Multi-behavior recommendation method based on knowledge interaction graph

Abstract

The invention belongs to the technical field of deep learning and recommendation systems, and particularly relates to a multi-behavior recommendation method based on a knowledge interaction graph, which comprises the following steps: acquiring data of a user and a project, constructing the knowledge interaction graph according to the data of the user and the project, and inputting the knowledge interaction graph into a trained user multi-behavior recommendation model, the interaction probability of the user and the project is obtained; the user multi-behavior recommendation model comprises a knowledge interaction graph embedding module, a behavior knowledge perception graph attention module and a prediction module; according to the invention, the unified knowledge interaction graph is constructed by combining the multi-behavior interaction graph and the project knowledge graph, modeling is carried out on a complex dependency relationship among multiple behaviors which are not influenced by the knowledge relationship, and the knowledge interaction graph is updated through the behavior knowledge perception graph attention module. The potential dependency relationship between behaviors and knowledge is revealed, the limitation of a traditional multi-behavior recommendation method in the aspect of behavior dependency relationship modeling is overcome, and the recommendation effect is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and recommendation systems, and specifically relates to a multi-behavior recommendation method based on a knowledge interaction graph. Background Art

[0002] Recommendation systems have become an important tool to alleviate information overload in various fields, such as e-commerce. The key to personalized recommendations is how to accurately model user preferences. The traditional recommendation system paradigm is collaborative filtering, which extracts user preferences from historical interactions between users and items. In recent years, many deep neural network-related recommendation models have been proposed, but most of them are limited to the study of a single behavior.

[0003] In order to make up for the data sparsity problem of a single behavior, a multi-behavior recommendation algorithm was proposed. In many actual recommendation scenarios, there are multiple types of behavioral interactions between users and items. For example, on e-commerce platforms, auxiliary behaviors such as "click", "collect", and "add to cart" can provide useful information for the recommendation of the target behavior "purchase", and the "purchase" behavior can directly generate revenue for the platform.

[0004] Given its significant practical value, the problem of multi-behavior recommendation has been studied in depth and extensively. Graph neural networks, contrastive learning, meta-learning, etc. have been introduced into the field of multi-behavior recommendation algorithms and have achieved certain success. On the other hand, some studies have introduced knowledge graphs into the field of recommendation algorithms, using the rich semantic information in knowledge graphs to enhance the embedding representation of the project side. Although numerous experiments have confirmed that the above methods have improved the recommendation performance and effect to a certain extent, there are still some problems that have not been solved.

[0005] Existing multi-behavior recommendation methods based on knowledge graphs usually explicitly model user multi-behavior dependencies in user-item interactions to enrich user representations, and then use item knowledge graphs to enrich item representations. However, user behaviors are affected by item knowledge, and there are complex dependencies between behaviors affected by item knowledge. The above-mentioned existing technologies do not take into account the potential dependencies between user-item interactions and item-item relationships in knowledge graphs. Summary of the invention

[0006] In order to solve the above-mentioned problems in the prior art, the present invention adopts a multi-behavior recommendation method based on a knowledge interaction graph, including: obtaining user and project data, constructing a knowledge interaction graph according to the user and project data, inputting the knowledge interaction graph into a trained user multi-behavior recommendation model, and obtaining the interaction probability of the user and the project; the user multi-behavior recommendation model includes: a knowledge interaction graph embedding module, a behavior knowledge perception graph attention module and a prediction module;

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

[0008] S1: Obtain the user-item dataset, build the user multi-behavior interaction graph and the item knowledge graph based on the user-item dataset, and build the knowledge interaction graph G = {(h, r, t) | h, t∈E based on the user multi-behavior interaction graph and the item knowledge graph ′ ,r∈R ′}, where E ′ is the entity set, R ′ is a relationship set, and the triple (h, r, t) indicates that there is a relationship r between the head entity h and the tail entity t;

[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 G according to the types of user behaviors in the user-project dataset k ; Where k is the index of the user behavior type;

[0011] S4: Each knowledge interaction subgraph G k Input the behavior knowledge perception graph attention module to obtain the final embedding representation of users and items;

[0012] S5: Input the final embedding representation of the user and item into the prediction module to obtain the interaction probability between the user and the item;

