Item recommendation method and system based on social guiding relationship

Abstract

The invention discloses an article recommendation method and system based on a social guidance relationship, and relates to the technical field of article recommendation, and the method comprises the following steps: embedding a knowledge graph into a graph convolutional network, obtaining the social guidance relationship between any two first nodes, and obtaining a similarity score between the two first nodes based on the social guidance relationship; distributing guiding relation weights for different first nodes through an attention mechanism to obtain guiding relation embedding, and combining the similarity score with the guiding relation embedding to obtain a guiding relation value between the two first nodes; establishing a guiding relation network based on the guiding relation value between any two first nodes; and in the guiding relation network, node information of a second node connected with the target node is aggregated to the target node, and article recommendation is performed on the recommendation node through the plurality of connection channels. The recommendation accuracy and novelty are remarkably improved, the information cocoon house problem is avoided, and the user view is widened.

Description

Technical Field

[0001] The present invention relates to the technical field of item recommendation, and particularly to an item recommendation method and system based on social-oriented relationships. Background Art

[0002] In today's digital age, as a key information filtering technology, item recommendation systems have played an indispensable role in processing massive amounts of data, especially having a profound impact on the behaviors of hundreds of millions of Internet users on various platforms such as short videos, live broadcasts, and fast social interactions.

[0003] The effectiveness of these recommendation systems largely depends on their ability to process and interpret complex user relationship networks, which usually exist in the form of higher-order relationship networks in modern society. As a structured data form, knowledge graphs have shown great potential in improving the accuracy, interpretability, and personalization of recommendation systems due to the rich relationships and attribute characteristics between their entities. Therefore, how to make full use of these rich information resources, especially how to reveal and utilize the indirect or implicit social relationships in knowledge graphs, has always been a difficult problem in the research of recommendation systems.

[0004] Traditional item recommendation systems, such as those based on collaborative filtering and content recommendation, often struggle when it comes to leveraging and understanding such higher-order relationship networks, especially when considering the rich semantic information in knowledge graphs. The current process of recommendation systems generally includes data collection and preprocessing, recall stage, ranking stage, re-ranking and filtering, and recommendation result generation and evaluation. However, a large amount of computation in traditional recommendation systems is consumed in the stages of collecting user data, ranking, and re-ranking and filtering. As a result, the recommended results thus formed usually have too strong a correlation with this user itself, and it is very easy to form an information cocoon. Summary of the Invention

[0005] Based on the defects existing in the above-mentioned prior art, the present invention provides an item recommendation method and system based on social-oriented relationships, which solves the problem that a large amount of computation in traditional recommendation systems is consumed in the stages of collecting user data, ranking, and re-ranking and filtering, and the recommended results thus formed usually have too strong a correlation with this user itself, and it is very easy to form an information cocoon.

[0006] The present invention adopts the following technical solutions: In a first aspect, the present invention provides an item recommendation method based on social-oriented relationships, including the following steps: Collect item data and user data of multiple users, and construct a knowledge graph based on the item data and user data; Embed the knowledge graph into a graph convolutional network, where multiple users serve as the first nodes and multiple items serve as the second nodes; Obtain the similarity score between two corresponding first nodes based on the social orientation relationship between any two users; assign orientation relationship weights to different first nodes through the attention mechanism, and obtain the orientation relationship value between the two first nodes based on the similarity score and the orientation relationship weight; establish an orientation relationship network based on the orientation relationship values between any two first nodes; In the orientation relationship network, take any one first node as the target node, and take another first node that has no orientation relationship with the target node as the recommended node, and establish multiple connection channels between the target node and the recommended node. Each connection channel is obtained by connecting multiple first nodes with all orientation relationships enabled; when the orientation relationship value between two first nodes is greater than the corresponding orientation relationship threshold, the orientation relationship is enabled; Aggregate the node information of the second node connected to the target node to the target node, and perform item recommendation for the recommended node through multiple connection channels.

[0007] Preferably, the similarity score is specifically as follows: ; In the formula, is the similarity score between the first node u and v , is the trainable weight, is the social orientation relationship between the two first nodes, is the node feature of the first node v , T is the matrix transpose operation.

[0008] Preferably, the step of assigning orientation relationship weights to different first nodes through the attention mechanism and obtaining the orientation relationship value between the two first nodes based on the similarity score and the orientation relationship weight specifically includes the following steps: Assign orientation relationship weights to different first nodes through the attention mechanism to obtain the orientation relationship embedding; Combine the similarity score with the orientation relationship embedding to obtain the orientation relationship value between the two first nodes.

[0009] Preferably, the orientation relationship embedding is specifically as follows: ; In the formula, is the orientation relationship embedding, is the embedding vector of the relationship , is the attention score, R is the set of all relationships r .

[0010] Preferably, the guiding relationship value is specifically as follows: ; In the formula, is the guiding relationship value, is the trainable weight parameter, is the guiding relationship embedding.

