;

[0013] S6: Calculate the loss function value based on the embedded knowledge interaction graph and the user-item interaction probability, update the model parameters based on the loss function value, and obtain the trained user multi-behavior recommendation model when the preset number of training times is reached.

[0014] User multi-behavior interaction graph ,in, Let be the set of edges, representing the user's... This behavior, Represents a user set, Indicates a set of items. Indicates user Under behavior type k and items There was interaction. Indicates user Under behavior type k and items There is no interaction; i and j are the indices for the user and project, respectively; project knowledge graph. ,in, Represents a set of entities. Represents a set of relations, a triple. Represents the head entity With tail body There is a relationship Constructing a knowledge interaction graph based on user multi-behavior interaction graphs and project knowledge graphs includes: [This involves] constructing a knowledge interaction graph from items in the user multi-behavior interaction graph. Aligning with entities in the knowledge graph yields a knowledge interaction graph. ,in, , .

[0015] The knowledge interaction graph embedding module is a TransR model, which includes a relation mapping matrix for each relation r. The TransR model processes the knowledge interaction graph by embedding it to obtain each triple. Embedded vector According to the relational mapping matrix embed vector Projection to correspondence In the space, the projection vector is obtained. , Among them, the TransR model is a relational projection model.

[0016] The behavioral knowledge perception graph attention module consists of multiple sequentially connected behavioral knowledge perception graph attention layers; the behavioral knowledge perception graph attention module processes each knowledge interaction subgraph. The processing includes:

[0017] S41. Knowledge Interaction Subgraph As the initial knowledge interaction subgraph Each knowledge interaction subgraph The first line of input is the knowledge perception graph attention layer, which yields the knowledge interaction subgraph. ;

[0018] S42. Transform the knowledge interaction subgraph The second line of input is the knowledge perception graph attention layer, which yields the knowledge interaction subgraph. ;

[0019] S43, Move the previous row to the knowledge perception map attention layer Output knowledge interaction subgraph Input current behavior knowledge perception graph attention layer Obtain the knowledge interaction subgraph ;

[0020] S44. Repeat step S43 until the knowledge interaction subgraph output by the last layer of the behavior knowledge perception graph attention layer is obtained.

[0021] S45. All knowledge interaction subgraphs The embedding vectors of the entities are concatenated to obtain the embedding representations of users and projects.

[0022] Each layer of the behavior knowledge perception graph attention layer includes a behavior knowledge perception attention layer, an information aggregation layer, and a cross-behavior dependent self-attention layer; the behavior knowledge perception graph attention layer Attention layer of behavioral knowledge perception map Output knowledge interaction subgraph The processing includes: dividing the knowledge interaction subgraph Input behavior knowledge perception attention layer l, and obtain the embedding vector of the head entity after the knowledge interaction subgraph is updated. , knowledge interaction subgraph Embedding vector of the head entity and the updated head entity embedding vector The input information aggregation layer yields the embedding vectors of the head entities after the knowledge interaction subgraph is aggregated. The embedding vector of the head entity after aggregating the knowledge interaction subgraph. Input cross-behavior dependencies are derived from the attention layer to obtain a knowledge interaction subgraph. and the embedding vector of its head entity .

[0023] Behavioral knowledge perception attention layer Knowledge interaction subgraph The processing includes:

[0024] in, Represents a knowledge interaction subgraph The tail entity vector, Indicates the relationship with the head entity The relevant set of triples, Represents the knowledge interaction subgraph tail entity Along the relationship To the head entity Attention weight in the spread.

[0025] Attention weight The calculation process includes: based on the knowledge interaction subgraph Calculate relational attention weights from the embedding vectors of entities. And behavioral knowledge perception attention weight , weight and weight Add them together to get the overall attention weight. For comprehensive attention weight Normalization is performed to obtain the attention weights. .

[0026] Calculate relational attention weights include:

[0027]

[0028] in, This is the activation function.

[0029] Calculate behavioral knowledge perception attention weights include:

[0030] in, and Let be the transformation matrix. This is the activation function.