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

[0014] User multi-behavior interaction diagram in, is a set of edges, representing K types of user behaviors. represents the user set, represents a set of items, Indicates user Under behavior type k and project There was interaction. Indicates user Under behavior type k and project There is no interaction, i and j are the indexes of the user and the project respectively; the project knowledge graph G2 = {(h, r, t) | h, t∈E, r∈R}, where E represents the entity set, R represents the relationship set, and the triple (h, r, t) represents that there is a relationship r between the head entity h and the tail entity t; constructing a knowledge interaction graph based on the user multi-behavior interaction graph and the project knowledge graph includes: by placing the project v in the user multi-behavior interaction graph jAlign with the entities in the knowledge graph to obtain the knowledge interaction graph G = {(h,r,t)|h,t∈E ′ ,r∈R ′},in, R ′ =R∪θ.

[0015] The knowledge interaction graph embedding module is the TransR model, which includes the relation mapping matrix W for each relation r r ; The TransR model processes the knowledge interaction graph by: embedding the knowledge interaction graph to obtain the embedding vector (x h ,x r ,x t ); According to the relationship mapping matrix W r Embed the vector x h 、x t Projected into the space of the corresponding relationship r, we get the projection vector Among them, the TransR model is a relational projection model.

[0016] The behavior knowledge perception graph attention module includes multiple layers of behavior knowledge perception graph attention layers connected in series; the behavior knowledge perception graph attention module performs a k Processing includes:

[0017] S41, the knowledge interaction subgraph G k As the initial knowledge interaction subgraph Each knowledge interaction subgraph Input the first line as the knowledge perception graph attention layer to obtain the knowledge interaction subgraph

[0018] S42, knowledge interaction subgraph Input the second behavior of the knowledge perception graph attention layer to obtain the knowledge interaction subgraph

[0019] S43, the knowledge interaction subgraph output by the attention layer l-1 of the previous behavior knowledge perception graph Input the current behavior knowledge perception graph attention layer l to obtain the knowledge interaction subgraph

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

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

[0022] Each behavior knowledge perception graph attention layer includes a behavior knowledge perception attention layer, an information aggregation layer, and a cross-behavior dependency self-attention layer; the behavior knowledge perception graph attention layer l outputs the knowledge interaction subgraph of the behavior knowledge perception graph attention layer l-1. The processing includes: transforming the knowledge interaction subgraph Input the behavior knowledge perception attention layer l to obtain the embedding vector of the head entity after the knowledge interaction subgraph is updated Knowledge Interaction Subgraph The embedding vector of the head entity and the updated embedding vector of the head entity Input the information aggregation layer to obtain the embedding vector of the head entity after the knowledge interaction subgraph is aggregated The embedding vector of the head entity after aggregating the knowledge interaction subgraph Input the cross-behavior dependency self-attention layer to obtain the knowledge interaction subgraph and its head entity embedding vector

[0023] Behavior knowledge-aware attention layer l on knowledge interaction subgraph Processing includes:

[0024]

[0025] in, Representing knowledge interaction subgraph The tail entity vector, N n ={(h,r,t)|(h,r,t)∈G k} represents the set of triples related to the head entity h, Represents the knowledge interaction subgraph The attention weight of the tail entity t propagated along the relation r to the head entity h.

[0026] Attention Weight The calculation process includes: according to the knowledge interaction subgraph The embedding vector of the entity in the relation is used to calculate the attention weight of the relation and behavioral knowledge-aware attention weights Weight and weight Add together to get the comprehensive attention weight For the comprehensive attention weight Normalize and get the attention weight

[0027] Calculating relation attention weights include:

[0028]

[0029] Among them, tanh is the activation function.

[0030] Calculating behavioral knowledge-aware attention weights include:

[0031]

[0032] Among them, W b and W k is the transformation matrix and tanh is the activation function.

