Gating aggregation mechanism-based knowledge graph and interaction graph fusion recommendation method and system
By constructing an item knowledge graph and a user-item interaction graph, and combining graph convolutional networks and gated aggregation units, the problems of cold start and data sparsity in recommendation systems are solved, achieving adaptive information fusion and efficient recommendation.
Patent Information
- Application Number
- CN202510723814.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-01
- Publication Date
- 2025-10-31
AI Technical Summary
Existing recommendation systems suffer from problems such as cold start, data sparsity, and insufficient model interpretability. In particular, existing methods cannot effectively integrate knowledge graphs and interaction graphs, resulting in insufficient recommendation accuracy and diversity, and a lack of dynamic adjustment capabilities.
A gated aggregation mechanism is adopted to construct an item knowledge graph and perform entity aggregation. Then, a graph convolutional network is used to aggregate features of the user-item interaction graph. A gated aggregation unit is designed to dynamically adjust the weights of different representation vectors to achieve adaptive information fusion.
It significantly improves the accuracy and robustness of the recommendation system, enhances the model's expressive power and generalization performance, increases the transparency and user trust of the recommendation system, and solves the problems of data sparsity and cold start.
Smart Images
Figure CN120873271A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to, but is not limited to, the field of recommendation systems, and particularly relates to a recommendation method and system based on the fusion of knowledge graphs and interaction graphs using a gating aggregation mechanism. Background Technology
[0002] Recommender systems have become a core component of digital life, not only solving the problem of information overload but also enhancing user experience and business value. However, traditional recommender systems still suffer from numerous problems, such as cold start, data sparsity, low model interpretability, and insufficient dynamism.
[0003] A knowledge graph is a structured semantic knowledge base that represents entities, attributes, and their relationships in the real world using a graph structure. Introducing knowledge graphs into recommender systems can, to some extent, address the cold start and data sparsity problems inherent in traditional recommender systems by leveraging the rich entity relationships within the graph. Furthermore, knowledge graphs can reveal the underlying logical paths of the recommender system (such as shared interests and similar attributes), further enhancing the transparency and interpretability of the system and making the recommendation results easier to understand and accept.
[0004] The closest existing technology is recommendation methods based on graph attention, such as KGAT (Knowledge Graph Attention Network). These methods first embed knowledge graph nodes, then apply graph attention mechanisms to the user-item interaction graph, fusing semantic information from the knowledge graph with structural information from the interaction graph. Finally, they use attention weights to weighted sums of the multi-source representations to predict user preferences. However, because they treat knowledge graph embedding and interaction graph aggregation as two relatively independent modules, and rely solely on neighbor attention to allocate global parameters, they cannot dynamically adjust the contribution of the two representation vectors for different item or user scenarios. This results in insufficient expressive power when fusing multi-source information and makes it difficult to fully exploit the complementary characteristics between the knowledge graph and interactive behavior.
[0005] Furthermore, existing attention fusion mechanisms lack fine-grained gating control. Noise or sparse regions in the knowledge graph or interaction graph may be introduced into the final representation to the same extent, further exacerbating over-smoothing and cold-start problems, and reducing the accuracy and diversity of recommendations. More importantly, these methods fail to assign learnable, independent weight channels to different types of representation vectors in their model design. They lack an aggregation unit that can both accommodate heterogeneous information and dynamically adjust according to actual interactions, thus limiting the model's adaptability and interpretability in complex recommendation scenarios. Summary of the Invention
[0006] To address the problems existing in the prior art, this invention provides a recommendation method and system based on the fusion of knowledge graphs and interaction graphs using a gating aggregation mechanism.
[0007] This invention includes the following steps:
[0008] Step 1: Construct an item knowledge graph. Use the distance-based embedding model TransE to vectorize the entities and relations in the knowledge graph and represent them in the same vector space, and then perform entity aggregation.
[0009] Step 2: Aggregate entities from the knowledge graph and use a graph convolutional network (GCN) to aggregate the user-item interaction graph;
[0010] Step 3: Design and implement a gated aggregation unit to receive the two types of item representation vectors from Step 1 and Step 2, and calculate the corresponding weight value for each type of representation vector;
[0011] Step 4: Use the final item representation vector to predict user preferences and make content recommendations accordingly;
[0012] Furthermore, step one specifically includes:
[0013] Step 1: Construct a knowledge graph based on the data and visualize it;
[0014] Step 2: Define the scoring function for TransE;
[0015] f(h,r,t)=||h+rt|| p
[0016] Where p is 1 or 2, corresponding to the L1 or L2 distance respectively.