[0031] Cross-behavior dependent on the head entity vector after aggregating the knowledge interaction subgraph by the attention layer The processing includes: aggregating the header entity vectors of each knowledge interaction subgraph. Concatenate the vectors to obtain the concatenated vector. According to the concatenated vector Calculate attention scores for cross-behavioral type dependencies Attention score Multiply by the concatenated vector This yields the final head entity vector. .

[0032] Beneficial effects:

[0033] This invention constructs a unified knowledge interaction graph by combining multi-behavior interaction graphs and project knowledge graphs, models the complex dependencies between various behaviors influenced by knowledge relationships, and processes the knowledge interaction graph through a behavior knowledge perception graph attention module, revealing the potential dependencies between behaviors and knowledge. This overcomes the limitations of traditional multi-behavior recommendation methods in modeling behavior dependencies and improves recommendation performance. Attached Figure Description

[0034] Figure 1 This is a flowchart of a multi-behavior recommendation method based on a knowledge interaction graph provided in an embodiment of the present invention;

[0035] Figure 2 This is a framework diagram of a multi-behavior recommendation method based on a knowledge interaction graph provided in an embodiment of the present invention;

[0036] Figure 3 This is a detailed diagram of the behavior knowledge perception graph attention module provided in an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] like Figure 1 , Figure 2 As shown in the figure, this embodiment of the invention proposes a multi-behavior recommendation method based on a knowledge interaction graph, which specifically includes the following: acquiring user and item data, constructing a knowledge interaction graph based on the user and item data, inputting the knowledge interaction graph into a trained user multi-behavior recommendation model, and obtaining the interaction probability of users and items; the user multi-behavior recommendation model includes: a knowledge interaction graph embedding module, a behavior knowledge perception graph attention module, and a prediction module;

[0039] The training process of a user multi-behavior recommendation model includes:

[0040] S1: Obtain the user project dataset, construct the user multi-behavior interaction graph and the project knowledge graph based on the user project dataset, and construct the knowledge interaction graph based on the user multi-behavior interaction graph and the project knowledge graph.

[0041] User multi-behavior interaction graph is defined as ,in, Let K be the set of K types of edges on the graph, including the observed ones. This behavior, Represents a user set, Indicates a set of items; Indicates user Under behavior type k and items There was interaction. Indicates user Under behavior type k and items There is no interaction; i and j are the indices for the user and the project, respectively.

[0042] Project Knowledge Graph ,in, This represents a collection of entities, specifically a collection of items. Represents a set of relations; each triple Represents the head entity With tail body There are semantic relationships between them For example, triples This indicates that director B is the director of film A.

[0043] To combine user multi-behavior interaction graphs and project knowledge graphs, this embodiment establishes a set of project-entity alignments. ,in, Represents items in a multi-behavior interaction diagram. Can be associated with entities in a knowledge graph Alignment; based on the alignment set By combining user multi-behavior interaction graphs and project knowledge graphs, a knowledge interaction graph is obtained. , where entity set Relation set .

[0044] S2: Input the knowledge interaction graph into the knowledge interaction graph embedding module to obtain the embedded knowledge interaction graph;

[0045] To parameterize entities and relations on the knowledge interaction graph into vector representations while preserving the overall structure of the graph, existing knowledge graph embedding techniques are employed. Considering the various types of relations between entities in the knowledge interaction graph, this embodiment utilizes the TransR (Relational Projection) model to model entities and relations. Specifically, entities and relations are embedded into the knowledge interaction graph to obtain each triple. Embedded representation For each relation r, a relation mapping matrix is ​​defined. This matrix will include the head entity. and tail body Projecting the original entity space onto a specific relation space yields the corresponding relation space. projection representation , According to the principle of translation .

[0046] The energy score of the triple Represented as:

[0047]

[0048] in, For relationship The relation mapping matrix, the relation will be the head entity and tail body from Projecting 3D space to 3D space The lower the score, the weaker the knowledge interaction graph. and The more likely there is a relationship between them ,vice versa.

[0049] Loss function based on embedded knowledge interaction graph The definition is as follows:

[0050]

[0051] Among them, set , These are triples that do not exist in the knowledge interaction graph, constructed by randomly replacing one entity in the real triple. This is the sigmoid function.

[0052] Since each type of behavior in this graph is treated as a relation in the knowledge graph, this module can capture the coarse-grained commonalities between multiple behaviors and knowledge relations, and the relation mapping matrix... This will be shared in subsequent modules.

