A Recommendation Method, System, Device and Storage Medium Based on Knowledge Graph

Abstract

The present invention provides a recommendation method, system, device and storage medium based on a knowledge graph. The method includes the following steps: S1: Obtain the embedded representation vector of user interests: Use the knowledge graph to obtain implicitly associated items according to the user's browsing records, calculate and aggregate the attention scores of the implicitly associated items to obtain the embedded representation vector of user interests; and obtain the embedded representation vector of candidate items: Use the knowledge graph to obtain candidate items similar to the user's preferences according to the item information, calculate and aggregate the attention scores of the candidate items to obtain the embedded representation vector of the candidate items; S2: Predict the interaction probability between the user and the candidate items according to the embedded representation vector of the user interests and the embedded representation vector of the candidate items, and make recommendations according to the predicted interaction probability. The present invention provides a recommendation method, system, device and storage medium based on a knowledge graph, which solves the problem of poor recommendation effect of existing knowledge graph recommendation methods.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge graphs, and more specifically, to a recommendation method, system, device and storage medium based on knowledge graphs. Background Art

[0002] In an information explosion society, the amount of information people face is growing exponentially. Personalized recommendation technology is a technology that helps users quickly find information they may be interested in. Its purpose is to study how to process massive amounts of heterogeneous data, build the optimal user preference model, and help users effectively alleviate information overload.

[0003] In recent years, with the rise of graph neural networks, the application of graph neural networks in the recommendation field has become a research hotspot. Considering that the relationship between users and items cannot be fully reflected by a two-dimensional structure, modeling data in the form of a graph structure can more truly reveal the relationship between data. In addition, the application of knowledge graphs can fill in the gaps in original data in the form of auxiliary information, to a certain extent solving the data sparsity problem that has plagued researchers for a long time. At the same time, it can also further mine implicit information from the semantic information level of the project, rather than just modeling through explicit data such as historical records, so as to more accurately portray user portraits and project features, and thus make more effective and explainable personalized recommendations.

[0004] However, most existing knowledge graph recommendation methods focus on using the knowledge graph on the project side to mine implicit information between projects, while ignoring the differences between the user side and the project side. They are unable to accurately calculate user preferences and project characteristics based on user preferences, resulting in poor recommendation results. Summary of the invention

[0005] In order to overcome the technical defect of poor recommendation effect of existing knowledge graph recommendation methods, the present invention provides a recommendation method, system, device and storage medium based on knowledge graph.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0007] A recommendation method based on knowledge graph includes the following steps:

[0008] S1: Get the embedding representation vector of user interests:

[0009] Use the knowledge graph to obtain implicitly related items based on the user's browsing history, calculate the attention scores of the implicitly related items and aggregate them to obtain the embedded representation vector of the user's interest;

[0010] And get the embedding representation vector of the candidate item:

[0011] Using a knowledge graph, candidate projects similar to the user's preferences are obtained based on project information, the attention scores of the candidate projects are calculated and aggregated to obtain an embedded representation vector of the candidate projects;

[0012] S2: Predict the interaction probability between the user and the candidate projects according to the embedded representation vector of the user's interests and the embedded representation vector of the candidate projects, and make recommendations based on the predicted interaction probability.

[0013] In the above solution, the attention scores are calculated and aggregated respectively from the user side and the project side according to the propagation link of the knowledge graph, and the embedded representation vectors of the user's interests and the candidate projects are obtained correspondingly. By combining the user side and the project side to predict the interaction probability between the user and the candidate projects and making recommendations based on the predicted interaction probability, the characteristic information of both the user side and the project side is reasonably considered, the recommendation effect is improved, and it has broad promotion space and application value.

[0014] Preferably, the user preference is the aggregated representation of all tail entity projects associated with the projects browsed by the user.

[0015] Preferably, the embedded representation vector of the user's interests is obtained through the following steps:

[0016] Sample a preset number of neighbor nodes at each layer of propagation in the knowledge graph, and obtain the set of the user's domain nodes of different orders through preference propagation:

[0017]

[0018] Suppose the set of projects with which user u has interactions is V u ={v|y u,v =1}, y u,v =1 indicates that there is an interaction between user u and project v, then the embedded representation of the user's interests is:

[0019]

[0020] By aggregating the embedded representation u of the user's interests in the set of the user's domain nodes of all orders Nu and the original representation of the user to obtain the embedded representation vector e u ;

[0021] Among them, represents the set of the user's domain nodes of the first order, represents the set of the user's domain nodes of the second order, represents the set of the user's domain nodes of the kth order, N(u) represents the set of the user's domain nodes of all orders, represents the attention score of the implied associated project, h represents the head entity project, r represents the relationship, t represents the tail entity project, πu r,v = v T e r e h represents the unnormalized attention score of the user side; e r represents the embedding representation vector of the relationship r in the knowledge graph triple, e h represents the embedding representation vector of the head entity item in the knowledge graph triple; e t represents the embedding representation vector of the tail entity item in the knowledge graph triple, represents the neighborhood information of the items with which the user has interacted, and G represents the knowledge graph on the item side.