[0033] Head entity vector after cross-behavior dependent self-attention layer aggregates knowledge interaction subgraphs The processing includes: aggregating the head entity vector of each knowledge interaction subgraph Perform splicing to obtain the splicing vector According to the splicing vector Computing attention scores across behavior type dependencies Score the attention Multiply by the concatenation vector Get the final head entity vector

[0034] Beneficial effects:

[0035] The present invention constructs a unified knowledge interaction graph by combining multiple behavior interaction graphs and project knowledge graphs, models the complex dependency relationships between multiple behaviors affected by knowledge relationships, and processes the knowledge interaction graph through the behavior knowledge perception graph attention module, revealing the potential dependency relationship between behavior and knowledge. This overcomes the limitations of traditional multi-behavior recommendation methods in modeling behavior dependency relationships and improves the recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of a multi-behavior recommendation method based on a knowledge interaction graph provided by an embodiment of the present invention;

[0037] Figure 2 It is a framework diagram of a multi-behavior recommendation method based on a knowledge interaction graph provided by an embodiment of the present invention;

[0038] Figure 3 It is a detailed diagram of the attention module of the behavioral knowledge perception graph provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in 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.

[0040] like Figure 1 , Figure 2 As shown, the embodiment of the present invention proposes a multi-behavior recommendation method based on a knowledge interaction graph, which specifically includes the following contents: obtaining user and project data, constructing a knowledge interaction graph according to the user and project data, inputting the knowledge interaction graph into a trained user multi-behavior recommendation model, and obtaining the interaction probability of the user and the project; the user multi-behavior recommendation model includes: a knowledge interaction graph embedding module, a behavior knowledge perception graph attention module, and a prediction module;

[0041] The training process of the user multi-behavior recommendation model includes:

[0042] S1: Obtain a user-project dataset, build a user multi-behavior interaction graph and a project knowledge graph based on the user-project dataset, and build a knowledge interaction graph based on the user multi-behavior interaction graph and the project knowledge graph;

[0043] The user multi-behavior interaction graph is defined as in, is a set of K edges on the graph, including K observed behaviors, represents the user set, Represents a set of items; Indicates user Under behavior type k and project There was interaction. Indicates user Under behavior type k and project There is no interaction, i and j are the indexes of the user and item respectively.

[0044] Project knowledge graph G2 = {(h,r,t)|h,t∈E,r∈R}, where E represents an entity set, i.e., a project set, and R represents a relationship set; each triple (h,r,t) represents a semantic relationship r between the head entity h and the tail entity t. For example, the triple (Titanic, directedBy, JanmesCameron) represents that JanmesCameron is the director of the movie Titanic.

[0045] In order to combine the user multi-behavior interaction graph with the project knowledge graph, this embodiment establishes a set of project-entity alignment sets Among them, (v j,e) represents the item v in the multi-behavior interaction graph j It can be aligned with the entity e in the knowledge graph; according to the alignment set A, the user multi-behavior interaction graph and the project knowledge graph are combined to obtain the knowledge interaction graph G = {(h, r, t) | h, t∈E ′ ,r∈R ′}, where the entity set Relation set R ′ =R∪θ.

[0046] S2: Input the knowledge interaction graph into the knowledge interaction graph embedding module to obtain the embedded knowledge interaction graph;

[0047] In order to parameterize the entities and relationships on the knowledge interaction graph into vector representations and retain the overall structure of the knowledge interaction graph, the existing knowledge graph embedding technology is used to embed the knowledge interaction graph. Considering that there are many types of relationships between entities in the knowledge interaction graph, this embodiment uses the TransR (relational projection) model to model entities and relationships. Specifically, the knowledge interaction graph is embedded with entities and relationships to obtain the embedded representation (x h ,x r ,x t ); For each relation r, a relation mapping matrix W is defined r , which projects the head entity h and the tail entity t from the original entity space to a specific relational space, and obtains the projection representation of the corresponding relational space r According to the translation principle

[0048] Then the energy score f(h,r,t) of the triple is expressed as:

[0049]

[0050] in, is the relation mapping matrix of relation r. The relation projects the head entity h and the tail entity t from the d-dimensional space to the k-dimensional space. The lower the score of f(h, r, t), the more likely it is that there is a relation r between h and t in the knowledge interaction graph, and vice versa.

[0051] Loss function L based on embedded knowledge interaction graph KIG The definition is as follows:

[0052]

[0053] Among them, the collection (h,r,t ′ ) is a triple that does not exist in the knowledge interaction graph and is constructed by randomly replacing one entity in the real triple. σ(·) is the sigmoid function.