[0053] S3: Divide the embedded knowledge interaction graph into multiple knowledge interaction subgraphs based on the types of user behaviors in the user project dataset. ;

[0054] To effectively explore the complex dependencies between individual behaviors and knowledge relationships, the embedded knowledge interaction graph is divided into... Zhang subgraph, where each subgraph contains only the first... Types of behavioral and knowledge relationships are formally defined as follows: .

[0055] S4: Connect each knowledge interaction subgraph The input behavior knowledge-aware graph attention module yields the final embedded representations of users and items.

[0056] like Figure 3 As shown, the behavior knowledge perception graph attention module includes multiple layers of behavior knowledge perception graph attention layers connected in sequence: each behavior knowledge perception graph attention layer includes a behavior knowledge perception attention layer, an information aggregation layer, and a cross-behavior dependent self-attention layer;

[0057] The behavioral knowledge perception graph attention module processes each knowledge interaction subgraph, including:

[0058] S41. Knowledge Interaction Subgraph As the initial knowledge interaction subgraph Each knowledge interaction subgraph The first line of input is the knowledge perception graph attention layer, which yields the knowledge interaction subgraph. ;

[0059] S42. Transform the knowledge interaction subgraph The second line of input is the knowledge perception graph attention layer, which yields the knowledge interaction subgraph. ;

[0060] S43, Move the previous row to the knowledge perception map attention layer Output knowledge interaction subgraph Input current behavior knowledge perception graph attention layer Obtain the knowledge interaction subgraph ;

[0061] Attention layer of behavioral knowledge perception graph Attention layer of behavioral knowledge perception map Output knowledge interaction subgraph The processing includes:

[0062] S431, Knowledge Interaction Subgraph Input behavior knowledge perception attention layer l, and obtain the updated head entity vector after knowledge interaction subgraph. :

[0063] in, Represents a knowledge interaction subgraph The tail entity vector, Indicates the relationship with the head entity The relevant set of triples, The knowledge interaction subgraph in the attention layer l represents the behavioral knowledge perception. tail entity Along the relationship To the head entity Attention weight in the spread.

[0064] Behavioral relationships and knowledge relationships convey different semantic information to the representation learning of knowledge interaction subgraphs. To capture the semantic differences between behavioral and knowledge relationships, attention weights are integrated. The calculation is based on the following formula for adaptive behavioral knowledge perception attention weights using a triplet:

[0065]

[0066]

[0067]

[0068] in, For relational attention weights, For behavioral knowledge, attention weights are perceived. These are the transformation matrices for behavioral relationships and knowledge relationships, respectively. It is a non-linear activation function.

[0069] To avoid bias, the softmax function is used to compare the current head entity. Normalize the attention weights of all relevant triples:

[0070]

[0071] in, For the head entity In relationship With tail body The attention weights were normalized. Representing relations lower head entity The number of tail entities j.

[0072] The above method mainly considers the propagation of two types of information in the knowledge interaction subgraph. On the one hand, the coarse-grained commonality between user behavior and knowledge relationships is represented by S2. On the one hand, modeling is used; on the other hand, behavioral relationships and knowledge relationships have different semantic information. For example, behavior represents the interaction between a user and an item, while knowledge relationships only connect different items and describe the attributes of the items. In order to capture the subtle semantic differences between the above-mentioned behavior and knowledge relationships, this embodiment considers them separately in the attention mechanism of information propagation. Specifically, it utilizes... and Model the differences if the relationship If it is one of the behaviors in a user multi-behavior interaction graph, then the user entity will be... and project entities By transforming the matrix Project it onto the behavior vector space; otherwise, pass it through... Projected into the vector space of knowledge relationships.

[0073] S432. Update the head entity vector of each knowledge interaction subgraph before it is updated. and the updated head entity vector The input information aggregation layer yields the header entity vector after aggregating each knowledge interaction subgraph. :

[0074] in, , For a trainable parameter matrix, To calculate the mean function, For activation function, Indicates splicing.

[0075] S433, Aggregate the head entity vector of each knowledge interaction subgraph The input cross-behavior depends on the attention layer, resulting in the final head entity vector. .

