A knowledge graph attention network recommendation method based on graph collaborative filtering

Abstract

This invention provides a knowledge graph attention network recommendation method based on graph collaborative filtering. The method includes: constructing a recommendation model; constructing a collaborative filtering layer for the recommendation model; constructing a knowledge graph attention embedding layer for the recommendation model; aggregating the collaborative filtering layer and the knowledge graph attention embedding layer to complete the recommendation model; inputting users into the recommendation model and outputting the predicted user click-through rate. This invention can utilize the collaborative filtering information of users and items to calculate the attention weights of related entities in the knowledge graph, thereby fully mining the similarity between users and objects, and between objects.

Description

Technical Field

[0001] This invention relates to the field of information recommendation technology, specifically to a knowledge graph attention network recommendation method based on graph collaborative filtering. Background Technology

[0002] The basic assumption of collaborative filtering is that similar users will show similar preferences for items. Since it has fully entered the field of deep learning, it is generally necessary to first learn the embeddings of users and items in the latent space, and then reconstruct the interaction between the two, such as matrix factorization to make inner product, and neural collaborative filtering to simulate higher-order interactions. Since knowledge graphs can effectively solve the sparsity and cold start problems of collaborative filtering, they have been widely studied and applied as side information in recommender systems. Social networks that reflect user relationships are used to enhance user representation, and knowledge graphs that express the relationships between items through attributes are used to enhance item representation. Integrating knowledge graphs into recommendations can bring two benefits: (1) the rich semantic relationships between items in the knowledge graph help to explore the connections between items and improve the representation of items; (2) the knowledge graph connects the user's historical interaction items with the recommended items, which enhances the interpretability of the results. Existing knowledge graph attention networks integrate knowledge graph relationship information and user-item interaction graphs into a graph space, so that collaborative filtering information and knowledge graph information can be integrated, and higher-order relationship information can be discovered through collaborative knowledge graphs.

[0003] Most existing knowledge graph-based recommendation methods focus on how to effectively encode knowledge associations in the knowledge graph, without highlighting potential key collaborative signals in user-item interactions and using the same strategy for both types of information. This leads to the learned entity vectors failing to accurately represent the features of the entities. Summary of the Invention

[0004] The purpose of this invention is to provide a knowledge graph attention network recommendation method based on graph collaborative filtering. This method can utilize the collaborative filtering information of users and items to calculate the attention weights of related entities of users and items in the knowledge graph, thereby fully mining the similarity between users and objects, and between objects.

[0005] A knowledge graph attention network recommendation method based on graph collaborative filtering includes:

[0006] Build a recommendation model;

[0007] The construction of the recommendation model includes: constructing a collaborative filtering layer for the recommendation model;

[0008] Construct a knowledge graph attention embedding layer for the recommendation model;

[0009] The collaborative filtering layer and knowledge graph attention embedding layer of the recommendation model are combined to complete the recommendation model;

[0010] Input users into the recommendation model and output the predicted user click-through rate.

[0011] The collaborative filtering layer used to build the recommendation model includes:

[0012] Construct a user-item bipartite graph to obtain the relevant itemsets for users and items. u and I v ;

[0013] Obtain initial entity embedding vectors for users and items from the relevant project set. and

[0014] Initial entity embedding vectors for users and items and The results are transformed into collaborative filtering vectors u representing users and items. cf and v cf .

[0015] The knowledge graph attention embedding layer used to build the recommendation model includes:

[0016] Construct a user-item bipartite graph to obtain the relevant itemsets for users and items. u and I v ;

[0017] Related itemset for users and items u and I v By aligning items and entities, an initial entity set U is generated and propagated within the knowledge graph. 0 and V 0 ;

[0018] Through user item interactions and knowledge associations in the knowledge graph, collaborative signals are propagated. After multiple recursive steps, the initial entity set yields the representation U of users and items. l and V l ;

[0019] Weighted representation of entities generated based on attention mechanism The knowledge graph attribute vectors u of users and items are obtained using a hierarchical aggregation mechanism. kg and v kg .