[0011] Preferably, the item recommendation for the recommendation node through multiple connection channels specifically includes the following steps: Train the guiding relationship network to obtain multiple trained trainable weights, and generate multiple first connection channels according to the multiple trained trainable weights; Set a first channel length threshold, retain multiple first connection channels smaller than the first channel length threshold to obtain multiple second connection channels, and perform item recommendation for the recommendation node through the multiple second connection channels; Set a second channel length threshold, retain multiple second connection channels smaller than the second channel length threshold to obtain multiple third connection channels, and perform item recommendation for the recommendation node through the multiple third connection channels; wherein, the first channel length threshold is smaller than the second channel length threshold.

[0012] Preferably, the user data includes browsing duration, user evaluation, and questionnaire feedback, and the item data includes the click volume, transaction volume, item category, and item attributes of the item or items.

[0013] In a second aspect, the present invention provides an item recommendation system based on a social guiding relationship, including: An acquisition unit for acquiring item data and user data of multiple users, and constructing a knowledge graph based on the item data and user data; An embedding unit for embedding the knowledge graph into a graph convolutional network, wherein multiple users are used as the first nodes and multiple items are used as the second nodes; A construction unit for obtaining a similarity score between corresponding two first nodes based on the social guiding relationship between any two users; allocating guiding relationship weights to different first nodes through an attention mechanism, and obtaining a guiding relationship value between the two first nodes based on the similarity score and the guiding relationship weights; establishing a guiding relationship network based on the guiding relationship values between any two first nodes; An establishment unit for, in the guiding relationship network, taking any one of the first nodes as the target node and taking another first node that has no guiding relationship with the target node as the recommendation node, establishing multiple connection channels between the target node and the recommendation node, and each connection channel is connected by multiple first nodes with all guiding relationships enabled; when the guiding relationship value between two first nodes is greater than the corresponding guiding relationship threshold, the guiding relationship is enabled; A recommendation unit is used to aggregate the node information of a second node connected to a target node to the target node, and perform item recommendation for the recommendation node through multiple connection channels.

[0014] Compared with the prior art, the above at least one technical solution adopted by the present invention can achieve the following beneficial effects: The present invention first constructs a knowledge graph based on the item data and user data of multiple users, and embeds the knowledge graph into a graph convolutional network. Then, the social orientation relationship between any two first nodes is obtained, and the interaction relationship between users in the knowledge graph is deeply mined to obtain the corresponding similarity score. At the same time, an attention mechanism is used to assign an orientation relationship weight to different first nodes, thereby obtaining an orientation relationship embedding. Based on the attention mechanism, relationship weights are assigned to different users, and the similarity score is combined with the orientation relationship embedding to obtain the orientation relationship value between two first nodes; an orientation relationship network is established based on the orientation relationship value between any two first nodes. The present invention deeply mines the complex social relationships of users in the knowledge graph, and combines the attention mechanism to assign relationship weights to different users, thereby establishing an orientation relationship network for users.

[0015] In the orientation relationship network, the concept of an orientation relationship threshold is proposed. This orientation relationship threshold is a control value for the priority of using this orientation relationship. Compared with existing recommendation methods, a recommendation technology with both breadth and depth is proposed. Based on the activation of the orientation relationship, multiple connection channels are established to construct a complete recommendation path, significantly improving the recommendation accuracy and novelty, avoiding the information cocoon problem, and broadening the user's vision. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1 It is a flowchart of an item recommendation method based on social orientation relationship of the present invention; Figure 2 It is a schematic structural diagram of a recommendation system involved in the present invention; Figure 3 It is a schematic diagram of various orientation relationships modeled by the present invention; Figure 4 It is a schematic diagram of an embodiment of a recommendation system involved in the present invention; Figure 5 It is a flowchart of movie recommendation involved in the present invention; Figure 6Schematic diagram of the relationship network constructed by the present invention for new users with scarce data; Figure 7 Schematic diagram of the complete traceable recommendation path constructed by the present invention; Figure 8 Schematic diagram of the recommendation principle of the accuracy and novelty content involved in the present invention; Among them, Figure 8 (a) of : Schematic diagram of the principle of focusing on accuracy recommendation, Figure 8 (b) of : Schematic diagram of the principle of focusing on novelty recommendation, Figure 8 (c) of : Schematic diagram of the principle of focusing on both accuracy and novelty recommendation; Figure 9 Interest index diagram of the recommendation mode of the existing recommendation system; Figure 10 Interest index diagram of the recommendation mode that takes into account both accuracy and novelty adopted by the present invention; Figure 11 Schematic diagram of the effect of reducing the system calculation amount by the present invention through threshold setting; Figure 12 Schematic diagram of the principle of the recommendation algorithm of the present invention for selecting recommended items. Detailed implementation manners

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] Explanation of relevant terms involved in the present invention: Neighborhood information enhancement: An image processing method that improves the image quality by analyzing and adjusting the relationship between each pixel point and its neighboring pixel points in the image, especially in terms of details and contrast. This technology is widely used in fields such as image restoration, medical imaging, and satellite images, and can enhance the visibility and key information of the image.

[0020] Social circle integration: A concept in social network analysis, which refers to integrating and modeling the strong and weak interpersonal relationships in different social circles or communities. This integration helps to more deeply understand the structure and dynamics of the social network, and is crucial for discovering key individuals in social circles, predicting social trends, and formulating effective social media marketing strategies.