[0022] Preferably, the aggregation formula for the embedding representation of user interests includes:

[0023] agg sum = sigmoid(W(u + u Nu ) + b)

[0024] agg concat = sigmoid(Wconcat(u, u Nu ) + b)

[0025] agg bi-intercation = LeakyRelu(W1(u + u Nu ) + b1) + LeakyRelu(W2(u + u Nu ) + b2)

[0026] where, agg sum represents the addition aggregator, b, b1, and b2 respectively represent different regularization terms, sigmoid represents the sigmoid function, agg concat represents the concatenation aggregator, Wconcat, W1, and W2 respectively represent different learnable training parameters, agg bi-intercation represents the bilinear aggregator, and LeakyRelu is a ReLU-based activation function used to assign a non-zero slope to all negative values.

[0027] Preferably, the embedding representation vector of the candidate item is obtained through the following steps:

[0028] Calculate the attention score of the candidate item

[0029]

[0030]

[0031] Calculate the set of tail entities associated with item v

[0032]

[0033] Aggregate the tail entity of project v itself and all domain nodes associated with the tail entity of project v to obtain the embedded representation vector of the entity

[0034]

[0035] Project the entity into the user's space to optimize the embedded representation of the entity, and fuse to obtain the vector of the single-layer associated entity

[0036]

[0037] Aggregate the vectors of the associated entities in each layer to obtain the embedded representation vector of the domain nodes in the corresponding layer:

[0038]

[0039] Aggregate the embedded representation vectors of the domain nodes in each layer to obtain the embedded representation vector e of the candidate project v ;

[0040] where, π k represents the unnormalized attention score, W0 and W2 represent different learnable training parameters, concat represents concatenating the vectors, represents the k-th head entity associated with the tail entity of project v, represents the k-th relationship associated with the tail entity of project v, represents the embedded representation vector of the tail entity, b0 represents the regularization term, h k represents the k-th head entity associated with the tail entity, N(t v ) represents the set of tail entities associated with project v, represents the embedded representation vector of the k-th head entity associated with the tail entity of project v, W1 represents the learnable training parameter, sigmoid represents the sigmoid function, u represents the user, b represents the regularization term, represents the final embedded representation vector of project v at the l-th layer, t v represents the tail entity t associated with project v, represents the set of all tail entities associated with project v at the l-th layer.

[0041] In the above solution, in combination with the triple structure characteristics of the knowledge graph, the attention scores of associated items are obtained from entity attention and relationship attention respectively, and the aggregated embedded representation vectors of candidate items are obtained; among them, entity attention models candidate items from the perspective of user interest preferences, effectively simulating the importance among different user preferences; relationship attention models candidate items from the inherent association relationships among items, and can better measure the similarity between different items and candidate items.

[0042] Preferably, the aggregation formula for the embedded representation of candidate items includes:

[0043] agg sum = sigmoid(W(v + v Nv )) + b)

[0044] agg concat = sigmoid(Wconcat(v, v Nv )) + b)

[0045] agg bi-intercation = LeakyRelu(W1(v + v Nv )) + b1) + LeakyRelu(W2(v + v Nv )) + b2)

[0046] where v Nv represents the set of associated items of item v.

[0047] In the above solution, by first finding all users who have interacted with item v, all the interacted items of these users are used as the set of associated items of item v.

[0048] Preferably, the following prediction function is used to predict the interaction probability between the user and the candidate item:

[0049]

[0050] where σ represents the sigmoid function, represents the transpose of the embedded representation vector e u of user interest.