[0054] Since each type of behavior in the graph is regarded as a relationship in the knowledge graph, this module can capture the coarse-grained commonalities between multiple behaviors and knowledge relationships, and the relationship mapping matrix W r Will be shared in subsequent modules.

[0055] S3: Divide the embedded knowledge interaction graph into multiple knowledge interaction subgraphs G according to the types of user behaviors in the user-project dataset k ;

[0056] In order to effectively explore the complex dependencies between a single behavior and knowledge relationship, the embedded knowledge interaction graph is divided into K subgraphs, where each subgraph contains only the kth type of behavior and knowledge relationship, formally defined as

[0057] S4: Each knowledge interaction subgraph G k Input the behavior knowledge perception graph attention module to obtain the final embedding representation of users and items;

[0058] like Figure 3 As shown, the behavior knowledge perception graph attention module includes multiple layers of behavior knowledge perception graph attention layers connected in series: each layer of behavior knowledge perception graph attention layer includes a behavior knowledge perception attention layer, an information aggregation layer, and a cross-behavior dependency self-attention layer;

[0059] The behavior knowledge perception graph attention module processes each knowledge interaction subgraph including:

[0060] S41, the knowledge interaction subgraph G k As the initial knowledge interaction subgraph Each knowledge interaction subgraph G k Input the first line as the knowledge perception graph attention layer to obtain the knowledge interaction subgraph

[0061] S42, knowledge interaction subgraph Input the second behavior of the knowledge perception graph attention layer to obtain the knowledge interaction subgraph

[0062] S43, the knowledge interaction subgraph output by the attention layer l-1 of the previous behavior knowledge perception graph Input the current behavior knowledge perception graph attention layer l to obtain the knowledge interaction subgraph

[0063] The knowledge interaction subgraph output by the attention layer l of the behavior knowledge perception graph and the attention layer l-1 of the behavior knowledge perception graph Processing includes:

[0064] S431, knowledge interaction subgraph Input the behavior knowledge perception attention layer l to obtain the updated head entity vector of the knowledge interaction subgraph

[0065]

[0066] in, Representing knowledge interaction subgraph The tail entity vector, N h ={(h,r,t)|(h,r,t)∈G k} represents the set of triples related to the head entity h, Represents the knowledge interaction subgraph in the behavior knowledge-aware attention layer l The attention weight of the tail entity t propagated along the relation r to the head entity h.

[0067] The behavior relationship and knowledge relationship have different semantic information for the representation learning of the knowledge interaction subgraph. In order to capture the semantic difference between behavior and knowledge relations, the comprehensive attention weight The calculation is based on the following triple adaptive behavior knowledge perception attention weight formula:

[0068]

[0069] in, is the relation attention weight, α s (h,r,t) is the behavioral knowledge perception attention weight, are the transformation matrices of behavior relationship and knowledge relationship respectively, and tanh is a nonlinear activation function.

[0070] To avoid bias, the softmax function is used to normalize the attention weights of all triplets related to the current head entity h:

[0071]

[0072] in, is the normalized weight of the attention weight of the head entity h in the relation r with the tail entity t, and n represents the number of tail entities j of the head entity h in the relation r.

[0073] 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 relations is represented by W in S2. rOn the other hand, since behavior relations and knowledge relations have different semantic information, for example, behavior represents the interaction between users and items, while knowledge relations only connect different items and describe the attributes of items. In order to capture the subtle semantic differences between the above behavior and knowledge relations, this embodiment considers them separately in the attention mechanism of information propagation. Specifically, using W b and W k To model the difference, if the relationship r is a behavior in the user multi-behavior interaction graph, the user entity h and the item entity t are transformed by the transformation matrix W b Project it into the behavior vector space, otherwise pass it through W k Projected into the vector space of knowledge relations.

[0074] S432: Update the head entity vector of each knowledge interaction subgraph before and the updated head entity vector Input the information aggregation layer to obtain the head entity vector of each knowledge interaction subgraph after aggregation

[0075]

[0076] in, is a trainable parameter matrix, mean is the mean calculation function, ReLU is the activation function, and || represents concatenation.