[0017] Step 3: Set a weight vector for the relation of each embedded triple:
[0018] w i =exp(-|d i |)i=1,2…n
[0019] Where i represents the embedding dimension of each entity, w i This represents the embedding vector for each triplet.
[0020] By embedding entities and relations on the knowledge graph, a vector representation with knowledge graph semantics is obtained.
[0021] Furthermore, step two specifically includes:
[0022] The first step, after the knowledge graph is embedded, is to update the entity representation by performing neighbor aggregation on the entities in the knowledge graph. The specific aggregation function is shown below:
[0023]
[0024] In the formula, i is the head entity, j is the tail entity, r is the relation, ε is the edge union, and e is the tail entity. j Let w be the embedding vector of entity j. r Let r be the embedding vector of relation r, and N(i) be the set of neighbors of entity i.
[0025] By aggregating entities on a knowledge graph, we obtain information representations of relationships with knowledge graph semantics.
[0026] The second step involves the Graph Convolutional Network (GCN) iteratively aggregating information from neighboring nodes through a multi-layered structure to update the node representation vector.
[0027] The Graph Convolutional Network (GCN) model is used to learn representations of user-item interaction information. The specific formula is shown below:
[0028]
[0029] Among them, H (l) The node information of layer l is represented, and a new representation is obtained after passing through the graph convolutional layer.
[0030] This represents a graph adjacency matrix that includes self-loop matrices. W represents the corresponding degree matrix. (l) Let represent the learnable weight matrix of the l-th layer, and σ(·) represent the nonlinear activation function.
[0031] Furthermore, step three specifically includes:
[0032] Representing items in a knowledge graph and interactive information representation Send into the fusion gate In this system, one can obtain a dual item representation that integrates interactive information related to knowledge. And integrate gate units The contributions of these two types of information can be adaptively adjusted, as shown in the following formula:
[0033]
[0034] in, Indicates fusion gate aggregation unit, This represents items that possess knowledge graph information. This represents an item representation with interactive information. Enhanced item representations are obtained through aggregation using a fusion gate.
[0035] Furthermore, step four specifically includes:
[0036] Prediction and optimization:
[0037] The final item representation vector is used to predict user preferences and make content recommendations accordingly.
[0038] The first step involves generating representations of user u and item i after multiple convolutions, and then summing them together to obtain the final user representation and item representation.
[0039]
[0040] In the formula, User e u The representation of the Lth layer, This represents the Lth layer of the fused item.
[0041] The second step, in order to predict the probability of user interaction with items, involves performing an inner product operation on the final representations of the user and the item, as shown in the following formula:
[0042]
[0043] The third step involves an enhanced information representation graph neural network model (Inact-Knowledge Graph Convolutional Network, IKGCN), which uses the Adam optimizer to learn the model parameters. In each iteration, the estimation of the first and second moments of the gradient is used to adjust the parameter update magnitude, as shown below:
[0044]
[0045] In the formula, θ t Represented as model parameters, f(θ) t ) represents the loss function, and λ represents the set gradient rate of change.
[0046] This invention also provides a recommendation system based on the fusion of knowledge graph and user-item interaction graph, comprising:
[0047] Knowledge graph construction module: used to construct a knowledge graph containing items and their relationships based on data, and generate initial item representation vectors through graph embedding methods;
[0048] Entity aggregation module: Used to aggregate entities in the knowledge graph and update the entity representation vectors;
[0049] Graph Neural Network Module: Used to aggregate user-item interaction graphs. It iteratively aggregates information of neighboring nodes through a multi-layered Graph Convolutional Network (GCN) to update the node representation vectors.
[0050] Gated aggregation module: used to fuse representation vectors from knowledge graph and user item interaction graph, and calculate corresponding weight values through gating units to generate the final item representation vector;
[0051] Recommendation prediction module: Used to predict user preferences based on the final item representation vector and generate recommended content.
[0052] Furthermore, the knowledge graph construction module includes:
[0053] Knowledge graph construction unit: used to extract entities and relationships from a dataset and build a knowledge graph;
[0054] Graph embedding unit: used to define the TransE scoring function, embed triples according to L1 or L2 distance and generate initial item representation vectors;
[0055] Weighting unit: Used to set a weight vector for each embedded triple relation to generate a representation vector for subsequent aggregation.