[0051] Based on the above-described recommendation method based on the knowledge graph, the present invention also proposes a recommendation system based on the knowledge graph, including:

[0052] a calculation module for the embedded representation vector of user interest, configured to use the knowledge graph to obtain implicit associated items according to the user browsing record, calculate the attention scores of the implicit associated items and aggregate them to obtain the embedded representation vector of user interest;

[0053] An embedded representation vector calculation module for candidate items, which is used to obtain candidate items similar to user preferences according to item information by using a knowledge graph, calculate and aggregate the attention scores of the candidate items to obtain an embedded representation vector of the candidate items;

[0054] An interaction probability prediction module, which is used to predict the interaction probability between the user and the candidate items according to the embedded representation vector of the user's interest and the embedded representation vector of the candidate items;

[0055] A recommendation module, which is used to make recommendations according to the predicted interaction probability.

[0056] Based on the above-mentioned recommendation method based on a knowledge graph, the present invention also proposes a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the processor executes the above-mentioned recommendation method based on a knowledge graph.

[0057] Based on the above-mentioned recommendation method based on a knowledge graph, the present invention also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the processor executes the above-mentioned recommendation method based on a knowledge graph.

[0058] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0059] The present invention provides a recommendation method, system, device and storage medium based on a knowledge graph, calculates and aggregates attention scores according to the propagation link of the knowledge graph from the user side and the item side respectively, and correspondingly obtains the embedded representation vector of the user's interest and the embedded representation vector of the candidate items, combines the user side and the item side to predict the interaction probability between the user and the candidate items, and makes recommendations according to the predicted interaction probability, reasonably considering the characteristic information of the user side and the item side, improving the recommendation effect, and having broad promotion space and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a flowchart of the implementation steps of the technical solution of the present invention;

[0061] Figure 2 It is a schematic diagram of the overall framework of the present invention;

[0062] Figure 3 It is a schematic diagram of the aggregation process of the embedded representation vector of the candidate items in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] The drawings are only for illustrative purposes and should not be construed as limiting the patent;

[0064] To better illustrate this embodiment, some components in the accompanying drawings are omitted, enlarged or reduced, which does not represent the dimensions of the actual product;

[0065] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the accompanying drawings may be omitted.

[0066] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0067] Embodiment 1

[0068] As Figure 1-2 shown, a recommendation method based on a knowledge graph includes the following steps:

[0069] S1: Obtain the embedded representation vector of the user's interest:

[0070] Use the knowledge graph to obtain implicit associated items according to the user's browsing records, calculate and aggregate the attention scores of the implicit associated items, and obtain the embedded representation vector of the user's interest;

[0071] And obtain the embedded representation vector of the candidate item:

[0072] Use the knowledge graph to obtain candidate items similar to the user's preferences according to the item information, calculate and aggregate the attention scores of the candidate items, and obtain the embedded representation vector of the candidate item;

[0073] S2: Predict the interaction probability between the user and the candidate item according to the embedded representation vector of the user's interest and the embedded representation vector of the candidate item, and make recommendations according to the predicted interaction probability.

[0074] In the specific implementation process, the attention scores are calculated and aggregated respectively from the user side and the item side according to the propagation link of the knowledge graph, and the embedded representation vector of the user's interest and the embedded representation vector of the candidate item are obtained correspondingly. By combining the user side and the item side to predict the interaction probability between the user and the candidate item and making recommendations according to the predicted interaction probability, the characteristic information of the user side and the item side is reasonably considered, the recommendation effect is improved, and it has broad promotion space and application value.

[0075] Embodiment 2

[0076] A recommendation method based on a knowledge graph includes the following steps:

[0077] S1: Obtain the embedded representation vector of the user's interest:

[0078] Use the knowledge graph to obtain implicit associated items according to the user's browsing records, calculate and aggregate the attention scores of the implicit associated items, and obtain the embedded representation vector of the user's interest;

[0079] And obtaining the embedded representation vectors of candidate items:

[0080] Using the knowledge graph, obtain candidate items similar to the user's preferences according to the item information, calculate and aggregate the attention scores of the candidate items to obtain the embedded representation vectors of the candidate items;

[0081] In the specific implementation process, the step of using the knowledge graph to obtain candidate items similar to the user's preferences according to the item information is specifically as follows: Using the idea of Trans, map the head entity item and the tail entity item to the relational space, calculate the similarity degree between the fitting entity vectors through the distance formula. If the calculated distance is less than the preset distance threshold, it means that the head entity item and the tail entity item are similar, and the smaller the distance, the closer they are, that is, the closer they are to the user's preferences.

[0082] S2: Predict the interaction probability between the user and the candidate items according to the embedded representation vectors of the user's interests and the candidate items, and make recommendations according to the predicted interaction probability.

[0083] More specifically, the user preference is the aggregated representation of all tail entity items associated with the user's browsed items.