[0021] Knowledge Graph: A complex data structure, which is structured data composed of "head entity - relationship - tail entity", and can be used to organize and express the complex relationships between entities. In a recommendation system, the Knowledge Graph can improve the accuracy and personalization level of recommendations and enhance the user experience by connecting relevant entities such as users, products, and services.

[0022] Strong and Weak Ties Theory: In sociological definitions, strong ties refer to a close interpersonal relationship, commonly seen among close friends, family members, and colleagues with close cooperation. This relationship plays a key role in social network analysis, especially showing its unique value in aspects such as information transmission, trust building, and group cohesion analysis. Different from strong ties, weak ties refer to less frequent and less in-depth social relationships, but they show their unique value in promoting information circulation across social circles, inspiring innovation, and constructing social capital. This theory has been widely applied in social network analysis, marketing strategies, and organizational behavior research.

[0023] Attention Mechanism: A feature of a computational model that can dynamically adjust the attention of the model to different parts of the input data. By assigning different degrees of attention to different parts of the model, this mechanism can help the model more effectively identify and process more critical information. In complex data environments such as natural language processing or image recognition, the attention mechanism can significantly improve the performance and accuracy of the model by capturing key information.

[0024] Graph Convolutional Network (GCN): A neural network structure specifically designed to process graph data. By learning the features of nodes and their connections in a graph, GCN can effectively process and interpret complex network structure data, and is often applied in fields such as image processing and social network analysis.

[0025] Guided Fusion Graph Convolutional Network: A neural network structure proposed in this invention that combines the advantages of sociological strong and weak ties theory, attention mechanism, and graph convolutional network, which can improve the model's ability to process complex user interaction data. This network structure can embed user neighbor relationship information into entities through convolutional aggregation operations, model the combination of strong and weak interpersonal ties in different social networks or social circles, and use user groups with similar interests in some aspects as the basic units for recommendation, so as to recommend high-quality recommended content within the same social circle and novel and surprising recommended content in different social circles for users, and achieve item recommendations closer to the structure of modern social relationship networks.

[0026] To solve the problems existing in the prior art, the present invention proposes an item recommendation method based on social-oriented relationships, adopting an innovative method that combines the knowledge graph neighborhood information enhancement technology and the social circle fusion technology based on the attention mechanism, explores the current and future interest development of users, and realizes the recommendation of various items that takes into account both accuracy and novelty. Refer to Figure 1 , including the following steps: S1: Collect the behavior data of users on different platforms, that is, the original data, which includes item data and user data.

[0027] The item data includes the click volume, transaction volume, item category, item attributes, etc. of the item, and the user data includes the browsing duration, user evaluation, and customer service / questionnaire feedback, etc. Clean, format, and standardize the original data to ensure the quality and efficiency of subsequent processing. Items include commodities, movies, books, videos, etc.

[0028] S2: Knowledge graph construction: It refers to the process of building the original data into a knowledge graph in the triple pattern of "head entity-relationship-tail entity".

[0029] In this embodiment, the head entity refers to the user, the relationship refers to the connection between the user and the user or the item, such as: data such as friends (user <-> user), evaluation (user -> item), and feedback information (user -> item), and the tail entity refers to the user or the item. And define the knowledge graph , where and represent the head entity and tail entity sets respectively, represents the relationship edge set.

[0030] First, embed the knowledge graph into the recommendation model. Multiple users are used as the first nodes, and multiple items are used as the second nodes.

[0031] S3: Oriented relationship modeling.

[0032] Before the oriented relationship modeling, a convolutional aggregation operation needs to be performed on multiple nodes. The convolutional aggregation operation represents a process of aggregating the features of neighbor nodes through convolutional operations. In graph data, nodes are connected by edges, and the convolutional operation is to perform information aggregation in the local neighborhood (that is, the nodes connected to the current node) to obtain a higher-level representation of the node.

[0033] The local neighborhood information refers to the information carried by the direct neighbor nodes of a certain node. In the graph, the representation of a node depends not only on the information of the node itself but also on the information of its neighbor nodes. By aggregating this neighbor information, the node can obtain richer context information.

[0034] The modeling of the guiding relationship is the core part of the present invention, which involves the definition of the social guiding relationship based on the distance correlation of the embedded relationship between any two users , that is: (1); In the formula, represents the distance correlation between and , where is a user pair, and the subscript

[0035] is the sequence number of the relationship. The effective guiding relationship value between any two users is calculated by formula (2). Then, by introducing the attention mechanism, the guiding relationship value is improved to obtain the final formula (7). Through the modeling of the guiding relationship, the present invention proposes an evaluation criterion that can quantitatively and intuitively describe the connection between users and nodes, as shown in formula (4).

[0036] The guiding relationship mining process can be defined as calculating the similarity score u between the first node v and : (2); In the formula, is the similarity score between the first node u and v , is the trainable weight, is the social guiding relationship between the two first nodes, is the node feature of the first node v , T is the matrix transpose operation. Among them, can be expressed as: (3); In the formula, is the attention weight parameter, is the node feature output by the network, L is a specific i value, represents the concatenation operation, is the output of the activation function.