[0056] Furthermore, the entity aggregation module includes:
[0057] Neighbor aggregation unit: used to aggregate and update the neighbor set of each entity after embedding the knowledge graph;
[0058] Aggregation function unit: Used to perform aggregation calculations on the neighbor set of an entity based on the embedding vectors of the entity and the relation, and generate a relation information representation with knowledge graph semantics.
[0059] Furthermore, the recommendation prediction module includes:
[0060] Representation generation unit: used to generate the final representation of users and items after multiple graph convolutions;
[0061] Interaction prediction unit: used to perform inner product operation on the final representation of user and item to calculate the probability of user interaction with item;
[0062] Optimization Unit: The Adam optimizer is used to iteratively update the model parameters based on the estimates of the first and second moments of the gradient to optimize the recommendation performance.
[0063] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:
[0064] First, in existing recommender systems, collaborative filtering algorithms based on user-item interactions and content-based recommender algorithms are two mainstream methods. However, collaborative filtering algorithms perform poorly in terms of data sparsity and the cold start problem, especially when user behavior data is insufficient or new items are introduced, making it difficult to accurately model user interests and preferences. Meanwhile, content-based recommender algorithms, due to their strong reliance on item feature representation, struggle to capture deep semantic relationships, resulting in insufficient generalization ability of recommendation performance. Furthermore, traditional methods, when modeling user-item relationships, typically rely on a single information source, lacking the ability to comprehensively utilize multimodal or heterogeneous data and perform semantic mining, thus limiting the accuracy and robustness of recommender systems.
[0065] This invention proposes a recommendation method based on the fusion of knowledge graphs and user-item interaction graphs to address issues such as data sparsity, cold start, and insufficient semantic information extraction in existing recommendation systems. By introducing a knowledge graph construction module, items and their relationships are structurally modeled, and initial item representation vectors are generated using graph embedding methods, thus providing the recommendation system with rich background knowledge and semantic information. Simultaneously, a graph convolutional network (GCN) is employed to perform feature aggregation and relationship reasoning on the user-item interaction graph, enabling the full mining of potential high-order relationships in heterogeneous data, thereby significantly improving the accuracy and robustness of recommendations.
[0066] Compared to traditional recommendation algorithms, this invention achieves effective integration of information sources and deep semantic mining by introducing a dual fusion mechanism of knowledge graph and user-item interaction graph. In this invention, a gated aggregation unit adaptively fuses representation vectors from different sources, dynamically adjusting the contribution ratio of knowledge associations and interaction information according to the specific task, thereby improving the expressive power and generalization performance of the recommendation model. Furthermore, the feature aggregation and optimization strategy based on GCN can better capture the potential associations between users and items, significantly improving the model's training efficiency and convergence speed.
[0067] The recommendation method of this invention is applicable to multiple fields such as e-commerce platforms, social media, online learning systems, and news recommendations, and shows significant technical advantages, especially in personalized recommendations and intelligent content delivery. By introducing the fusion technology of knowledge graphs and user-item interaction graphs, a more refined relationship can be established between complex user behaviors and item features, providing users with more accurate, timely, and personalized content recommendations. Furthermore, the algorithm architecture of this invention has good scalability and portability, and can be widely applied to different types of datasets and recommendation scenarios, possessing high industrial promotion value and application prospects.
[0068] Secondly, this invention proposes a recommendation method that integrates user behavior data and item knowledge graphs, overcoming the limitations of traditional systems that use behavior data or graph information in isolation. This approach combines the basic attributes of items and their multi-level relationships in the knowledge graph with user interaction behavior for joint modeling, effectively improving the accuracy and information richness of the recommendation system.
[0069] By introducing knowledge graphs, recommendation systems can clearly display recommendation paths, such as common interests among user preferences or semantic attribute relationships between items, thereby improving the transparency and interpretability of the system, making it easier for users to understand and accept recommendation results, and enhancing trust and user stickiness.
[0070] This invention employs an entity aggregation method in the process of constructing a knowledge graph. While preserving key semantic information, it significantly reduces the computational complexity of redundant nodes and edges in the graph, improves the overall model's operating efficiency, and provides an efficient foundation for large-scale recommendation.
[0071] This solution incorporates a gated aggregation unit to dynamically control the fusion ratio of information from different data sources. This mechanism effectively integrates the initial knowledge graph representation with user-item interaction features learned through a graph convolutional network (GCN), enabling a flexible and refined fusion strategy and significantly improving recommendation performance.
