A personalized recommendation method fusing knowledge enhancement and noise suppression strategy

Abstract

The present application relates to the technical field of knowledge graph-based recommendation methods, in particular to a personalized recommendation method fusing knowledge enhancement and noise suppression strategies, which solves the technical problems of poor graph noise suppression, insufficient fine-grained interest capture and unbalanced information utilization of existing recommendation models, and comprises the following steps: constructing multiple noise reduction strategies to suppress graph noise interference, adopting multi-dimensional fusion enhancement representation strategies to obtain the final item representation in the user-item side view and the final item representation in the item-entity side view, so as to improve the recommendation performance; introducing a view contrast learning method to mine the implicit association patterns between different views; and using a matching degree function to calculate the similarity between users and items for evaluation, and then generating a TOP-K recommendation list. The present application suppresses noise interference and significantly improves data representation quality by introducing multiple noise reduction mechanisms and graph structure enhancement strategies, and fully mines the implicit association patterns between different views using a contrast learning mechanism.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph-based recommendation methods, and in particular to a personalized recommendation method that integrates knowledge enhancement and noise suppression strategies. Background Technology

[0002] In the era of big data and artificial intelligence, information data, as an indispensable resource for national economic development, is experiencing explosive growth. How to anchor valuable information amidst massive redundancy has become a bottleneck that urgently needs to be overcome in the digital economy era. Personalized recommendation systems, as core tools for information filtering and precise matching, have become key technologies for improving user experience and optimizing resource allocation. Currently, personalized recommendation algorithms are mainly divided into three categories: content-based recommendation algorithms, collaborative filtering-based recommendation algorithms, and hybrid recommendation algorithms. Knowledge graphs, as an effective auxiliary information in hybrid recommendation algorithms, have attracted considerable attention from researchers in recent years. A knowledge graph is a directed heterogeneous graph network composed of triples consisting of head entities, relations, and tail entities, where nodes and edges correspond to entities and relations, respectively. Introducing knowledge graphs containing a large number of entities and relations as additional auxiliary information into recommendation systems, and utilizing their external knowledge to enhance user and item representations, effectively alleviates problems such as algorithm cold start and data sparsity that exist in traditional algorithm application scenarios when new users or items are added.

[0003] Currently, knowledge graph-based recommendation models have shortcomings in both the graph and recommendation aspects. From the graph perspective, firstly, the triple information in the graph contains noise, i.e., inaccurate, redundant, or irrelevant information. Redundant triple data can lead to the incorrect enhancement of the weights of certain relations, thus causing weight dilution. Irrelevant information can distract the algorithm, reducing its efficiency when processing truly relevant information. Secondly, knowledge graph-driven algorithms often ignore user interaction history, or when integrating user behavior data, user modeling and item modeling are highly coupled. This coupling can introduce invalid information, resulting in a large amount of additional noise. From the recommendation perspective, the ability to fully integrate user behavior data, context awareness, and information from the graph into the user-item representation determines the recommendation system's ability to match user needs and interests, and existing methods (such as RippleNet and KGCN) have limited capabilities in fine-grained user modeling. Furthermore, user interaction information is sparse, while knowledge graph information is redundant. The fusion of the two results in the CF (Collaborative Filtering) signal, which directly determines user preferences, being weakened in the final embedded representation, while the KG (Knowledge Graph) information, which serves as auxiliary information, is given more emphasis in the final user representation. Summary of the Invention

[0004] To overcome the technical shortcomings of existing recommendation models, such as poor suppression of graph noise, insufficient capture of fine-grained interests, and unbalanced information utilization, this invention provides a personalized recommendation method that integrates knowledge enhancement and noise suppression strategies.