[0037] Score prediction and attention mechanism: Using the Sigmoid function as the activation function, the score can be expressed as: (4); Actually, after obtaining the similarity scores of the first node u and v , the influence of time series on similarity can be improved through the attention mechanism. The present invention uses the attention mechanism to assign relationship weights to different users, and uses to represent that the attention mechanism creates a guiding relationship embedding: (5); In the above formula, is the embedding vector of the relationship , and is assigned an attention score to represent its importance, R is the set of all relationships r ; (6); Here, is the trainable weight of the relationship and the guiding relationship , is R except r among other relationships.

[0038] Single-path relationship: target user - guiding user - item to be recommended, where the guiding user is the user used to expand the social circle of the target user to effectively collect items to be recommended.

[0039] Furthermore, the attention weight and the embedding of the guiding relationship are introduced, that is: (7); In the above formula, is the guiding relationship value, is the trainable weight parameter.

[0040] Using the Sigmoid function as the activation function again, the score can be expressed as: (8); Finally, a simple binary cross-entropy loss function is defined to measure the error between the label and the measured score : (9).

[0041] By optimizing the loss function, the model can obtain better recommendation results among the guiding user groups.

[0042] Overall, the present invention mainly utilizes the structural information and user-oriented information of graph data, combines node features and user-oriented information, and optimizes feature learning and relationship mining based on the attention mechanism through a graph convolutional network, ultimately solving problems such as cold start, weak dynamic adaptability, poor interpretability, low utilization rate of user information, and easy formation of user information cocoons in existing recommendation systems.

[0043] S4: Setting the threshold for guiding relationship modeling.

[0044] In the process of guiding relationship modeling, only when the threshold of a certain type of guiding relationship is less than 1, will this guiding relationship be generated and accessed. The smaller the threshold, the higher the priority of this guiding relationship relative to other guiding relationships, that is, the model will form recommendation results through this type of guiding relationship more preferentially. The threshold here is the control value for the priority of using this guiding relationship, and it is effective when set to less than 1, that is, when the calculated by formula (7) is greater than this threshold, this guiding relationship will be enabled. Even if no type of guiding relationship can be generated and accessed, the algorithm can still generate recommendation results through the strong and weak connection channels in the original data. The specific threshold parameters are automatically generated during the training of the guiding fusion graph convolutional network.

[0045] Taking the graph attention network (GAT) that incorporates the attention mechanism as an example, its training process can be divided into the following key steps: Data selection: Select a dataset suitable for the graph structure, such as social networks, molecular structures, or knowledge graphs, etc.

[0046] Data preprocessing: Construct the adjacency matrix and node feature matrix of the graph. For graph data, it is usually necessary to calculate the neighbor information of nodes and normalize the features.

[0047] Model architecture: GAT introduces the attention mechanism to weight and aggregate the features of neighbor nodes.

[0048] Multi-head attention: To enhance the expression ability of the model, GAT usually adopts the multi-head attention mechanism, and splices or averages the attention results of different heads.

[0049] Loss function: Usually, the cross-entropy loss function is used to measure the difference between the predicted value and the true label of the model.

[0050] Optimizer: Commonly used optimizers include Adam, and its hyperparameters such as learning rate and weight decay need to be adjusted according to the task.

[0051] Forward propagation: Input node features and graph structure information, and calculate the representation of each node through the graph attention layer.

[0052] Backpropagation: Calculate the gradient of the loss function and update the model parameters through backpropagation.

[0053] Training loop: Repeat the forward propagation and backpropagation processes until the model converges or reaches the preset number of training epochs.

[0054] Validation set evaluation: Evaluate the model performance using the validation set. Common metrics include accuracy, recall, etc.

[0055] Early stopping mechanism: If the performance on the validation set no longer improves, stop training early to avoid overfitting.

[0056] Test set evaluation: Evaluate the final performance of the model on the test set to ensure that the model has good generalization ability.

[0057] Model saving: Save the trained model as a file for subsequent use.

[0058] S5: Setting the length threshold for strong and weak connection channels.

[0059] In the process of guiding relationship modeling, in the knowledge graph composed of original data, the process of accessing the length threshold of strong and weak connection channels. A channel refers to the path connected by a specific relationship between two specific nodes (the target node and the recommended node). There can be multiple channels between two specific nodes (i.e., connected through different relationships and nodes). The channel length refers to the number of relationships included in the path corresponding to a specific channel. The channel length threshold is a hyperparameter. When the channel length exceeds the threshold, the guiding relationship fails (the weight is assigned 0). When the channel length is less than the threshold, the guiding relationship is valid and the weight is valid. See Figure 11 The schematic diagram of the guiding relationship taking effect / failing as described.

[0060] The length of a certain strong and weak connection channel is proportional to the number of entities included in the channel. Only when the threshold is greater than 0 can the subsequent recommendation algorithm access the strong and weak connection channels. If the threshold is equal to 0, all relationships between all entities will not be accessible, and no recommendation results can be generated at this time. This threshold parameter is automatically generated during network training.

[0061] S6: Recall.