[0072] To address the issues of data sparsity and cold start, the solution leverages the rich entity attributes and relationships of knowledge graphs, combined with behavioral data, and employs a gating mechanism to mitigate sparsity and enhance cold start recommendation capabilities. Simultaneously, it introduces a time decay function and GCN to model user interest dynamics, capturing trends in interest changes and ensuring that recommendations better match users' current needs.
[0073] This technical solution systematically addresses three common biases in recommender systems: first, "data utilization bias" that relies solely on behavioral data; second, "model selection bias" due to a single model structure; and third, "information fusion bias" resulting from simple splicing and fusion. Through a combination of techniques including TransE graph construction, GCN aggregation, and gating and weighting, a multi-dimensional, structurally innovative, and high-performance recommender system architecture is constructed, providing new insights for the development of recommender systems. Attached Figure Description
[0074] Figure 1 This is a flowchart illustrating the method steps of an example of the present invention.
[0075] Figure 2 This is a flowchart illustrating the knowledge graph and interaction graph fusion method and system based on a gating aggregation mechanism provided in this embodiment of the invention.
[0076] Figure 3 This invention provides a method for constructing a knowledge graph and generating initial item representation vectors.
[0077] Figure 4 This invention provides prediction and optimization in its embodiments.
[0078] Figure 5-7 This is relevant evidence of the technical effects obtained by the embodiments of the present invention. Detailed Implementation
[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0080] See Figure 1 As shown, the method steps of this invention include the following key steps:
[0081] S101: Construct a knowledge graph, embed the knowledge graph into a graph, and aggregate entities.
[0082] S102: Aggregate the user-item interaction graph.
[0083] S103: Use a gating aggregation unit to perform information fusion.
[0084] Gated aggregation units are used to selectively extract high-quality information from knowledge graphs and interaction graphs to generate dual-item representations. This step aims to better capture the full informational content of the course and user behavior patterns.
[0085] S104: Prediction and Optimization
[0086] The final item representation vector is used to predict user preferences and make content recommendations accordingly.
[0087] Combined with appendix Figure 2 Further description.
[0088] Figure 2 The document demonstrates the specific implementation process, including data acquisition, knowledge graph construction, the application of gating aggregation units, and recommendation generation. Each step emphasizes how to effectively integrate data from different sources to improve the accuracy and relevance of user preference predictions. The steps include:
[0089] like Figure 3 As shown, the knowledge graph is constructed and initial item representation vectors are generated. The specific method is as follows:
[0090] S1: Knowledge Graph Construction:
[0091] Build and visualize a course knowledge graph using data sources;
[0092] S2: Define the scoring function of TransE and use the transE method for graph embedding.
[0093] f(h,r,t)=||h+rt|| p
[0094] Where p is 1 or 2, corresponding to the L1 or L2 distance respectively.
[0095] S3: Set a weight vector for the relation of each embedded triple:
[0096] w i =exp(-|d i |)i=1,2…n
[0097] Where i represents the embedding dimension of each entity, w i This represents the embedding vector for each triplet.
[0098] S4: Update the entity representation by performing neighbor aggregation on entities in the knowledge graph. The specific aggregation function is as follows:
[0099]
[0100] Where i is the head entity, j is the tail entity, r is the relation, ε is the edge set, and e is the tail entity. j Let w be the embedding vector of entity j. r Let r be the embedding vector of relation r, and N(i) be the set of neighbors of entity i.
[0101] 2. Entity aggregation is performed on the knowledge graph, and a graph neural network (GCN) is used to aggregate the user-item interaction graph.
[0102] 2.1 Based on user interaction records, a user-item interaction graph is constructed. A graph neural network (GCN) model is used to learn the representation of user-item interaction information. The specific formula is shown below:
[0103]
[0104] Among them, H (l) The node information of layer l is represented, and a new representation is obtained after passing through the graph convolutional layer.
[0105] This represents a graph adjacency matrix that includes self-loop matrices. W represents the corresponding degree matrix. (l) Let represent the learnable weight matrix of the l-th layer, and σ(·) represent the nonlinear activation function.
[0106] 2.2 Multi-layer information aggregation
[0107] 2.3 Output Representation
[0108] After L layers of stacking:
[0109] The user ultimately stated: h u =H (L) [u,:]
[0110] The final representation of the item is: h i =H (L) [i,:]
[0111] h u The final representation of user u is a d-dimensional vector.