[0005] This invention provides a personalized recommendation method that integrates knowledge enhancement and noise suppression strategies, the steps of which are as follows:

[0006] Step S1: Construct a multiple denoising strategy to suppress spectral noise interference. Its sub-steps are as follows:

[0007] Step S101: Construct a learnable binary mask matrix Q to perform dynamic noise filtering on the original interaction matrix Y to obtain the denoised user-item graph;

[0008] Step S102: Through the weighting function To assess the importance of all triples in the knowledge graph, the bottom-k method is used to select the k triples with the lowest value to form a noise set. And mask the noise set on the knowledge graph. After recoding, an enhanced project-knowledge graph was obtained;

[0009] Step S103: Connect the denoised user-item graph with the enhanced item-knowledge graph to form a user-item-entity graph. Then, convert the user-item-entity graph into a low-dimensional continuous vector through knowledge representation learning.

[0010] Step S2: Employ a multi-dimensional fusion enhancement representation strategy to obtain the final project representation in the user-project side view. Final project representation in the project-entity side view ;

[0011] Step S3: Introduce the view comparison learning method to uncover implicit relationship patterns between different views. Its sub-steps are as follows:

[0012] Step S301: Represent the final project in the Project-Entity Side View. Final project representation in the user-project side view Perform cross-view comparative learning;

[0013] Step S302: Construct positive and negative sample pairs based on cross-view node correspondence: where the representation of the same node in different views forms a positive sample pair, and the remaining nodes are used as negative samples;

[0014] Step S303: Calculate the loss function :

[0015] ,

[0016] Step S4: Use a matching degree function to calculate the similarity between users and items for evaluation, and then generate a TOP-K recommendation list.

[0017] Preferably, in step S101, an L1 norm regularization constraint is first applied to the learnable binary mask matrix Q. This suppresses excessive generation of noise edges by controlling the number of non-zero elements in the learnable binary mask matrix Q. The calculation formula is as follows:

[0018] ,

[0019] ,

[0020] In the formula, This represents the user's project interaction index, where [·] represents 0 or 1; These are the corresponding network layer learning parameters used to calculate interaction importance. i,j ~Bern( i,j ), i,j It follows a Bernoulli distribution;

[0021] Then, the Hadamard product operation is used to multiply the learnable binary mask matrix Q, which has been constrained by L1 norm regularization, element by element with the original interaction matrix Y to generate the denoised reconstructed interaction matrix Y'=Q⊙Y, which is the denoised user-item graph. The learnable binary mask matrix Q is used to dynamically determine whether the edges between nodes belong to noise interference.

[0022] Preferably, in step 102, the similarity score of the edges from the same origin is calculated. Softmax normalization is performed, and topological feature enhancement is achieved by incorporating the degree information of the central node, ultimately generating triplet saliency weights with global consistency. The similarity score Significance weights of triples The calculation formula is:

[0023] ,

[0024] ,

[0025] In the formula, , and These are the embedded representations of the head entity, relation, and tail entity, respectively. It is a trainable attention parameter matrix. It is the hidden layer dimension. This indicates the number of connections between the head entity and other entities. A triplet where the leading entity is the same, but the relation and trailing entity are different;

[0026] According to the weighting function The values ​​of all triples in the knowledge graph are derived, and the bottom-k method is used to select the k triples with the lowest values ​​to form a noise set. , The calculation formula is:

[0027] .

[0028] Preferably, the sub-step of step 2 is as follows:

[0029] Step S201: Through learnable attention weights Create historical interactive projects and candidate projects The semantic relevance between them, among which The calculation formula is:

[0030] ,

[0031] In the formula, f( () is a correlation evaluation function based on bilinear transformation;

[0032] Attention-weighted aggregation operations are used to learn representations of users' historical behavioral interactions, where the weight distribution is dynamically adjusted through a gating mechanism, ultimately generating shallow interest levels with goal-adaptive interaction hierarchies. , The calculation formula is:

[0033] ;

[0034] Step S202: Implement neighborhood information fusion based on a two-order lightweight graph convolutional architecture. First, neighborhood aggregation of node features is achieved through layer-by-layer propagation rules:

[0035] ,

[0036] In the formula, This represents the interaction interest representation of the l-th layer. , Represents neighbors connected to users and project nodes;

[0037] Employing a cross-layer representation fusion strategy to generate deep interest at the interaction level , The calculation formula is:

[0038] ;