[0062] Recall refers to the process of using the established guiding relationship network to obtain as many valuable items for users as possible. In this link, the optimal user relationship network route will be automatically searched during the network training process and trainable weights will be generated. 。Here, it refers to the channel weight parameter, which is a result value obtained by summing up the weights of all relationships within a channel. It is calculated and generated during the process of finding the optimal user relationship network route, sorted according to the sum of the weights of each channel, and the several channels with the largest weights are selected when screening several channels. The channel with the largest channel weight is the optimal user relationship network route. See Figure 11 。

[0063] S7: Coarse ranking.

[0064] Coarse ranking refers to the process of using a small part of the established guiding relationship network to obtain as many valuable items for users as possible, sorting the predicted values of the items, and removing duplicates of the items. In this link, through the relationship network established in the previous link, duplicate items recommended for the same user are identified and this recommended route is trimmed to avoid duplication.

[0065] In this embodiment, by setting a relatively small channel length threshold, only a small part of the channels are activated, thereby reducing the computational amount when screening items and ensuring to a certain extent that the screened items are of interest to users. The nodes and relationships finally screened by coarse ranking will be applied to the fine ranking stage.

[0066] S8: Fine ranking.

[0067] It refers to the process of using most of the established guiding relationship network to obtain a small part of the most valuable items for users, sorting the predicted values of the items, and removing duplicates of the items. This link will amplify the contribution degree of the total score of some key guiding relationships to the recommendation result (increase its weight). This process is different from the previous process. This process targets the recommendation route after the network is basically formed, prioritizes the recommended items that users may be more interested in according to the value (the size of the sum of the weights of the guiding relationship networks of each item), and further screens out duplicate items.

[0068] In this embodiment, by setting a relatively large channel length threshold, most of the channels are activated, and thus an optimal path with a larger channel weight than that in the coarse ranking stage may be found (that is, a channel with a channel length greater than the channel length threshold in the coarse ranking but with a larger channel weight). Since the nodes and relationships applied in the fine ranking stage all come from the coarse ranking stage, the complexity of the guiding relationship network when bearing a relatively large channel length threshold is also within an acceptable range.

[0069] S9: Re-ranking: It refers to the process of reweighing the accuracy and novelty of the recommended items and carefully adjusting the sorting of the recommendation results after inserting other advertisements and other content according to the marketing strategy.

[0070] S10: Recommendation result: It refers to the process of generating the final recommendation result using the result after re-ranking.

[0071] S11: User operation: refers to the operation behaviors of users on the recommended results, such as liking, collecting, and commenting.

[0072] S12: Operation data: refers to the processing process of generating user operation behavior data based on user operation behaviors.

[0073] S13: Update: refers to the process of adding a new relationship network that causes effective user operation behaviors to the original data.

[0074] The above 13 steps together constitute the recommendation process of the present invention. Among them, compared with the existing recommendation system algorithms (such as the user-based collaborative filtering algorithm), the setting of the guiding relationship modeling threshold and the strong and weak connection channel length threshold for users proposes a recommendation technology with both breadth and depth. Based on the user relationship, it can form a guiding network among users, more comprehensively depict the user portrait, and perform in-depth recommendations according to the browsing information of other users with the same preferences as this user.

[0075] Refer to Figure 12 , first select a user 1 to be recommended, model the guiding relationship through the recommendation algorithm proposed by the present invention, and generate the N guiding relationship channels with the largest sum of guiding relationship weights. Then, according to the commodity purchase relationships of the users pointed to by these guiding relationship channels, after commodity sorting (including rough sorting, fine sorting, and re-sorting) and duplicate checking (synchronized with sorting), generate at most n recommended objects, and then conduct a trial recommendation for these n commodities: If the user feedback is not good (such as low click-through rate or blacklisting of the recommended commodities), then regenerate the guiding relationship; if the user feedback is good (such as high click-through rate or browsing rate of the recommended commodities), then insert information such as advertisements based on these commodities according to the marketing strategy to generate the final recommendation result. Finally, the user operates on the recommended commodities (such as clicking, purchasing, evaluating, collecting, and blacklisting, etc.) to form new operation behavior data, which will be used to update the database and execute subsequent recommendations.

[0076] The present invention relates to guiding relationship modeling. Refer to Figure 3 , and the defined various different guiding relationships are as follows (by default, there are relatively frequent interaction relationships among individuals in the same social circle, and the double-headed arrow represents a strong connection relationship): ① The intra-layer strong connection guiding relationship established through the interest relationship in the same social circle: If there is at least one strong connection relationship channel that does not cross the social circle between an individual and other individuals in the same social circle, there is a probability that the two parties form this guiding relationship.

[0077] ② is the within-layer weak connection-oriented relationship established through cross-social-layer interest relationships: If there is at least one strong connection relationship channel spanning social layers between an individual and other individuals in the same social layer, there is a probability that the two parties will form this oriented relationship.

[0078] ③ is the between-layer strong connection-oriented relationship established through cross-social-layer interest relationships: If there is at least one strong connection relationship channel spanning social layers between an individual and other individuals in different social layers, there is a probability that the two parties will form this oriented relationship.

[0079] ④ is the between-layer weak connection-oriented relationship established through same-social-layer interest relationships: If there are no or only a small number of strong connection relationships (usually cold start problems) between an individual and other individuals in the same social layer, but there are strong connection relationship channels spanning social layers between other individuals and a large number of individuals in another social layer, there is a probability that this individual and the individuals in that other social layer will form this oriented relationship.