[0112] H (L) The output representation matrix of the Lth layer of the graph neural network has dimensions (N+M)×d.
[0113] N: Number of users
[0114] M: Quantity of items
[0115] d: Embedding dimension (representing the length of the vector)
[0116] [u,:]: Select matrix H (L) The row corresponding to user u (i.e., the embedding vector of that user).
[0117] h i The final representation of item i is also a d-dimensional vector.
[0118] [i,:]: Select matrix H (L) The row corresponding to item i
[0119] By stacking multiple graph neural network aggregation layers, user item interaction information and user item feature information are aggregated. Neighbor information is aggregated to update the user item feature information, and finally, the information representation of the user's items is output.
[0120] 2.4 A Graph Neural Network (GCN) is used to aggregate the user-item interaction graph.
[0121] The graph neural network model GCN is used to learn representations of user-item interaction information. The specific formula is shown below:
[0122]
[0123] Among them, H (l) The node information of layer l is represented, and a new representation is obtained after passing through the graph convolutional layer.
[0124] This represents a graph adjacency matrix that includes self-loop matrices. W represents the corresponding degree matrix. (l)Let represent the learnable weight matrix of the l-th layer, and σ(·) represent the nonlinear activation function.
[0125] 3. Design and implement a gating aggregation unit to receive item representation vectors from steps one and two, and calculate the corresponding weight value for each type of representation vector.
[0126] The final item representation vector is used to predict user preferences and make content recommendations accordingly.
[0127] Representing items in a knowledge graph and interactive information representation Send into the fusion gate In this system, one can obtain a dual item representation that integrates interactive information related to knowledge. And integrate gate units The contributions of these two types of information can be adaptively adjusted, as shown in the following formula:
[0128]
[0129] in, Indicates fusion gate aggregation unit, This represents items that possess knowledge graph information. This represents an item representation with interactive information. Enhanced item representations are obtained through aggregation using a fusion gate.
[0130] 4 such as Figure 4 As shown, the prediction and optimization include:
[0131] The final item representation vector is used to predict user preferences and make content recommendations accordingly.
[0132] S21, after multiple convolutions, finally generate representations of user u and item i, and then add them together as the final user representation and item representation.
[0133]
[0134] In the formula, User e u The representation of the Lth layer, This represents the Lth layer of the fused item.
[0135] S22, In order to predict the probability of user interaction with items, an inner product operation is performed on the final representations of the user and the items, as shown in the following formula:
[0136]
[0137] y uiThis represents the predicted interaction score (such as click-through rate, purchase probability, etc.) between user u and target item i.
[0138] The final embedding representation of user u is a d-dimensional column vector (∈R) d ).
[0139] The final embedding representation of the target item i is also a d-dimensional column vector.
[0140] S23, IKGCN uses the Adam optimizer to learn model parameters. In each iteration, it uses estimates of the first and second moments of the gradient to adjust the parameter update magnitude, as shown below:
[0141]
[0142] In the formula, θ t Represented as model parameters, f(θ) t ) represents the loss function, and λ represents the set gradient rate of change.
[0143] The recommendation method of this invention is applicable to multiple fields such as e-commerce platforms, social media, online learning systems, and news recommendations, and shows significant technical advantages, especially in personalized recommendations and intelligent content push. When used in online education platforms, it can provide personalized learning paths and course recommendations based on users' learning materials, course completion status, and course relationships in a knowledge graph. For example, it can recommend suitable subsequent courses or supplementary materials based on a student's learning progress and interests. When used in e-commerce platforms such as Taobao and JD.com, it can help users discover products they may be interested in by analyzing their browsing history, purchasing behavior, and the relationships between products. For example, it can recommend related or complementary products by combining the user's shopping cart content and favorites information. When used in film and music platforms, it can accurately recommend movies, TV series, or music works based on users' viewing habits and listening preferences. For example, it can provide personalized recommendation lists by utilizing users' historical playback records and rating data, combined with metadata of film works or songs (such as director, actors, style, etc.). Through the above application examples in different fields, it can be seen that this invention not only improves user experience but also helps enterprises better understand user needs, thereby providing more accurate service and product recommendations. This marks an important step forward in the development and application of intelligent recommendation systems.