[0020] Obtain initial entity embedding vectors for users and items from the relevant project set. and include:

[0021] Assume the set of users is I u ={u1,u2,...,u m} and the set of items I v ={v1,v2,...,v n};

[0022] Based on historical interactions, the user-item interaction matrix Y∈R is obtained. m×n , where y uv =1 indicates that there is an observed interaction between user u and item v; if there is no interaction between the user and the item, y = 1. uv =0;

[0023] Initial embedding vectors for users and items are obtained using the PinSAGE algorithm. and

[0024] Initial entity embedding vectors for users and items and The results are transformed into collaborative filtering vectors u representing users and items. cf and v cf include:

[0025] The collaborative filtering layer aggregates neighborhood information of nodes in the user-item bipartite graph through simplified graph convolution operations to obtain collaborative filtering vector representations of users and items.

[0026] Graph convolutional networks can achieve convergence by skipping message passing in infinite layers. When convergence is achieved, the embedding can be expressed as follows:

[0027]

[0028] Among them, e u Let d be the user feature vector after infinite iterations. u and d v Represents the originality of user and item nodes, following standard GCN design.

[0029] The following convergence state is derived, and the collaborative filtering vector u is obtained. cf v cf :

[0030]

[0031]

[0032] Related itemset for users and items u and I v By aligning items and entities, an initial entity set U is generated and propagated within the knowledge graph. 0 and V 0 include:

[0033] User and item related items set I u and I v By aligning items and entities, an initial entity set U is generated and propagated within the knowledge graph. 0 ={u|u∈G,y uv =1} and V 0 ={v|v∈G,y uv =1};

[0034] Where G represents the set of all entities and relations in the knowledge graph, y uv =1 indicates that user u has interacted with item v; if the user has not interacted with the item, y = 1. uv =0.

[0035] Through user item interactions and knowledge associations in the knowledge graph, collaborative signals are propagated. After multiple recursive steps, the initial entity set yields the representation U of users and items. l and V l include:

[0036] The entity set definition of user u and item v at level l is recursively expressed as:

[0037] U l ={t|(h,r,t)∈G,h∈U l-1}

[0038] V l ={t|(h,r,t)∈G,h∈V l-1}

[0039] Here, the triple (h,r,t) is a triple containing an entity t in the l-th level attribute entity set of user u;

[0040] The l-level recursive triple representation is obtained:

[0041]

[0042]

[0043] Weighted representation of entities generated based on attention mechanism The knowledge graph attribute vectors u of users and items are obtained using a hierarchical aggregation mechanism. kg and v kg include:

[0044] Suppose that the triple (h, r, t) is a triple containing an entity t in the l-level attribute entity set of user u, and define the vector representation of t after adding attention weights as a. i :

[0045] a i =π(h) i ,ri )t i

[0046] Among them, h i It is a head entity embedding, r i It is relational embedding, t i It is a tail entity embedding, π(h) i ,r i The attention weights generated by the head entity and the head-to-tail relationship are controlled by a neural network with an attention mechanism, implementing the π(·) function as follows:

[0047] π(h i ,r i )=σ(W2ReLU(W1ReLU(W0(h i ||r i )+b0)+b1)+b2)

[0048] In this context, ReLU serves as the nonlinear activation function, W and b are the weight matrix and bias to be learned, respectively, with different subscripts indicating that they are parameters of different layers. The softmax function is used to normalize the coefficients of the entire triplet.

[0049]

[0050] in, This is a general representation of multi-level recursive triples. When calculating the triple representation of the l-th level, attention is calculated for each triple (h′, r′, t′) in the l-th level, and the multi-level attribute information is aggregated to obtain the l-th level attribute vector representation of the user / item.

[0051]

[0052]

[0053] in, and A vector representation of the attention weights for users and items;

[0054] Concatenate the l-level attribute vectors of user u and item v to obtain the knowledge graph attribute vectors of user and item:

[0055]

[0056]

[0057] Where || represents the concatenation operation.