[0076] Since different knowledge interaction subgraphs contain only one behavioral relationship, the same node often exhibits different feature representations in different subgraphs. The previous section revealed the potential dependency between single behaviors and knowledge relationships. Considering factors such as user behavior habits, there are always complex interdependencies between behaviors. Therefore, in order to reveal the potential dependencies between multiple behaviors influenced by knowledge relationships, a cross-behavior dependency self-attention mechanism was designed.

[0077] Specifically, the header entity vector after aggregating each knowledge interaction subgraph Perform concatenation (Concat) to obtain the concatenated vector. ;

[0078] Based on the splicing vector Calculate attention scores for cross-behavioral type dependencies :

[0079]

[0080] in , These are two behavior-related parameters.

[0081] Attention score Multiply by the concatenated vector This yields the final head entity vector. :

[0082]

[0083] S44. Repeat step S43 until the knowledge interaction subgraph output by the last layer of the behavior knowledge perception graph attention layer is obtained.

[0084] S45, All knowledge interaction subgraphs The embedding vectors of the entities are concatenated to obtain the embedding representations of users and items:

[0085]

[0086]

[0087] in, For users Embedded representation, For the project Embedded representation.

[0088] S5: Input the final embedded representations of users and items into the prediction module to obtain the interaction probability between users and items;

[0089] Finally, the target behavior is predicted by taking the inner product of the user embedding and the item embedding. Interaction probability :

[0090]

[0091] S6: Calculate the loss function value based on the embedded knowledge interaction graph and the user-item interaction probability, update the model parameters based on the loss function value, and obtain the trained user multi-behavior recommendation model when the preset number of training times is reached.

[0092] The BPRLoss (Bayesian Personalized Ranking Loss) loss function is defined as follows:

[0093]

[0094] in, Represents the user project dataset. Indicates user In target behavior Below and Project There is interaction. Indicates user In target behavior Below and Project No interaction; finally, joint loss function The definition is as follows:

[0095]

[0096] in, This represents the set of parameters for a multi-behavior recommendation model. As weight.

[0097] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-behavior recommendation method based on knowledge interaction graphs, characterized in that, include: Acquire user and project data, construct a knowledge interaction graph based on the user and project data, input the knowledge interaction graph into a trained user multi-behavior recommendation model, and obtain the interaction probability of users and projects; The user multi-behavior recommendation model includes: a knowledge interaction graph embedding module, a behavior knowledge perception graph attention module, and a prediction module; The training process of a user multi-behavior recommendation model includes: S1: Obtain the user project dataset, construct a user multi-behavior interaction graph and a project knowledge graph based on the user project dataset, and construct a knowledge interaction graph based on the user multi-behavior interaction graph and the project knowledge graph. ,in, For a collection of entities, For a set of relations, triples Represents the head entity With tail body There is a relationship ; S2: Input the knowledge interaction graph into the knowledge interaction graph embedding module to obtain the embedded knowledge interaction graph; S3: Divide the embedded knowledge interaction graph into multiple knowledge interaction subgraphs based on the types of user behaviors in the user project dataset. Where k is the index of the user behavior type; S4: Connect each knowledge interaction subgraph The input behavior knowledge-aware graph attention module yields the final embedded representations of users and items. The behavioral knowledge perception graph attention module consists of multiple sequentially connected behavioral knowledge perception graph attention layers; each behavioral knowledge perception graph attention layer includes a behavioral knowledge perception attention layer, an information aggregation layer, and a cross-behavioral dependent attention layer. S5: Input the final embedded representations of users and items into the prediction module to obtain the interaction probability between users and items; S6: Calculate the loss function value based on the embedded knowledge interaction graph and the user-item interaction probability, update the model parameters based on the loss function value, and obtain the trained user multi-behavior recommendation model when the preset number of training times is reached.

2. The multi-behavior recommendation method based on knowledge interaction graphs according to claim 1, characterized in that, User multi-behavior interaction graph ,in, Let be the set of edges, representing the user's... This behavior, Represents a user set, Indicates a set of items. Indicates user Under behavior type k and items There was interaction. Indicates user Under behavior type k and items There is no interaction; i and j are the indices for the user and project, respectively; project knowledge graph. ,in, Represents a set of entities. Represents a set of relations, a triple. Represents the head entity With tail body There is a relationship Constructing a knowledge interaction graph based on user multi-behavior interaction graphs and project knowledge graphs includes: [This involves] constructing a knowledge interaction graph from items in the user multi-behavior interaction graph. Aligning with entities in the knowledge graph yields a knowledge interaction graph. ,in, , .