[0084] In the specific implementation process, the user preference is reflected in the user's browsing history and its neighborhood information. Taking movie recommendation as an example, if the user has watched "Iron Man" which belongs to the science fiction movie series, it can be considered that the user is interested in science fiction movies. Obtain the neighborhood information of the user's viewing history, and generate the embedded representation vector of the user's interest through the vector generation neural network and aggregation. For the item node, the item neighborhood contains the expansion of the item association information. The star of the movie "Iron Man" is Robert Downey Jr., then the information of this actor can be expanded along the link in the knowledge graph, add the user preference in the attention network to obtain the neighborhood information that better meets the user's preferences, and aggregate the neighborhood information to represent the item.

[0085] More specifically, the embedded representation vector of the user's interest is obtained through the following steps:

[0086] In order to reduce the introduction of noise and at the same time reduce the number of domain node sets, sample a preset number of neighbor nodes in each layer of propagation in the knowledge graph, and obtain the domain node sets of different orders of the user through preference propagation:

[0087]

[0088] Suppose the set of items with which user u has interactions is V u ={v|y u,v =1}, y u,v =1 indicates that there is an interaction between user u and item v, then the embedded representation of the user's interest is:

[0089]

[0090] By aggregating the embedded representations u of the user's interests in the set of domain nodes at all orders Nu and the original representation of the user to obtain the embedded representation vector e of the user's interests u ;

[0091] Among them, represents the set of domain nodes at the first order of the user, represents the set of domain nodes at the second order of the user, represents the set of domain nodes at the k-th order of the user, and N(u) represents the set of domain nodes at all orders of the user, represents the attention score of the implicit associated item, h represents the head entity item, r represents the relationship, t represents the tail entity item, π u r,v = v T e r e h represents the unnormalized attention score at the user side; e r represents the embedded representation vector of the relationship r in the knowledge graph triple, e h represents the embedded representation vector of the head entity item in the knowledge graph triple; e t represents the embedded representation vector of the tail entity item in the knowledge graph triple, represents the neighborhood information of the items with which the user has interacted, and G represents the knowledge graph at the item side.

[0092] In the specific implementation process, the triple in the knowledge graph is represented as (h, r, t). According to the user's browsed items, all tail entity items are found in the knowledge graph with the limitation that the number of layers does not exceed 3. Then, h, r, and t are input into the attention network to obtain the attention score.

[0093] More specifically, the aggregation formula for the embedded representation of the user's interests includes:

[0094] agg sum = sigmoid(W(u + u Nu ) + b)

[0095] agg concat = sigmoid(Wconcat(u, u Nu ) + b)

[0096] agg bi-intercation = LeakyRelu(W1(u + u Nu ) + b1) + LeakyRelu(W2(u + u Nu ) + b2)

[0097] Among them, agg sum represents an additive aggregator, b, b1, and b2 respectively represent different regularization terms, sigmoid represents the sigmoid function, and agg concat represents a concatenation aggregator, Wconcat, W1, and W2 respectively represent different learnable training parameters, and agg bi-intercation represents a bilinear aggregator, and LeakyRelu is a ReLU-based activation function used to assign a non-zero slope to all negative values.

[0098] In the specific implementation process, the embedded representations u of the user interests in the set of domain nodes of all orders of the user are aggregated through three different aggregators Nu and the original representation of the user to obtain the embedded representation vector e of the user interests u .

[0099] More specifically, as Figure 3 shown, the embedded representation vector of the candidate item is obtained through the following steps:

[0100] Calculate the attention score of the candidate item

[0101]

[0102]

[0103] Calculate the set of tail entities associated with the item v

[0104]

[0105] Aggregate the tail entity itself of the item v and all domain nodes associated with the tail entity of the item v to obtain the embedded representation vector of the entity

[0106]

[0107] During the process of aggregating entity information, a vector of user preferences is added to promote the propagation along the user-specific direction during the preference propagation in the knowledge graph, so as to capture a set of domain nodes that better conform to the user preferences. Specifically, the embedded representation of the entity is optimized by projecting the entity into the user's space, and the vector of the single-layer associated entity is fused

[0108]

[0109] Aggregate the vectors of the associated entities of each layer to obtain the embedded representation vector of the domain nodes of the corresponding layer:

[0110]

[0111] Through the aggregation of domain nodes at a single layer, the final representation of an entity is determined by itself and the domain nodes, and it is extended from one layer to multiple layers, modeling the potential information of the project in ways of delving deeper longitudinally and expanding horizontally respectively;