[0039] Step S203: For candidate projects Recommendation prediction is achieved through a relation-aware inner product operation mechanism. Quantify triples under specific relational paths Middle head entity With candidate projects The semantic association strength generates a relation-dimensional sensitive weighting factor. , The calculation formula is:

[0040] ,

[0041] The relevant entity features are obtained by weighted summation of the relevant entities. , The calculation formula is:

[0042] ;

[0043] Step S204: After jointly encoding the project attribute vector and the original representation, input the input into the graph convolutional network to achieve feature propagation and iterative optimization. After processing through multiple layers of stacked network, the feature representation of entity attributes is generated. ;

[0044] Step S205: Model the first-order neighborhood relationship between users and items using a graph convolutional network, and extract high-order interaction features between users and items by stacking multiple layers of graph convolutional operations. The interaction features propagated through layer l are accumulated to obtain the feature representation of entity relationships. ;

[0045] Step S206, End-user representation in the user-project side view and final project statement They are respectively:

[0046] ,

[0047] ,

[0048] Final project representation in the project-entity side view for:

[0049] .

[0050] Preferably, in step S204, in order to eliminate semantic coupling or semantic pollution caused by relationship intersections in the knowledge graph, heterogeneous relationships are modeled and mapped into independent attribute subspaces, the calculation formula of which is:

[0051] ,

[0052] ,

[0053] In the formula, It is entity i in the corresponding attribute space r m Embedded within;

[0054] The attribute information of entities is fused through cross-relational vector concatenation operations, and the calculation formula is as follows:

[0055] ,

[0056] A primitive representation based on first-order neighborhood feature integration is constructed, and then extended to higher-order neighborhood modeling: a graph convolutional architecture is used to implement a neighborhood information propagation mechanism, and hierarchical convergence of contextual information is achieved through multi-hop connections. In the l-th layer of the network, the item entity representation is updated according to a recursive formula, the calculation formula of which is:

[0057] ,

[0058] Therefore, the feature representation of entity attributes for:

[0059] .

[0060] Preferably, in step S205, in the user-project-entity diagram, the project entity is used to connect users and entities, and entities associated with the project entity are considered as attribute entities. If we represent the neighborhood information of item t, then item t is represented as:

[0061] ,

[0062] This represents the embedded representation of a project after aggregating its neighborhood information. It is an aggregate function;

[0063] Since the same entity may have multidimensional feature representations, and its attribute expressions exhibit dynamic semantic differences as the relational context changes, the relational context is modeled as follows:

[0064] ,

[0065] After modeling the first-order neighborhood relationship between users and items using a graph convolutional network, high-order interaction features between users and items are effectively extracted by stacking multiple layers of graph convolutional operations. The calculation formula is as follows:

[0066] ,

[0067] The interaction features after propagation through layer l are accumulated to obtain the final feature representation of entity relationships. :

[0068] .

[0069] Preferably, in step S4,

[0070] After optimizing the representation by comparing the information from the learning layer, the project representation is obtained. , ),Will and Summing yields the final project representation ,

[0071] To the end user and final project statement A matching operation is performed to obtain a matching score, which is used for recommendation prediction. The formula for calculating the matching operation is as follows:

[0072] .

[0073] The technical solution provided by this invention has the following technical effects compared with the prior art:

[0074] This invention addresses the problems of poor graph noise suppression and insufficient fine-grained interest capture by proposing a personalized recommendation method that combines multiple denoising strategies and multi-dimensional enhancement of user-item representations. It establishes a joint denoising and enhancement strategy by constructing three dimensions: an input-side denoising layer, an internal model denoising layer, and a contrastive learning denoising layer. Simultaneously, it utilizes graph convolutional networks and weighted layers to enhance the final representations of users and items from the perspectives of user-side interactive items and related entities, and item-side entity attributes and relationships, aiming to improve recommendation performance. To address the information imbalance problem in existing recommendation models, this invention performs comparative learning between the user-item view and the item-entity view. By constructing positive and negative sample pairs, it adaptively adjusts the weights of different information sources, mitigating the information imbalance problem. Attached Figure Description

[0075] 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.

[0076] 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.