[0080] ⑤ is the within-layer weak connection-oriented relationship established through same-social-layer interest relationships: If there are no or only a small number of strong connection relationships (usually cold start problems) between an individual and other individuals in the same social layer, there is a probability that the two parties will form this oriented relationship.

[0081] ⑥ is the within-layer strong connection-oriented relationship established through cross-social-layer interest relationships: If there is at least one strong connection relationship channel that does not span social layers between an individual and other individuals in the same social layer, and both parties have strong connection relationships with the same individual in another social layer or two individuals connected by at least one strong connection relationship channel that does not span social layers, there is a probability that the two parties will form this oriented relationship.

[0082] ⑦ is the between-layer weak connection-oriented relationship established through cross-social-layer interest relationships: If there is at least one strong connection relationship channel spanning social layers that needs to be continued by other strong connection-oriented relationships between an individual and other individuals in different social layers, there is a probability that the two parties will form this oriented relationship.

[0083] ⑧ is the between-layer strong connection-oriented relationship established through same-social-layer interest relationships: If there are a large number of individuals in a social layer who have strong connection relationship channels with a large number of individuals in another social layer, there is a probability that the individuals in these two social layers will form this oriented relationship.

[0084] The differences in social layers affect the user's interests and the strength of connections, which in turn affect the degree of freshness of the recommended content for the user, from shallow to deep.

[0085] Among them, the so-called "with probability" means that when there is a certain interest relationship between users, if only one of the proposed guiding relationship establishment conditions is met, the probability of forming this guiding relationship is 100%; when there is more than one interest relationship, according to the actual model training situation, weights are assigned to each guiding relationship.

[0086] The result of the guiding relationship modeling is the result obtained by calculating based on the attention mechanism according to the importance level (weight). Each relationship occupies a certain proportion and plays a corresponding role in it. Through this guiding relationship modeling, it is possible to recommend the content that a user is interested in to another user due to accuracy or novelty through the guiding relationship. Whether these two users are acquaintances with a considerable number of common or similar interests in the same social circle or complete strangers from two completely different social circles, the above effects can be achieved, and item recommendations that take into account both accuracy and novelty can be completed.

[0087] The following is a user emotion judgment criterion for reference (i.e., similar effects can be achieved by modifying each threshold): Very interested: If among the user's recent 100 browsing records, the browsing records of a certain type of content account for more than 70, it is determined that the user is very interested in this type of content. For users with less than 100 but more than 30 browsing records, according to the proportion of browsing types, the content types accounting for more than 70% are judged as the content that the user is very interested in. For users with less than 30 browsing records, such a judgment is not made.

[0088] Rather interested: If among the user's recent 100 browsing records, the browsing records of a certain type of content account for more than 30, it is judged that the user is rather interested in this type of content. For users with less than 100 but more than 10 browsing records, according to the proportion of browsing types, the content types accounting for more than 30% are judged as the content that the user is rather interested in. For users with less than 10 browsing records, such a judgment is not made.

[0089] Rather novel: If among the user's recent 100 browsing records, the browsing records of a certain type of content are less than 20, but according to the emotional judgment of similar users associated with the social circle, the users have a rather interested or higher emotional judgment on this type of content, it is judged that the user is rather novel about this type of content. For users with less than 100 browsing records, if the number of times the user browses this type of content is less than 20% of the total browsing times and there are browsing records of this type of content in the recent 5 browsing records, it is judged that the user is rather novel about this type of content. For users with less than 10 browsing records, such a judgment is not made.

[0090] Very novel: If among the user's last 10 browsing records, the browsing records of a certain type of content account for more than 3, but among the user's last 100 browsing records, the browsing records of this type account for less than 30, then it is determined that the user is very novel about this type of content. For users with less than 100 browsing records, such a judgment is not made.

[0091] It should be noted that the above sentiment judgment is not an independent relationship and is only used to explain the effect of the guiding relationship proposed in the present invention, which has nothing to do with the recommendation process.

[0092] Through meticulous utilization of the self-attention mechanism, the present invention quantifies the importance of combinations of strong and weak interpersonal connection relationships in different social circles and models these social relationships. This process reveals the changing trend of user interests in the nature of multi-level social circles. Among them, based on the rich semantic relationships of the knowledge graph, the present invention embeds local neighbor information into entities through advanced convolutional aggregation operations, realizing the conversion from individual users to user groups; and realizes neighborhood selection according to the attention weights, effectively avoiding the problem of increased model complexity caused by too large a neighborhood, expanding the space of specific parameters of the artificial control system (artificially regulating the model calculation amount, artificially regulating the actual utilization rate of information, and artificially balancing the novelty and accuracy of recommended content, etc.), and effectively improving the practicality and generality of the model.

[0093] Based on the same concept, the present invention also provides an item recommendation system based on social guiding relationships, including a collection unit, an embedding unit, a construction unit, an establishment unit, and a recommendation unit.

[0094] The collection unit is used to collect item data and user data of multiple users and construct a knowledge graph based on the item data and user data.

[0095] The embedding unit is used to embed the knowledge graph into a graph convolutional network, where multiple users are used as the first nodes and multiple items are used as the second nodes.