[0077] S433, aggregate the head entity vector of each knowledge interaction subgraph Input the cross-behavior dependency self-attention layer to obtain the final head entity vector

[0078] Since different knowledge interaction subgraphs contain only one behavior relationship, the same node often presents different feature representations in different subgraphs. The previous part revealed the potential dependency between a single behavior 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 under the influence of knowledge relationships, a cross-behavior dependency self-attention mechanism is designed.

[0079] Specifically, the head entity vector of each knowledge interaction subgraph is aggregated Concatenate (Concat) to get the concatenated vector

[0080] According to the splicing vector Computing attention scores across behavior type dependencies

[0081]

[0082] in There are two behavior-related parameters.

[0083] Score the attention Multiply by the concatenation vector Get the final head entity vector

[0084]

[0085] S44, repeat step S43 until the knowledge interaction subgraph output by the attention layer of the last layer of behavior knowledge perception graph is obtained;

[0086] S45. All knowledge interaction subgraphs Concatenate the embedding vectors of the entities to get the embedding representation of users and items:

[0087]

[0088] in, For user u i The embedding representation of For project v j The embedded representation of .

[0089] S5: Input the final embedding representation of the user and item into the prediction module to obtain the interaction probability between the user and the item;

[0090] Finally, the interaction probability under the target behavior k is predicted by taking the inner product of the user embedding and the item embedding

[0091]

[0092] S6: Calculate the loss function value based on the embedded knowledge interaction graph and the interaction probability between the user and the item, and update the model parameters based on the loss function value. When the pre-set number of training times is reached, a trained user multi-behavior recommendation model is obtained.

[0093] The BPRLoss (Bayesian personalized ranking loss) loss function is defined as follows:

[0094]

[0095] in, represents the user-item dataset, I + Represents user u i Under target behavior k and item v j There is interaction, I - Represents user u i Under target behavior k and item v zThere is no interaction; the final joint loss function L is defined as follows:

[0096]

[0097] Among them, Φ represents the set of parameters of the user multi-behavior recommendation model, and λ is the weight.

[0098] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes 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 in the protection scope of the present invention.

Claims

1. A multi-behavior recommendation method based on knowledge interaction graph, characterized in that: include: Obtain user and project data, build a knowledge interaction graph based on the user and project data, input the knowledge interaction graph into the trained user multi-behavior recommendation model, and obtain the interaction probability between users and projects; The user multi-behavior recommendation model includes: knowledge interaction graph embedding module, behavior knowledge perception graph attention module and prediction module; The training process of the user multi-behavior recommendation model includes: S1: Obtain the user-item dataset, build the user multi-behavior interaction graph and the item knowledge graph based on the user-item dataset, and build the knowledge interaction graph G = {(h, r, t) | h, t∈E based on the user multi-behavior interaction graph and the item knowledge graph ′ ,r∈R ′ }, where E ′ is the entity set, R ′ is a relationship set, and the triple (h, r, t) indicates that there is a relationship r between the head entity h and the tail entity t; 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 G according to the types of user behaviors in the user-project dataset k ; Where k is the index of the user behavior type; S4: Each knowledge interaction subgraph G k Input the behavior knowledge perception graph attention module to obtain the final embedding representation of users and items; S5: Input the final embedding representation of the user and item into the prediction module to obtain the interaction probability between the user and the item; S6: Calculate the loss function value based on the embedded knowledge interaction graph and the interaction probability between the user and the item, and update the model parameters based on the loss function value. When the pre-set number of training times is reached, a trained user multi-behavior recommendation model is obtained.

2. According to the multi-behavior recommendation method based on knowledge interaction graph in claim 1, it is characterized in that: User multi-behavior interaction diagram in, is a set of edges, representing K types of user behaviors. represents the user set, represents a set of items, Indicates user Under behavior type k and project There was interaction. Indicates user Under behavior type k and project There is no interaction, i and j are the indexes of the user and the project respectively; the project knowledge graph G2 = {(h, r, t) | h, t∈E, r∈R}, where E represents the entity set, R represents the relationship set, and the triple (h, r, t) represents that there is a relationship r between the head entity h and the tail entity t; constructing a knowledge interaction graph based on the user multi-behavior interaction graph and the project knowledge graph includes: by placing the project v in the user multi-behavior interaction graph j Align with the entities in the knowledge graph to obtain the knowledge interaction graph G = {(h,r,t)|h,t∈E ′ ,r∈R ′ },in, R ′ =R∪θ.