[0077] Figure 1 This is a flowchart of a personalized recommendation method that integrates knowledge enhancement and noise reduction strategies according to a certain embodiment of the present invention;

[0078] Figure 2This is a model framework diagram of a personalized recommendation method that integrates knowledge enhancement and noise reduction strategies according to a certain embodiment of the present invention;

[0079] Figure 3 This is a schematic diagram of user item interaction noise reduction according to a certain embodiment of the present invention;

[0080] Figure 4 This is a schematic diagram of knowledge graph noise reduction and enhancement in a certain embodiment of the present invention;

[0081] Figure 5 This is a schematic diagram of the fusion and embedding portion in a certain embodiment of the present invention. Detailed Implementation

[0082] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0083] In this description, it should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. It should also be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "joint" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0084] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0085] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0086] In one embodiment, such as Figure 1 As shown, a personalized recommendation method integrating knowledge enhancement and noise suppression strategies is disclosed, with the following steps:

[0087] Step S1: Construct a multi-layered noise reduction strategy consisting of an input noise reduction layer, an internal noise suppression layer, and a contrastive noise compression layer to suppress spectral noise interference. The sub-steps are as follows:

[0088] Step S101: Construct a learnable binary mask matrix Q to dynamically filter noise from the original interaction matrix Y, obtaining a denoised user-item graph. To ensure the model's sparse representation capability, an L1 norm regularization constraint is first applied to the learnable binary mask matrix Q. This is done by controlling the number of non-zero elements in the learnable binary mask matrix Q to suppress the excessive generation of noise edges. The calculation formula is as follows:

[0089] ,

[0090] ,

[0091] In the formula, This represents the user's project interaction index, where [·] represents 0 or 1; These are the corresponding network layer learning parameters used to calculate interaction importance. i,j ~Bern( i,j ), i,j It follows a Bernoulli distribution;

[0092] Then, the Hadamard product operation is used to multiply the learnable binary mask matrix Q, which has been constrained by L1 norm regularization, element by element with the original interaction matrix Y to generate the denoised reconstructed interaction matrix Y'=Q⊙Y, which is the denoised user-item graph. The learnable binary mask matrix Q is used to dynamically determine whether the connection between nodes belongs to noise interference.

[0093] Step S102: Through the weighting function To assess the importance of all triples in the knowledge graph, the bottom-k method is used to select the k triples with the lowest value to form a noise set. And mask the noise set on the knowledge graph. After re-encoding, an enhanced item-knowledge graph is obtained; similarity scores are given for edges with the same origin. Softmax normalization is performed, and topological feature enhancement is achieved by incorporating the degree information of the central node, ultimately generating triplet saliency weights with global consistency. The similarity score Significance weights of triples The calculation formula is:

[0094] ,

[0095] ,

[0096] In the formula, , and These are the embedded representations of the head entity, relation, and tail entity, respectively. It is a trainable attention parameter matrix. It is the hidden layer dimension. This indicates the number of connections between the head entity and other entities. A triplet where the leading entity is the same, but the relation and trailing entity are different;

[0097] According to the weighting function The values ​​of all triples in the knowledge graph are derived, and the bottom-k method is used to select the k triples with the lowest values ​​to form a noise set. , The calculation formula is:

[0098] ;

[0099] Step S103: Since projects correspond to entities in the knowledge graph, the denoised user-project graph is connected to the enhanced project-knowledge graph to form a user-project-entity graph. Through this connection operation, the user-project-entity graph can simultaneously capture user interaction information and project knowledge information, providing a foundation for subsequent multi-dimensional fusion. Knowledge representation learning is then used to convert the user-project-entity graph into a low-dimensional continuous vector for further processing by subsequent modules.