[0112] Finally, the embedded representation vectors of the domain nodes at each layer are aggregated to obtain the embedded representation vector e of the candidate project v ;

[0113] where, π k represents the unnormalized attention score, W0 and W2 represent different learnable training parameters, concat represents concatenating vectors, represents the k-th head entity associated with the tail entity of project v, represents the k-th relation associated with the tail entity of project v, represents the embedded representation vector of the tail entity, b0 represents the regularization term, h k represents the k-th head entity associated with the tail entity, N(t v ) represents the set of tail entities associated with project v, represents the embedded representation vector of the k-th head entity associated with the tail entity of project v, W1 represents the learnable training parameter, sigmoid represents the sigmoid function, u represents the user, b represents the regularization term, represents the final embedded representation vector of project v at the l-th layer, t v represents the tail entity t associated with project v, represents the set of all tail entities associated with project v at the l-th layer.

[0114] In the specific implementation process, combined with the triple structure characteristics of the knowledge graph, the attention scores of associated projects are obtained from entity attention and relation attention respectively, and the embedded representation vector of the candidate project is aggregated. Among them, entity attention models the candidate project from the perspective of user interest preferences, effectively simulating the importance among different user preferences. Taking the example of users watching movies, some users will choose corresponding movies mainly based on the actors, and some users will also choose movies mainly based on the movie theme or director factors. Therefore, it is considered that the focus of user preferences will affect the user's choice of entities; specifically, an attention neural network is used to fit the similarity degree between entity vectors, and then the softmax function is used to normalize the similarity calculation results to obtain the attention weight coefficients. Relation attention models the candidate project from the inherent association relationship between projects, and can better measure the similarity degree between different projects and the candidate project.

[0115] More specifically, the aggregation formula for the embedded representation of the candidate project includes:

[0116] agg sum = sigmoid(W(v + v Nv ) + b)

[0117] agg concat = sigmoid(Wconcat(v, v Nv ) + b)

[0118] agg bi-intercation = LeakyRelu(W1(v + v Nv ) + b1) + LeakyRelu(W2(v + v Nv ) + b2)

[0119] where v Nv represents the set of associated items of item v.

[0120] In the specific implementation process, by first finding all users who have interacted with item v, all the interacted items of these users are used as the set of associated items of item v.

[0121] More specifically, the following prediction function is used to predict the interaction probability between the user and the candidate item:

[0122]

[0123] where σ represents the sigmoid function, represents the transpose of the embedded representation vector e u of the user's interest.

[0124] In the specific implementation process, the loss function of the prediction function is:

[0125]

[0126] Through the constraint of the loss function, it continuously converges to achieve the effect of training self-optimization;

[0127] where Γ is the cross-entropy loss function, and p is negative sampling that follows a normal distribution.

[0128] Embodiment 3

[0129] Based on the above-mentioned recommendation method based on the knowledge graph, the present invention also proposes a recommendation system based on the knowledge graph, including:

[0130] A calculation module for the embedded representation vector of the user's interest, which is used to obtain implicit associated items according to the user's browsing records by using the knowledge graph, calculate the attention scores of the implicit associated items and aggregate them to obtain the embedded representation vector of the user's interest;

[0131] The embedding representation vector calculation module for candidate items is used to obtain candidate items similar to the user's preferences according to the item information by using the knowledge graph, calculate and aggregate the attention scores of the candidate items, and obtain the embedding representation vector of the candidate items;

[0132] The interaction probability prediction module is used to predict the interaction probability between the user and the candidate items according to the embedding representation vector of the user's interests and the embedding representation vector of the candidate items;

[0133] The recommendation module is used to make recommendations according to the predicted interaction probability.

[0134] Embodiment 4

[0135] Based on the above-mentioned recommendation method based on the knowledge graph, the present invention also proposes a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the processor executes the recommendation method based on the knowledge graph described in Embodiment 1 or Embodiment 2.

[0136] Embodiment 5

[0137] Based on the above-mentioned recommendation method based on the knowledge graph, the present invention also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the processor executes the recommendation method based on the knowledge graph described in Embodiment 1 or Embodiment 2.