[0144] This invention uses the MOOCCube dataset for experimental evaluation, employing it as the interactive information for user courses. The data is randomly divided into an 80% training set, a 10% validation set, and a 10% test set.
[0145] The following is combined Figure 5-7 Further explanation of the parameter experiments for this invention is provided. The performance of the MOOCCube dataset on two evaluation dimensions—recall and normalized depreciation cumulative gain (ndcg)—is presented, revealing the dynamic impact of parameter adjustments on model performance.
[0146] Recall reflects the extent to which user-preferred items are covered in a recommender system, and its expression is shown below.
[0147]
[0148] In the formula, R(u) represents the items recommended by user u, and T(u) represents the set of items that user u likes on the test set.
[0149] Normalized Discounted Cumulative Gain (NDCG) is a metric used to evaluate the performance of recommender systems, particularly useful in measuring the correlation between the ranking of recommended items and actual user feedback. This metric calculates a weighted score by assigning differentiated weight coefficients to items at different positions in the recommendation sequence and then standardizing the result. Quantitatively, NDCG ranges from [0,1], and its value is positively correlated with the performance of the recommender system; that is, the closer the value is to 1, the better the performance of the recommender algorithm. The specific calculation is shown below:
[0150]
[0151] In the formula, rel i This represents the relevance score of the i-th ranked item. Let |U| represent the relevance score of the i-th item in the ideal ranking, and |U| represent the set of users.
[0152] Figure 5Comparing the impact of different numbers of convolutional layers on model performance, we found that as network depth increases, the system can capture more complex feature interactions and semantic associations, thereby improving the accuracy of user behavior prediction. However, excessively deep network structures may lead to decreased model generalization ability and excessive computational resource consumption. Therefore, this study uses a controlled variable method to conduct experiments within four levels {1, 2, 3, 4} to determine the optimal convolutional layer configuration to achieve the best balance between model prediction performance and computational efficiency. As shown in the figure, model performance and network depth exhibit a non-linear relationship. Experiments demonstrate that appropriately deepening the network structure can effectively improve model performance, thanks to the effective supplementation of second-order neighborhood information by higher-order connections. By aggregating collaborative features and knowledge associations from multi-hop neighborhoods, the model can learn more representative node embeddings, thus more accurately characterizing user interest features. Experimental results on the MOOCCube dataset show that the model achieves optimal performance when the number of convolutional layers is set to 3. It is noteworthy that further increasing network depth leads to performance degradation, a phenomenon that may be related to excessive smoothing of node features.
[0153] Figure 6 Training was performed with three different learning rate parameters: 0.0001, 0.001, and 0.01. (The sentence is incomplete and requires further context.) Figure 6 The experimental results show that the model's performance declines significantly as the learning rate increases. On the MOOCCube dataset, the model's recall performance decreased by 12% when the learning rate increased from 0.0001 to 0.01. Furthermore, the learning rate has a more significant impact on model performance than the number of convolutional layers. The main reason is that a larger learning rate leads to larger step sizes in parameter updates during training, causing the model to skip optimal solutions and fail to converge effectively.
[0154] Figure 7 First, we analyze the impact of embedding dimension in the model. Low dimension offers high computational efficiency and small content footprint, saving processing time. However, excessively low dimension can lead to overly strong feature compression, potentially losing information from local or the entire graph, resulting in underfitting. Conversely, while high dimension can capture more complex graph structures and node features, improving the model's expressive power, it also incurs excessive computational costs and may introduce noise, leading to overfitting. Therefore, a suitable embedding dimension is crucial. We conducted experiments with embedding dimensions {4, 8, 16, 32, 64, 128}. Experimental results show that... Figure 7 As shown, model performance and embedding dimension settings exhibit a non-linear relationship.
[0155] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.
[0156] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A recommendation method based on the fusion of knowledge graph and user-item interaction graph, characterized in that, Includes the following steps: Step 1: Construct an item knowledge graph. Use the distance-based embedding model TransE to vectorize the entities and relations in the knowledge graph and represent them in the same vector space, and then perform entity aggregation. Step 2: Use a graph neural network (GCN) to aggregate the user-item interaction graph; Step 3: Design and implement a gated aggregation unit to calculate the corresponding weight values for the representation vectors obtained in Step 1 and Step 2, and merge the two vectors; Step 4: Use the final item representation vector to predict user preferences and make content recommendations accordingly.