3. The multi-behavior recommendation method based on knowledge interaction graphs according to claim 2, characterized in that, The knowledge interaction graph embedding module is a TransR model, which includes a relation mapping matrix for each relation r. ; The TransR model processes knowledge interaction graphs by embedding them to obtain each triplet. Embedded vector According to the relational mapping matrix embed vector Projection to correspondence In the space, the projection vector is obtained. , Among them, the TransR model is a relational projection model.

4. The multi-behavior recommendation method based on knowledge interaction graphs according to claim 3, characterized in that, The behavioral knowledge perception graph attention module focuses on each knowledge interaction subgraph. The processing includes: S41. Knowledge Interaction Subgraph As the initial knowledge interaction subgraph Each knowledge interaction subgraph The first line of input is the knowledge perception graph attention layer, which yields the knowledge interaction subgraph. ; S42. Transform the knowledge interaction subgraph The second line of input is the knowledge perception graph attention layer, which yields the knowledge interaction subgraph. ; S43, Move the previous row to the knowledge perception map attention layer Output knowledge interaction subgraph Input current behavior knowledge perception graph attention layer Obtain the knowledge interaction subgraph ; S44. Repeat step S43 until the knowledge interaction subgraph output by the last layer of the behavior knowledge perception graph attention layer is obtained. S45, All knowledge interaction subgraphs The embedding vectors of the entities are concatenated to obtain the embedding representations of users and projects.

5. The multi-behavior recommendation method based on knowledge interaction graphs according to claim 4, characterized in that, Attention layer of behavioral knowledge perception graph Attention layer of behavioral knowledge perception map Output knowledge interaction subgraph The processing includes: dividing the knowledge interaction subgraph Input behavior knowledge perception attention layer l, and obtain the embedding vector of the head entity after the knowledge interaction subgraph is updated. , knowledge interaction subgraph Embedding vector of the head entity and the updated head entity embedding vector The input information aggregation layer yields the embedding vectors of the head entities after the knowledge interaction subgraph is aggregated. The embedding vector of the head entity after aggregating the knowledge interaction subgraph. Input cross-behavior dependencies are derived from the attention layer to obtain a knowledge interaction subgraph. and the embedding vector of its head entity .

6. The multi-behavior recommendation method based on knowledge interaction graphs according to claim 5, characterized in that, Behavioral knowledge perception attention layer Knowledge interaction subgraph The processing includes: ； in, Represents a knowledge interaction subgraph The tail entity vector, Indicates the relationship with the head entity The relevant set of triples, Represents the knowledge interaction subgraph tail entity Along the relationship To the head entity Attention weight in the spread.

7. The multi-behavior recommendation method based on knowledge interaction graphs according to claim 6, characterized in that, Attention weight The calculation process includes: based on the knowledge interaction subgraph Calculate relational attention weights from the embedding vectors of entities. And behavioral knowledge perception attention weight , weight and weight Add them together to get the overall attention weight. For comprehensive attention weight Normalization is performed to obtain the attention weights. .

8. The multi-behavior recommendation method based on knowledge interaction graphs according to claim 6, characterized in that, Calculate relational attention weights include: ； in, This is the activation function.

9. A multi-behavior recommendation method based on a knowledge interaction graph according to claim 6, characterized in that, Calculate behavioral knowledge perception attention weights include: ； in, and The transformation matrix is... This is the activation function.

10. A multi-behavior recommendation method based on a knowledge interaction graph according to claim 5, characterized in that, Cross-behavior dependent on the head entity vector after aggregating the knowledge interaction subgraph by the attention layer The processing includes: aggregating the header entity vectors of each knowledge interaction subgraph. Concatenate the vectors to obtain the concatenated vector. According to the concatenated vector Calculate attention scores for cross-behavioral type dependencies Attention score Multiply by the concatenated vector This yields the final head entity vector. .

Images