[0096] The construction unit is used to obtain the similarity score between two corresponding first nodes based on the social guiding relationship between any two users; assign guiding relationship weights to different first nodes through the attention mechanism, and obtain the guiding relationship value between two first nodes based on the similarity score and the guiding relationship weight; establish a guiding relationship network based on the guiding relationship value between any two first nodes.

[0097] The establishment unit is used to, in the guiding relationship network, take any one of the first nodes as the target node and another first node that has no guiding relationship with the target node as the recommended node, establish multiple connection channels between the target node and the recommended node, and each connection channel is connected by multiple first nodes for which all guiding relationships are enabled; when the guiding relationship value between two first nodes is greater than the corresponding guiding relationship threshold, this guiding relationship is enabled.

[0098] The recommendation unit is used to aggregate the node information of the second node connected to the target node to the target node, and perform item recommendation on the recommendation node through multiple connection channels.

[0099] Based on the same concept, referring to Figure 2 and Figure 4 , the recommendation system of the present invention further includes a client and a server. Among them, the client includes a client - operation module, a client - processing module, a client - sending module, a client - receiving module, a client - verification module, and a client - display module. The server includes a server - receiving module, a server - setting module, a server - verification module, a server - recommendation module, a server - update module, and a server - sending module.

[0100] The client - operation module is used to support user operation behaviors, such as liking, favoriting, commenting, and searching. Compared with the rating data of specific books or movies provided by the MovieLens dataset and the Jester dataset, the behavioral data used in this module is more diverse, more in line with user viewing habits, and can mine more user relationships, which helps to improve the utilization rate of user behavioral data.

[0101] The client - processing module is used to perform three data - processing stages of extraction, transformation, and loading on the user operation behavior, and package the user operation behavior in the operation module into text data that can be accepted and used by the server.

[0102] The client - sending module is used to send the user operation behavior data processed by the processing module to the server through the network.

[0103] The server - receiving module is used for the server to receive the packaged user operation behavior text dataset through the network.

[0104] The server - setting module is used to switch or modify the recommendation model parameters (including the threshold for guiding relationship modeling and the threshold for the length of strong and weak connection channels) according to changes in data such as user behavior, item list, or the business purpose of the system user. This module is only enabled when the system is updated. This module is similar to the online - layer server - side processing technology mentioned above. Using online data - processing technology can ensure the pertinence and timeliness of the recommendation model.

[0105] The server - side verification module is used to check whether there are errors in the recommended model after switching and modification, and is only enabled during system updates. This module is the verification module of the setting module, which can perform recommendation verification based on the data stored in the system. The verification process includes: extracting the user's operation behavior data, checking whether the content that has been repeatedly viewed by the user in a short period (such as viewed three times within the last week), or the content that has been blacklisted by the user, or the content that has been repeatedly recommended by the system in a short period (such as recommended ten times within the last week) is included in the blacklist for the next recommendation of the recommended model, that is, the above - mentioned content will not be recommended again, and all generated recommendation results are composed of relationship weights that meet the requirements of the setting module. As a guarantee for the correctness and rationality of the recommended model before going online, this module can avoid errors in the updated model.

[0106] The server - side recommendation module is used to call the recommended model to complete personalized recommendation of information or items. After the server - side receives the user behavior data and when the system does not need to be updated, the data is directly transmitted to this module for online processing of the user operation data, and several items that the user may be most interested in are selected on the server - side, and the recommendation reasons for the corresponding items are generated according to the composition of the relationship weights.

[0107] The server - side update module is used to pack the recommended items and recommendation reasons and other recommendation results selected by the recommendation module into data that can be accepted and used by the client through data processing, and update information such as user relationships in the database. After the recommendation module makes an effective recommendation, new relationship network information will be generated. It is possible to judge whether the previous recommendation is effective by the user's operation behavior on the previous recommendation result. If it is effective, the relationship network information generated by the previous recommendation can be updated.

[0108] The server - side sending module is used to send the packed recommended items and recommendation reasons and other recommendation result data to the client through the network.

[0109] The client - side receiving module is used to receive the packed recommended items and recommendation reasons and other recommended content through the network.

[0110] The client - side verification module is used to use a fixed format to judge whether there are situations such as packet loss or format chaos in the recommendation result data, and then by trial - pushing part of the recommendation results, through operation behavior data such as user clicks and browsing feedback, to check whether the recommendation results meet the requirements, further judge whether other advertisements need to be inserted according to the marketing strategy, etc., and generate the final complete recommendation results. Existing recommendation systems usually do not have the "trial recommendation" link.

[0111] The client - side display module is used to display the final recommendation results to the user and provide subsequent operations such as liking and viewing for the user.

[0112] The present invention proposes a graph convolutional neural network item recommendation algorithm based on a knowledge graph and a social strong and weak connection attention mechanism. Compared with traditional algorithms, it can extract high-order guiding relationship features of users in the same social circle at the strong connection level, extract high-order guiding relationship features of users in different social circles at the weak connection level, and provide accurate recommendations and novel recommendations to users based on strong connections and weak connections respectively, taking into account the emotional preferences of users and the diversity of recommendations.