[0058] The collaborative filtering layer and knowledge graph attention embedding layer of the aggregation recommendation model are combined to complete the recommendation model, which includes:

[0059] By combining the embeddings of the collaborative filtering layer and the knowledge graph attention layer, we obtain the final vector representations of user u and item v:

[0060] u=σ(W·(u cf ||u kg )+b)

[0061] v=σ(W·(v cf ||v kg )+b)

[0062] The recommendation model uses the dot product of u and v to predict the probability that user u will choose item v:

[0063]

[0064] A knowledge graph attention network recommendation system based on graph collaborative filtering includes:

[0065] The first data processing module is used to build the recommendation model;

[0066] The second data processing module is used to take user input into the recommendation model and output the predicted user click-through rate.

[0067] This invention utilizes a collaborative filtering propagation layer and a knowledge graph attention propagation layer in the knowledge graph-based collaborative filtering knowledge graph recommendation model to ensure the effective fusion of collaborative filtering information and entity-related information in the knowledge graph, thus preserving the accuracy of the recommendation results. Then, in the knowledge graph attention embedding propagation layer, the collaborative filtering information of users and items is used to calculate the attention weights of related entities of users and items in the knowledge graph, thereby fully mining the similarity between users and objects, and between objects. Attached Figure Description

[0068] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 This is a flowchart of the present invention;

[0071] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0072] 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0073] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0074] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0075] Existing knowledge graph-based recommendation methods mostly focus on effectively encoding knowledge associations within the knowledge graph, without highlighting potential key collaborative signals in user-item interactions. Furthermore, using the same strategy for both types of information leads to learned entity vectors that fail to accurately represent entity features. Introducing graph convolutional networks into recommendation algorithms, however, can effectively extract high-order connection information between users and items.

[0076] Therefore, in the knowledge graph attention network recommendation model based on graph collaborative filtering, a collaborative filtering propagation layer and a knowledge graph attention propagation layer are used to ensure the effective fusion of collaborative filtering information and entity-related information in the knowledge graph, so that the accuracy of the recommendation results is not affected. Then, in the knowledge graph attention embedding propagation layer, the collaborative filtering information of users and items is used to calculate the attention weights of the entities associated with users and items in the knowledge graph, so as to fully explore the similarity between users and objects, and between objects.

[0077] Example 1

[0078] A knowledge graph attention network recommendation method based on graph collaborative filtering includes:

[0079] S100, Build a recommendation model;

[0080] The construction of the recommendation model includes: S101, constructing a collaborative filtering layer for the recommendation model;

[0081] S102, Construct the knowledge graph attention embedding layer for the recommendation model;

[0082] S103, which integrates the collaborative filtering layer and knowledge graph attention embedding layer of the recommendation model to complete the recommendation model;

[0083] S200 takes user input as input to the recommendation model and outputs the predicted user click-through rate.

[0084] The collaborative filtering layer used in S101 to build the recommendation model includes:

[0085] S1011, Construct a user-item bipartite graph and obtain the relevant itemsets for users and items. u and I v ;

[0086] S1012, Obtain initial entity embedding vectors for users and items from the relevant project set. and

[0087] S1013, embed the initial entities of users and items into vectors. and The results are transformed into collaborative filtering vectors u representing users and items. cf and v cf .

[0088] In recommendation scenarios, we often have historical user-item interactions (e.g., purchases and clicks). Taking the Movielens dataset as an example, let's assume the set of users is I. u ={u1,u2,...,u m} and the collection of movies I v ={v1,v2,...,v n Based on historical interactions, we can obtain the user-movie interaction matrix Y∈R. m×n , where y uv =1 indicates that there is an observed interaction between user u and movie v; otherwise, y uv =0.

[0089] The collaborative filtering layer aggregates neighborhood information of nodes in the user-item bipartite graph through simplified graph convolution operations to obtain collaborative filtering vector representations of users and movies.