[0100] Step S2: Employ a multi-dimensional fusion enhancement representation strategy to obtain the final project representation in the user-project side view. Final project representation in the project-entity side view In order to improve recommendation performance; its sub-steps are as follows:

[0101] Step S201: Through learnable attention weights Create historical interactive projects and candidate projects The semantic relevance between them, among which The calculation formula is:

[0102] ,

[0103] In the formula, f( () is a correlation evaluation function based on bilinear transformation;

[0104] Attention-weighted aggregation operations are used to learn representations of users' historical behavioral interactions, where the weight distribution is dynamically adjusted through a gating mechanism, ultimately generating shallow interest levels with goal-adaptive interaction hierarchies. , The calculation formula is:

[0105] ;

[0106] Step S202: Implement neighborhood information fusion based on a two-order lightweight graph convolutional architecture. First, neighborhood aggregation of node features is achieved through layer-by-layer propagation rules:

[0107] ,

[0108] In the formula, This represents the interaction interest representation of the l-th layer. , Represents neighbors connected to users and project nodes;

[0109] Employing a cross-layer representation fusion strategy to generate deep interest at the interaction level , The calculation formula is:

[0110] ;

[0111] Step S203: For candidate projects Recommendation prediction is achieved through a relation-aware inner product operation mechanism. Quantify triples under specific relational paths Middle head entity With candidate projects The semantic association strength generates a relation-dimensional sensitive weighting factor. , The calculation formula is:

[0112] ,

[0113] The relevant entity features are obtained by weighted summation of the relevant entities. , The calculation formula is:

[0114] ;

[0115] Step S204: After jointly encoding the project attribute vector and the original representation, input the input into the graph convolutional network to achieve feature propagation and iterative optimization. After processing through multiple layers of stacked network, the feature representation of entity attributes is generated. Specifically, in a knowledge graph, the relationships between an entity and its neighboring nodes are often different. To eliminate semantic coupling or semantic pollution caused by overlapping relationships, heterogeneous relationships are modeled and mapped as independent attribute subspaces, the calculation formula of which is:

[0116] ,

[0117] ,

[0118] In the formula, It is entity i in the corresponding attribute space rm Embedded within;

[0119] The attribute information of entities is fused through cross-relational vector concatenation operations, and the calculation formula is as follows:

[0120] ,

[0121] A primitive representation based on first-order neighborhood feature integration is constructed, and then extended to higher-order neighborhood modeling: a graph convolutional architecture is used to implement a neighborhood information propagation mechanism, and hierarchical convergence of contextual information is achieved through multi-hop connections. In the l-th layer of the network, the item entity representation is updated according to a recursive formula, the calculation formula of which is:

[0122] ,

[0123] Therefore, the feature representation of entity attributes for:

[0124] ;

[0125] Step S205: Model the first-order neighborhood relationship between users and items using a graph convolutional network, and extract high-order interaction features between users and items by stacking multiple layers of graph convolutional operations. The interaction features propagated through layer l are accumulated to obtain the feature representation of entity relationships. Specifically, in the user-item-entity graph, the item entity constitutes one of the core components of a knowledge graph-based recommendation system, connecting users and entities. In this method, entities associated with the item entity are treated as attribute entities. If we represent the neighborhood information of item t, then item t is represented as:

[0126] ,

[0127] This represents the embedded representation of a project after aggregating its neighborhood information. It is an aggregate function;

[0128] Since the same entity may have multidimensional feature representations, and its attribute expressions exhibit dynamic semantic differences as the relational context changes, the relational context is modeled as follows:

[0129] ,

[0130] After modeling the first-order neighborhood relationship between users and items using a graph convolutional network, high-order interaction features between users and items are effectively extracted by stacking multiple layers of graph convolutional operations. The calculation formula is as follows:

[0131] ,

[0132] The interaction features after propagation through layer l are accumulated to obtain the final feature representation of entity relationships. :

[0133] ;

[0134] Step S206, End-user representation in the user-project side view and final project statement They are respectively:

[0135] ,

[0136] ,

[0137] Final project representation in the project-entity side view for:

[0138] ;

[0139] Step S3: Introduce a view contrast learning method to fully explore the implicit correlation patterns between different views. This method uses mutual supervision and complementarity to alleviate the information imbalance problem and suppress view noise. Its sub-steps are as follows:

[0140] Step S301: Represent the final project in the Project-Entity Side View. Final project representation in the user-project side view Perform cross-view comparative learning;

[0141] Step S302: Construct positive and negative sample pairs based on cross-view node correspondence: where the representation of the same node in different views forms a positive sample pair, and the remaining nodes are used as negative samples;