2. The recommendation method based on the fusion of knowledge graph and user-item interaction graph as described in claim 1, characterized in that, Step one specifically includes: Step 1: Construct a knowledge graph based on the known dataset; The second step is to use the distance-based embedding model TransE to vectorize the entities and relations in the knowledge graph and represent them in the same vector space. Step 3: Set weight vectors for each embedded triplet relation. w i =exp(-|d i |) i=1,2…n Where i represents the embedding dimension of each entity, w i This represents the embedding vector for each triplet. Step 4: After the knowledge graph is embedded, neighbor aggregation is performed on the entities in the knowledge graph to update the entity representation. The specific aggregation function is shown in Equation (1): In the formula, i is the head entity, j is the tail entity, r is the relation, ε is the edge union, and e is the tail entity. j Let w be the embedding vector of entity j. r Let be the embedding vector of relation r, and N(i) be the set of neighbors of entity i. Information representations of relations with knowledge graph semantics are obtained by aggregating entities on the knowledge graph.
3. The recommendation method based on the fusion of knowledge graph and user-item interaction graph as described in claim 1, characterized in that, Step two specifically includes: The graph neural network model GCN is used to learn representations of user-item interaction information. The specific formula is shown below: Among them, H (l) The node information of layer l is represented, and a new representation is obtained after passing through the graph convolutional layer. This represents a graph adjacency matrix that includes self-loop matrices. W represents the corresponding degree matrix. (l) Let represent the learnable weight matrix of the l-th layer, and σ(·) represent the non-linear activation function. Graph Neural Networks (GCNs) iteratively aggregate information from neighboring nodes through a multi-layered structure to update the node representation vectors.
4. The method according to claim 1, characterized in that, The gated aggregation unit in step three includes: Representing items in a knowledge graph and interactive information representation Send into the fusion gate In this system, one can obtain a dual item representation that integrates interactive information related to knowledge. And integrate gate units The contributions of these two types of information can be adaptively adjusted, as shown in the following formula: in, Indicates fusion gate aggregation unit, This represents items that possess knowledge graph information. This represents an item representation with interactive information. Enhanced item representations are obtained through aggregation using a fusion gate.
5. The method according to claim 1, characterized in that, The prediction and optimization in step four include: The final item representation vector is used to predict user preferences and make content recommendations accordingly. The first step involves generating representations of user u and item i after multiple convolutions, and then summing them together to obtain the final user and item representations. In the formula, User e u The representation of the Lth layer, This represents the Lth layer of the fused item. The second step, in order to predict the probability of user interaction with items, involves performing an inner product operation on the final representations of the user and the item, as shown in the following formula: Third, IKGCN uses the Adam optimizer to learn the model parameters. In each iteration, it uses the estimates of the first and second moments of the gradient to adjust the parameter update magnitude, as shown in Figure 4-8. In the formula, θ t Represented as model parameters, f(θ) t ) represents the loss function, and λ represents the set gradient rate of change.
6. A recommendation system based on the fusion of knowledge graph and user-item interaction graph, comprising: Knowledge graph construction module: used to construct a knowledge graph containing items and their relationships based on data, and generate initial item representation vectors through graph embedding methods; Entity aggregation module: Used to aggregate entities in the knowledge graph and update the entity representation vectors; Graph Neural Network Module: Used to aggregate user-item interaction graphs. It iteratively aggregates information from neighboring nodes using a multi-layered Graph Neural Network (GCN) to update the node representation vectors. Gated aggregation module: used to fuse representation vectors from knowledge graph and user item interaction graph, and calculate corresponding weight values through gating units to generate the final item representation vector; Recommendation prediction module: Used to predict user preferences based on the final item representation vector and generate recommended content.
7. The recommendation system according to claim 6, characterized in that, The entity aggregation module includes: Neighbor aggregation unit: used to aggregate and update the neighbor set of each entity after embedding the knowledge graph; Aggregation function unit: Used to perform aggregation calculations on the neighbor set of an entity based on the embedding vectors of the entity and the relation, and generate a relation information representation with knowledge graph semantics.
8. The recommendation system according to claim 6, characterized in that, The recommendation prediction module includes: Representation generation unit: used to generate the final representation of users and items after multiple graph convolutions; Interaction prediction unit: used to perform inner product operation on the final representation of user and item to calculate the probability of user interaction with item; Optimization Unit: The Adam optimizer is used to iteratively update the model parameters based on the estimates of the first and second moments of the gradient to optimize the recommendation performance.