[0090] Graph convolutional networks (GCNNs) have been widely used in recommender systems as an effective collaborative filtering method. However, message propagation slows down the convergence speed of GCNNs during training. After an infinite number of message propagation layers, the final result tends to a fixed value. By skipping this infinite layer of message propagation, the convergence state of the model can be approximated. When convergence is achieved, the embedding can be written as follows:

[0091]

[0092] Among them, e u Let d be the user feature vector after infinite iterations. u and d v Represents the originality of users and movie nodes, following standard GCN design.

[0093] After some simplification, the following convergence state is derived, yielding the collaborative filtering vector u. cf v cf :

[0094]

[0095]

[0096] To facilitate optimization, sigmoid activation and negative log-likelihood were further combined. Although this greatly simplified the model, the current loss is still affected by oversmoothing. To ensure the model's performance, negative sampling was added, resulting in the following loss function:

[0097]

[0098] in, and Let σ represent the set of positive pairs and randomly sampled negative pairs, where σ is the sigmoid function.

[0099] We use BCE (Binary Cross Entropy) as the main loss function:

[0100]

[0101] Both losses are based on the User-Item graph, therefore we can now obtain a basic form of the loss function:

[0102] L = L o +λL c

[0103] Where λ is a hyperparameter that controls the importance weights of the two loss terms.

[0104] The knowledge graph attention embedding layer for building the recommendation model in S102 includes:

[0105] S1021, Construct a user-item bipartite graph and obtain the relevant itemsets for users and items. u and I v ;

[0106] S1022, link the relevant itemset I for users and items u and I v By aligning items and entities, an initial entity set U is generated and propagated within the knowledge graph. 0 and V 0 ;

[0107] S1023, through user item interactions and knowledge associations in the knowledge graph, collaborative signals are propagated, and the initial entity set is recursively processed through multiple layers to obtain the representations U of users and items. l and V l ;

[0108] S1024, Weighted representation of entities generated based on attention mechanism The knowledge graph attribute vectors u of users and items are obtained using a hierarchical aggregation mechanism. kg and v kg .

[0109] Knowledge graph-based recommendation leverages the rich information in knowledge graphs, based on known user-item interaction information, to estimate user preferences for items by explicitly capturing the correlations between products.

[0110] User and film related project set I u and I v By aligning items and entities, an initial entity set U can be generated and propagated within the knowledge graph. 0 ={u|u∈G,y uv =1} and V 0 ={v|v∈G,y uv =1}.

[0111] Where G represents the set of all entities and relations in the knowledge graph; y uv =1 indicates that user u has interacted with movie v; otherwise, y = 1. uv =0.

[0112] S1012 obtains the initial entity embedding vectors of users and items in the relevant project set. and include:

[0113] Assume the set of users is I u ={u1,u2,...,u m} and the set of items I v ={v1,v2,...,v n};

[0114] Based on historical interactions, the user-item interaction matrix Y∈R is obtained. m×n , where y uv =1 indicates that there is an observed interaction between user u and item v; if there is no interaction between the user and the item, y = 1. uv =0;

[0115] Initial embedding vectors for users and items are obtained using the PinSAGE algorithm. and

[0116] S1023 embeds the initial entity embedding vectors of users and items. and The results are transformed into collaborative filtering vectors u representing users and items. cf and v cf include:

[0117] The collaborative filtering layer aggregates neighborhood information of nodes in the user-item bipartite graph through simplified graph convolution operations to obtain collaborative filtering vector representations of users and items.

[0118] Graph convolutional networks can achieve convergence by skipping message passing in infinite layers. When convergence is achieved, the embedding can be expressed as follows:

[0119]

[0120] Among them, e u Let d be the user feature vector after infinite iterations. u and d v Represents the originality of user and item nodes, following standard GCN design.