[0138] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made on the basis of the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A recommendation method based on a knowledge graph, characterized in that, Including the following steps: S1: Obtain the embedded representation vector of the user's interest: Use the knowledge graph to obtain implicitly associated items based on the user's browsing history, calculate and aggregate the attention scores of the implicitly associated items to obtain the embedded representation vector of the user's interest, including: Sample a preset number of neighbor nodes at each layer of propagation in the knowledge graph, and obtain the set of neighborhood nodes of the user at different orders through preference propagation: 、 、…、 Suppose the user u The set of items that generate interactions is , indicating that the user u has an interaction with the item v , then the embedded representation of the user's interest is: By aggregating the embedded representations of the user interests in the sets of neighborhood nodes of all orders and the original representation of the user to obtain the embedded representation vector of the user interests ; Among them, represents the set of neighborhood nodes of the user at the 1st order, represents the set of neighborhood nodes of the user at the 2nd order, represents the set of neighborhood nodes of the user at the kth order, represents the set of neighborhood nodes of all orders of the user, represents the attention score of the implicit associated item, h represents the head entity item, r represents the relationship, t represents the tail entity item, represents the unnormalized attention score at the user side; represents the relationship in the knowledge graph triple r 's embedded representation vector, represents the embedded representation vector of the head entity item in the knowledge graph triple; represents the embedded representation vector of the tail entity item in the knowledge graph triple, represents the neighborhood information of the items with which the user has interacted, G represents the knowledge graph at the item side; And obtain the embedded representation vector of the candidate item: Use the knowledge graph to obtain candidate items similar to the user's preference based on the item information, calculate and aggregate the attention scores of the candidate items to obtain the embedded representation vector of the candidate item, including: Calculate the attention score of candidate items : Calculate the project v The set of tail entities associated : Aggregation project v The tail entity itself of v and all neighborhood nodes associated with the tail entity of the project are used to obtain the embedded representation vector of the entity : Project the entity into the user's space to optimize the embedded representation of the entity, and fuse to obtain the vector of a single-layer associated entity : Aggregate the vectors of the associated entities at each layer to obtain the embedded representation vector of the neighborhood nodes at the corresponding layer: Aggregate the embedded representation vectors of the neighborhood nodes in each layer to obtain the embedded representation vector of the candidate item ; Among them, represents the unnormalized attention score, , represent different learnable training parameters, represents concatenating vectors, represents related to item v the k th head entity associated with the tail entity of the item, represents related to item v the k th relation associated with the tail entity of the item, represents the embedding representation vector of the tail entity, represents the regularization term, represents the k th head entity associated with the tail entity, represents related to item v the set of tail entities associated with the item, represents related to item v the k th embedding representation vector of the head entity associated with the tail entity of the item, represents a learnable training parameter, represents the sigmoid function, u represents the user, represents the regularization term, represents item v at the l th layer's final embedding representation vector, represents related to item v the tail entity t , represents related to item v at the l th layer, the set of all associated tail entities; S2: Predict the interaction probability between the user and the candidate item according to the embedded representation vector of the user's interest and the embedded representation vector of the candidate item, and make recommendations according to the predicted interaction probability.

2. The recommendation method based on a knowledge graph according to claim 1, wherein The user preference is the aggregated representation of all tail entity items associated with the user's browsed items.

3. The recommended method based on a knowledge graph according to claim 1, characterized in that The aggregation formula for the embedded representation of the user's interest includes: Among them, represents an additive aggregator, b、 , respectively represent different regular terms, represents the sigmoid function, represents a concatenation aggregator, , , respectively represent different learnable training parameters, represents a bilinear aggregator, is a ReLU-based activation function that assigns a non-zero slope to all negative values.

4. The recommendation method based on a knowledge graph according to claim 1, characterized in that The aggregation formula for the embedded representation of the candidate item includes: Among them, represents the v set of associated items of the item.

5. A recommendation method based on a knowledge graph according to claim 1, wherein Use the following prediction function to predict the interaction probability between the user and the candidate item: Among them, represents the sigmoid function, represents the transposed embedding representation vector of the user interest .

6. A recommendation system based on a knowledge graph, applied to the recommendation method based on a knowledge graph according to any one of claims 1 to 5, characterized in that Including: An embedded representation vector calculation module for the user's interest, which is used to use the knowledge graph to obtain implicitly associated items based on the user's browsing history, calculate and aggregate the attention scores of the implicitly associated items to obtain the embedded representation vector of the user's interest; An embedded representation vector calculation module for the candidate item, which is used to use the knowledge graph to obtain candidate items similar to the user's preference based on the item information, calculate and aggregate the attention scores of the candidate items to obtain the embedded representation vector of the candidate item; An interaction probability prediction module, which is used to predict the interaction probability between the user and the candidate item according to the embedded representation vector of the user's interest and the embedded representation vector of the candidate item; A recommendation module, which is used to make recommendations according to the predicted interaction probability.

7. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 5.

Images