[0142] Step S303: Calculate the loss function :

[0143] ,

[0144] Step S4: The similarity between users and items is calculated using a matching degree function for evaluation, resulting in a matching score used for recommendation prediction, thereby generating a TOP-K recommendation list; specifically, after optimizing the representation through the contrastive learning layer, the item representation is obtained ( , ),Will and Summing yields the final project representation ,

[0145] To the end user and final project statement A matching operation is performed to obtain a matching score, which is used for recommendation prediction. The formula for calculating the matching operation is as follows:

[0146] .

[0147] A specific embodiment of the present invention also discloses a terminal device, which includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of a personalized recommendation method that integrates knowledge enhancement and noise reduction strategies as described in the present invention.

[0148] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. A personalized recommendation method integrating knowledge enhancement and noise suppression strategies, characterized in that, The steps are as follows: Step S1: Construct a multiple denoising strategy to suppress spectral noise interference. Its sub-steps are as follows: Step S101: Construct a learnable binary mask matrix Q to perform dynamic noise filtering on the original interaction matrix Y to obtain the denoised user-item graph; Step S102: Through the weighting function To assess the importance of all triples in the knowledge graph, the bottom-k method is used to select the k triples with the lowest value to form a noise set. And mask the noise set on the knowledge graph. After recoding, an enhanced project-knowledge graph was obtained; Step S103: Connect the denoised user-item graph with the enhanced item-knowledge graph to form a user-item-entity graph. Then, convert the user-item-entity graph into a low-dimensional continuous vector through knowledge representation learning. Step S2: Employ a multi-dimensional fusion enhancement representation strategy to obtain the final project representation in the user-project side view. Final project representation in the project-entity side view ; Step S3: Introduce the view comparison learning method to uncover implicit relationship patterns between different views. Its sub-steps are as follows: Step S301: Represent the final project in the Project-Entity Side View. Final project representation in the user-project side view Perform cross-view comparative learning; Step S302: Construct positive and negative sample pairs based on cross-view node correspondence: where the representation of the same node in different views forms a positive sample pair, and the remaining nodes are used as negative samples; Step S303: Calculate the loss function : ， Step S4: Use a matching degree function to calculate the similarity between users and items for evaluation, and then generate a TOP-K recommendation list.

2. The personalized recommendation method integrating knowledge enhancement and noise suppression strategies according to claim 1, characterized in that, In step S101, an L1 norm regularization constraint is first applied to the learnable binary mask matrix Q. This suppresses excessive generation of noise edges by controlling the number of non-zero elements in the learnable binary mask matrix Q. The calculation formula is as follows: ， ， In the formula, This represents the user's project interaction index, where [·] represents 0 or 1; These are the corresponding network layer learning parameters used to calculate interaction importance. i,j ~Bern( i,j ), i,j It follows a Bernoulli distribution; Then, the Hadamard product operation is used to multiply the learnable binary mask matrix Q, which has been constrained by L1 norm regularization, element by element with the original interaction matrix Y to generate the denoised reconstructed interaction matrix Y'=Q⊙Y, which is the denoised user-item graph. The learnable binary mask matrix Q is used to dynamically determine whether the edges between nodes belong to noise interference.