[0121] The following convergence state is derived, and the collaborative filtering vector u is obtained. cf v cf :

[0122]

[0123]

[0124] S1022 will set up relevant items for users and items. u and I v By aligning items and entities, an initial entity set U is generated and propagated within the knowledge graph. 0 and V 0 include:

[0125] User and item related items set I u and I vBy aligning items and entities, an initial entity set U is generated and propagated within the knowledge graph. 0 ={u|u∈G,y uv =1} and V 0 ={v|v∈G,y uv =1};

[0126] Where G represents the set of all entities and relations in the knowledge graph, y uv =1 indicates that user u has interacted with item v; if the user has not interacted with the item, y = 1. uv =0.

[0127] S1023 propagates collaborative signals through user item interactions and knowledge associations in the knowledge graph, obtaining the representations U of users and items after multi-level recursion of the initial entity set. l and V l include:

[0128] The entity set definition of user u and item v at level l is recursively expressed as:

[0129] U l ={t|(h,r,t)∈G,h∈U l-1}

[0130] V l ={t|(h,r,t)∈G,h∈V l-1}

[0131] Here, the triple (h,r,t) is a triple containing an entity t in the l-th level attribute entity set of user u;

[0132] The l-level recursive triple representation is obtained:

[0133]

[0134]

[0135] Adjacent entities in a knowledge graph are always strongly related. Propagating along links within the knowledge graph as units yields extended entity sets and triplet sets with varying distances from the initial entity set, effectively expanding the latent vector representations of users and items. Therefore, we can obtain the representations of users and items after multiple recursions, where the tail entity t of layer l aggregates information from the head entity h of layer (l-1). The recursive definition of the entity set of user u and movie v at layer l is as follows:

[0136] U l ={t|(h,r,t)∈G,h∈U l-1}

[0137] V l ={t|(h,r,t)∈G,h∈Vl-1}

[0138] Here, the triple (h,r,t) is a triple containing an entity t in the l-th level attribute entity set of user u.

[0139] Similarly, we obtain the l-level recursive triple representation:

[0140]

[0141]

[0142] Users who have interacted with the same item can have similar behavioral preferences for the objects they interact with, and these preferences may contribute to the feature representation of that object. By definition, the initial entity set of item v contains entities directly related to item v. A knowledge graph-based attention embedding method generates different attention weights for tail entities to indicate their different meanings when acquiring different head entities and relationships. Finally, a weighted average is used to obtain the representation of the tail entities (i.e., the user / item representation), followed by multi-hop propagation to generate multi-level representations of users and items.

[0143] S1024 generates weighted representations of entities based on an attention mechanism. The knowledge graph attribute vectors u of users and items are obtained using a hierarchical aggregation mechanism. kg and v kg include:

[0144] Suppose that the triple (h, r, t) is a triple containing an entity t in the l-level attribute entity set of user u, and define the vector representation of t after adding attention weights as a. i :

[0145] a i =π(h) i ,r i )t i

[0146] Among them, h i It is a head entity embedding, r i It is relational embedding, t i It is a tail entity embedding, π(h) i ,r i The attention weights generated by the head entity and the head-to-tail relationship are controlled by a neural network with an attention mechanism, implementing the π(·) function as follows:

[0147] π(h i ,r i )=σ(W2ReLU(W1ReLU(W0(h i ||r i )+b0)+b1)+b2)

[0148] In this context, ReLU serves as the nonlinear activation function, W and b are the weight matrix and bias to be learned, respectively, with different subscripts indicating that they are parameters of different layers. The softmax function is used to normalize the coefficients of the entire triplet.

[0149]

[0150] in, This is a general representation of multi-level recursive triples. When calculating the triple representation of the l-th level, attention is calculated for each triple (h′, r′, t′) in the l-th level, and the multi-level attribute information is aggregated to obtain the l-th level attribute vector representation of the user / item.

[0151]

[0152]

[0153] in, and A vector representation of the attention weights for users and items;

[0154] Concatenate the l-level attribute vectors of user u and item v to obtain the knowledge graph attribute vectors of user and item:

[0155]

[0156]

[0157] Where || represents the concatenation operation.