3. The multi-behavior recommendation method based on knowledge interaction graph according to claim 2 is characterized in that: The knowledge interaction graph embedding module is the TransR model, which includes the relation mapping matrix W for each relation r r ; The TransR model processes the knowledge interaction graph by: embedding the knowledge interaction graph to obtain the embedding vector (x h ,x r ,x t ); According to the relationship mapping matrix W r Embed the vector x h 、x t Projected into the space of the corresponding relationship r, we get the projection vector Among them, the TransR model is a relational projection model.

4. The multi-behavior recommendation method based on knowledge interaction graph according to claim 3 is characterized in that: The behavior knowledge perception graph attention module includes multiple layers of behavior knowledge perception graph attention layers connected in series; The behavior knowledge perception graph attention module is used for each knowledge interaction subgraph G k Processing includes: S41, the knowledge interaction subgraph G k As the initial knowledge interaction subgraph Each knowledge interaction subgraph Input the first line as the knowledge perception graph attention layer to obtain the knowledge interaction subgraph S42, knowledge interaction subgraph Input the second behavior of the knowledge perception graph attention layer to obtain the knowledge interaction subgraph S43, the knowledge interaction subgraph output by the attention layer l-1 of the previous behavior knowledge perception graph Input the current behavior knowledge perception graph attention layer l to obtain the knowledge interaction subgraph S44, repeat step S43 until the knowledge interaction subgraph output by the attention layer of the last layer of behavior knowledge perception graph is obtained; S45. All knowledge interaction subgraphs The embedding vectors of the entities are concatenated to obtain the embedding representation of users and items.

5. The multi-behavior recommendation method based on knowledge interaction graph according to claim 4 is characterized in that: Each behavior knowledge perception graph attention layer includes a behavior knowledge perception attention layer, an information aggregation layer, and a cross-behavior dependency self-attention layer; The knowledge interaction subgraph output by the attention layer l of the behavior knowledge perception graph and the attention layer l-1 of the behavior knowledge perception graph The processing includes: transforming the knowledge interaction subgraph Input the behavior knowledge perception attention layer l to obtain the embedding vector of the head entity after the knowledge interaction subgraph is updated Knowledge Interaction Subgraph The embedding vector of the head entity and the updated embedding vector of the head entity Input the information aggregation layer to obtain the embedding vector of the head entity after the knowledge interaction subgraph is aggregated The embedding vector of the head entity after aggregating the knowledge interaction subgraph Input the cross-behavior dependency self-attention layer to obtain the knowledge interaction subgraph and the embedding vector of its head entity 6. The multi-behavior recommendation method based on knowledge interaction graph according to claim 5, characterized in that: Behavior knowledge-aware attention layer l on knowledge interaction subgraph Processing includes: in, Representing knowledge interaction subgraph The tail entity vector, N h ={(h,r,t)|(h,r,t)∈G k } represents the set of triples related to the head entity h, Represents the knowledge interaction subgraph The attention weight of the tail entity t propagated along the relation r to the head entity h.

7. The multi-behavior recommendation method based on knowledge interaction graph according to claim 6, characterized in that: Attention Weight The calculation process includes: according to the knowledge interaction subgraph The embedding vector of the entity in the relation is used to calculate the attention weight of the relation and behavioral knowledge-aware attention weights Weight and weight Add together to get the comprehensive attention weight For the comprehensive attention weight Normalize and get the attention weight 8. The multi-behavior recommendation method based on knowledge interaction graph according to claim 6, characterized in that: Calculating relation attention weights include: Among them, tanh is the activation function.

9. The multi-behavior recommendation method based on knowledge interaction graph according to claim 6, characterized in that: Calculating behavioral knowledge-aware attention weights include: Among them, W b and W k is the transformation matrix and tanh is the activation function.

10. The multi-behavior recommendation method based on knowledge interaction graph according to claim 5, characterized in that: Head entity vector after cross-behavior dependent self-attention layer aggregates knowledge interaction subgraphs The processing includes: aggregating the head entity vector of each knowledge interaction subgraph Perform splicing to obtain the splicing vector According to the splicing vector Computing attention scores across behavior type dependencies Score the attention Multiply by the concatenation vector Get the final head entity vector

Images