3. The personalized recommendation method integrating knowledge enhancement and noise suppression strategies according to claim 2, characterized in that, In step 102, the similarity score of the edges from the same origin is calculated. Softmax normalization is performed, and topological feature enhancement is achieved by incorporating the degree information of the central node, ultimately generating triplet saliency weights with global consistency. The similarity score Significance weights of triples The calculation formula is: ， ， In the formula, , and These are the embedded representations of the head entity, relation, and tail entity, respectively. It is a trainable attention parameter matrix. It is the hidden layer dimension. This indicates the number of connections between the head entity and other entities. A triplet where the leading entity is the same, but the relation and trailing entity are different; According to the weighting function The values ​​of all triples in the knowledge graph are derived, and the bottom-k method is used to select the k triples with the lowest values ​​to form a noise set. , The calculation formula is: 。 4. The personalized recommendation method integrating knowledge enhancement and noise suppression strategies according to claim 3, characterized in that, The sub-steps of step 2 are as follows: Step S201: Through learnable attention weights Create historical interactive projects and candidate projects The semantic relevance between them, among which The calculation formula is: ， In the formula, f( () is a correlation evaluation function based on bilinear transformation; Attention-weighted aggregation operations are used to learn representations of users' historical behavioral interactions, where the weight distribution is dynamically adjusted through a gating mechanism, ultimately generating shallow interest levels with goal-adaptive interaction hierarchies. , The calculation formula is: ； Step S202: Implement neighborhood information fusion based on a two-order lightweight graph convolutional architecture. First, neighborhood aggregation of node features is achieved through layer-by-layer propagation rules: ， In the formula, This represents the interaction interest representation of the l-th layer. , Represents neighbors connected to users and project nodes; Employing a cross-layer representation fusion strategy to generate deep interest at the interaction level , The calculation formula is: ； Step S203: For candidate projects Recommendation prediction is achieved through a relation-aware inner product operation mechanism. Quantify triples under specific relational paths Middle head entity With candidate projects The semantic association strength generates a relation-dimensional sensitive weighting factor. , The calculation formula is: ， The relevant entity features are obtained by weighted summation of the relevant entities. , The calculation formula is: ； Step S204: After jointly encoding the project attribute vector and the original representation, input the input into the graph convolutional network to achieve feature propagation and iterative optimization. After processing through multiple layers of stacked network, the feature representation of entity attributes is generated. ; Step S205: Model the first-order neighborhood relationship between users and items using a graph convolutional network, and extract high-order interaction features between users and items by stacking multiple layers of graph convolutional operations. The interaction features propagated through layer l are accumulated to obtain the feature representation of entity relationships. ; Step S206, End-user representation in the user-project side view and final project statement They are respectively: ， ， Final project representation in the project-entity side view for: 。 5. The personalized recommendation method integrating knowledge enhancement and noise suppression strategies according to claim 4, characterized in that, In step S204, to eliminate semantic coupling or semantic pollution caused by overlapping relationships in the knowledge graph, heterogeneous relationships are modeled and mapped into independent attribute subspaces, the calculation formula of which is as follows: ， ， In the formula, It is entity i in the corresponding attribute space r m Embedded within; The attribute information of entities is fused through cross-relational vector concatenation operations, and the calculation formula is as follows: ， A primitive representation based on first-order neighborhood feature integration is constructed, and then extended to higher-order neighborhood modeling: a graph convolutional architecture is used to implement a neighborhood information propagation mechanism, and hierarchical convergence of contextual information is achieved through multi-hop connections. In the l-th layer of the network, the item entity representation is updated according to a recursive formula, the calculation formula of which is: ， Therefore, the feature representation of entity attributes for: 。 6. The personalized recommendation method integrating knowledge enhancement and noise suppression strategies according to claim 5, characterized in that, In step S205, in the user-project-entity diagram, the project entity is used to connect users and entities. Entities associated with the project entity are considered attribute entities. If we represent the neighborhood information of item t, then item t is represented as: ， This represents the embedded representation of a project after aggregating its neighborhood information. It is an aggregate function; Since the same entity may have multidimensional feature representations, and its attribute expressions exhibit dynamic semantic differences as the relational context changes, the relational context is modeled as follows: ， After modeling the first-order neighborhood relationship between users and items using a graph convolutional network, high-order interaction features between users and items are effectively extracted by stacking multiple layers of graph convolutional operations. The calculation formula is as follows: ， The interaction features after propagation through layer l are accumulated to obtain the final feature representation of entity relationships. : 。 7. The personalized recommendation method integrating knowledge enhancement and noise suppression strategies according to claim 6, characterized in that, In step S4, After optimizing the representation by comparing the information from the learning layer, the project representation is obtained. , ),Will and Summing yields the final project representation , To the end user and final project statement A matching operation is performed to obtain a matching score, which is used for recommendation prediction. The formula for calculating the matching operation is as follows: 。

Images