[0158] The S103 aggregation recommendation model incorporates a collaborative filtering layer and a knowledge graph attention embedding layer, completing the recommendation model as follows:

[0159] By combining the embeddings of the collaborative filtering layer and the knowledge graph attention layer, we obtain the final vector representations of user u and item v:

[0160] u=σ(W·(u cf ||u kg )+b)

[0161] v=σ(W·(f cf ||v kg )+b)

[0162] The recommendation model uses the dot product of u and v to predict the probability that user u will choose item v:

[0163]

[0164] Example 2

[0165] A knowledge graph attention network recommendation system based on graph collaborative filtering includes:

[0166] The first data processing module is used to build the recommendation model;

[0167] The second data processing module is used to take user input into the recommendation model and output the predicted user click-through rate.

[0168] This invention utilizes a collaborative filtering propagation layer and a knowledge graph attention propagation layer in the knowledge graph-based collaborative filtering knowledge graph recommendation model to ensure the effective fusion of collaborative filtering information and entity-related information in the knowledge graph, thus preserving the accuracy of the recommendation results. Then, in the knowledge graph attention embedding propagation layer, the collaborative filtering information of users and items is used to calculate the attention weights of related entities of users and items in the knowledge graph, thereby fully mining the similarity between users and objects, and between objects.

[0169] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A knowledge graph attention network recommendation method based on graph collaborative filtering, characterized in that, include: Build a recommendation model; The construction of the recommendation model includes: constructing a collaborative filtering layer for the recommendation model; Construct a knowledge graph attention embedding layer for the recommendation model; The collaborative filtering layer and knowledge graph attention embedding layer of the aggregation recommendation model are combined to complete the recommendation model; The user is input into the recommendation model, and the predicted user click-through rate is output; the collaborative filtering layer for constructing the recommendation model includes: Construct a user-item bipartite graph to obtain the relevant itemsets for users and items. and ; Obtain initial entity embedding vectors for users and items from the relevant project set. and ; Initial entity embedding vectors for users and items and The collaborative filtering process transforms the data into collaborative filtering vectors for users and items. and ; The knowledge graph attention embedding layer for constructing the recommendation model includes: Construct a user-item bipartite graph to obtain the relevant itemsets for users and items. and ; Sets of related items for users and items and By aligning items and entities, an initial set of entities is generated and propagated within the knowledge graph. and ; Through user interactions and knowledge associations in the knowledge graph, collaborative signals are propagated, and the initial entity set is recursively multi-level to obtain representations of users and items. and ; Weighted representation of entities generated based on attention mechanism The hierarchical aggregation mechanism is used to obtain the knowledge graph attribute vectors of users and items. and .

2. The knowledge graph attention network recommendation method based on graph collaborative filtering according to claim 1, characterized in that, The step involves obtaining initial entity embedding vectors for users and items from relevant project sets. and include: Assume the set of users is and a collection of items ; Based on historical interactions, a user-item interaction matrix is ​​obtained. ,in Indicates user With items There is observed interaction between the user and the item; if there is no interaction between the user and the item, ; Initial embedding vectors for users and items are obtained using the PinSAGE algorithm. and .

3. The knowledge graph attention network recommendation method based on graph collaborative filtering according to claim 1, characterized in that, The initial entity embedding vector of users and items and The collaborative filtering process transforms the data into collaborative filtering vectors for users and items. and include: The collaborative filtering layer aggregates neighborhood information of nodes in the user-item bipartite graph through simplified graph convolution operations to obtain collaborative filtering vector representations of users and items. Graph convolutional networks can achieve convergence by skipping message passing in an infinite number of layers. When convergence is achieved, the embedding can be expressed as follows: ； in, This is the user feature vector after infinite iterations. and Represents the originality of user and item nodes, following the standard GCN design. ; The following convergence state is derived, and the collaborative filtering vector is obtained. 、 : ； 。 4. The knowledge graph attention network recommendation method based on graph collaborative filtering according to claim 1, characterized in that, The related item set of users and items and By aligning items and entities, an initial set of entities is generated and propagated within the knowledge graph. and include: User and item related project set and By aligning items and entities, an initial set of entities is generated and propagated within the knowledge graph. and ; in, It represents the set of all entities and relations in a knowledge graph. Indicates user Already with items Interaction, if the user does not interact with the item, .

5. The knowledge graph attention network recommendation method based on graph collaborative filtering according to claim 1, characterized in that, The process involves propagating collaborative signals through user item interactions and knowledge associations in the knowledge graph, and obtaining representations of users and items after multi-level recursion of the initial entity set. and include: User u and item v l The entity set definition of a layer is recursively expressed as: ； ； Wherein, the triple (h,r,t) is the first digit of user u. l A triple containing a specific entity t in the layer attribute entity set; get l Layered recursive triple representation: ； 。 6. The knowledge graph attention network recommendation method based on graph collaborative filtering according to claim 1, characterized in that, The weighted representation of entities generated based on the attention mechanism The hierarchical aggregation mechanism is used to obtain the knowledge graph attribute vectors of users and items. and include: Suppose that the triple (h, r, t) is a triple containing an entity t in the l-level attribute entity set of user u, and the vector representation of t after adding attention weights is defined as follows: : ； in, It is a head entity embedding. It is relational embedding. It is a tail-end entity embedding, The attention weights generated by the head entity and the head-to-tail relationship are controlled through a neural network with an attention mechanism. The function, formula is as follows: ； In this context, ReLU serves as the nonlinear activation function, W and b are the weight matrix and bias to be learned, respectively, with different subscripts indicating that they are parameters of different layers. The softmax function is used to normalize the coefficients of the entire triplet. ； in, and A vector representation of the attention weights for users and items; Each of user u and item v l By concatenating the layer attribute vectors, we obtain the knowledge graph attribute vectors for users and items. ； ； Where || represents the concatenation operation.

7. The knowledge graph attention network recommendation method based on graph collaborative filtering according to claim 1, characterized in that, The collaborative filtering layer and knowledge graph attention embedding layer of the aggregated recommendation model complete the recommendation model as follows: By combining the embeddings of the collaborative filtering layer and the knowledge graph attention layer, we obtain the final vector representations of user u and item v: ； ； The recommendation model uses the dot product of u and v to predict the probability that user u will choose item v: 。 8. A knowledge graph attention network recommendation system based on graph collaborative filtering, characterized in that, include: The first data processing module is used to build the recommendation model; The construction of the recommendation model includes: constructing a collaborative filtering layer for the recommendation model; Construct a knowledge graph attention embedding layer for the recommendation model; The collaborative filtering layer and knowledge graph attention embedding layer of the aggregation recommendation model are combined to complete the recommendation model; The second data processing module is used to take user input into the recommendation model and output the predicted user click-through rate. The collaborative filtering layer for constructing the recommendation model includes: Construct a user-item bipartite graph to obtain the relevant itemsets for users and items. and ; Obtain initial entity embedding vectors for users and items from the relevant project set. and ; Initial entity embedding vectors for users and items and The collaborative filtering process transforms the data into collaborative filtering vectors for users and items. and The knowledge graph attention embedding layer used to construct the recommendation model includes: Construct a user-item bipartite graph to obtain the relevant itemsets for users and items. and ; Sets of related items for users and items and By aligning items and entities, an initial set of entities is generated and propagated within the knowledge graph. and ; Through user interactions and knowledge associations in the knowledge graph, collaborative signals are propagated, and the initial entity set is recursively multi-level to obtain representations of users and items. and ; Weighted representation of entities generated based on attention mechanism The hierarchical aggregation mechanism is used to obtain the knowledge graph attribute vectors of users and items. and .

Images