[0113] The system classifies and collects according to the browsing preferences of users, forms a social graph in the background, and further differentiates according to the hobby characteristics of each user's social graph. After differentiation, users will form strong and weak relationship connections among themselves according to different favorite characteristics, and new relationships among users will be embedded according to the connection weights between users in different circles.

[0114] Through the guiding relationship network, the background will push relevant content to users, and then continuously refine the user portrait according to the tendency of users to click, enrich the user's social circle, deeply cross the social connection relationships among users, and conduct in-depth recommendations to increase the diversity of recommended content.

[0115] In the way of forming a user social circle by information differentiation, the user preset operations include one or more of the following operations: multiple browsing, repeatedly browsing the same type of web pages, commenting operations, sharing operations, and follow lists, etc.

[0116] The algorithm provided by the present invention is not inherently related to any specific computer, virtual system or other device. In addition, the embodiments of the present invention are not directed to any specific programming language. It should be understood that the content of the present invention described herein can be implemented using various programming languages, and the description of a specific language above is for better demonstrating the system functions.

[0117] Appendix Figure 2Depicts a schematic diagram of the structure of a recommendation system with hardware such as a power supply component. In the figure, the core function of the communication interface is to collect users' behavioral data on different platforms through input and output devices, including clickstream, browsing duration, transaction data, evaluation and feedback information, etc., and to send the recommendation results generated by the recommendation system to the users' clients. The above-mentioned behavioral data also includes social interaction data, such as the content posted by users, interaction frequency (liking, commenting, or favoriting, etc.), topics participated in discussions, and structural dynamics in the social network (being friends with others, blocking others, etc.). The power supply component is involved in monitoring the usage patterns of other hardware and power consumption to ensure the normal operation of the system. The data preprocessing sub-module comes into play after data collection and is used for data purification. The normalization process converts the data into a format more suitable for analysis, and formatting further ensures the consistency of the data between different sources. For example, the data presentation forms on different social media platforms are uniformly processed to prepare for cross-platform analysis.

[0118] Appendix Figure 1 Details the recommendation process in the recommendation system. This recommendation process is based on the construction of the guiding relationship in the knowledge graph and uses graph analysis technology to construct the social network graph of users, establishing the complex connections between users and other individuals. The position of the user in the graph is determined by calculating network metrics such as the degree, betweenness, and closeness of the nodes. These metrics reflect the importance and influence of the user in their social network. After that, any process that uses social network analysis (such as recall, rough ranking, fine ranking, and re-ranking) will identify the complex guiding relationship network components of the user, including defining the user's main social circle, secondary circles, and various marginal connections, etc. Through in-depth analysis of the personal social graph, this process can also identify and quantify the user's social influence, which can help the system understand what kind of content the user may be more sensitive or interested in and predict what kind of content they are expecting.

[0119] Obviously, Figure 1 Belongs to Figure 4 The sub-content of. For example, Figure 1 Both the "setting of the guiding relationship modeling threshold" and the "setting of the strong and weak connection channel length threshold" in Figure 4 Belong to the "server-side - setting module" in Figure 1 The "update" operation in Figure 4 Belongs to the "server-side - update module" in Figure 1 The "knowledge graph construction", "guiding relationship modeling", "recall", "rough ranking", "fine ranking", and "re-ranking", etc. in Figure 4 Belong to the "server-side - recommendation module" in Figure 1 The "user operations" in Figure 4 Are included in the "client-side - operation module" in Figure 1 The "operation data" inFigure 4 The "client - processing module" in Figure 1 The "recommended result" in Figure 4 is the data sent from the server - side to the client - side in

[0120] Meanwhile, all the above - mentioned server - side processing procedures are implemented in Figure 2 the "program" and "processor" in ; all the server - side data involved are stored in Figure 2 the "memory" in ; the data transmission process between the server - side and the client - side is all implemented through Figure 2 the "communication interface" in ; the sequence of operations of each server - side module is implemented by Figure 2 the "controller" in ; all the internal data communication of the server - side is based on Figure 2 the "communication bus" in ; all the hardware operations of the server - side are based on Figure 2 the support of the "power supply component" in

[0121] Due to the high uncertainty of the user device configuration of the client - side hardware, the present invention does not need to make special restrictions on it.

[0122] Specifically, Figure 4 This invention focuses on the operating mechanisms of the client - side and the server - side, involving how to integrate the behaviors and roles of users in different social circles into the recommendation system. Further, after generating the recommended results and feeding them back to the users, the database is updated to generate new social relationship information for the next recommendation. This operating mechanism will analyze the types and frequencies of interactions of users in these circles. This process specifically considers the social identity theory, identifies the behaviors of users in different circles as specific social identities, and analyzes how these identities interact with each other and how they affect the preferences and decisions of users. In addition, the process of social circle integration will also evaluate the participation and influence of users in their social circles, including quantifying the reactions of users to events inside and outside the social circles, such as comments on news events or adoption of popular trends. Through in - depth analysis of these behaviors, this process can identify which new content users may tend to accept. After completing a recommendation, the system will judge the effect of the previous recommendation (such as whether it is liked, watched, collected, etc. by users), select valid data to update the database, update the guiding relationships (involving the establishment of some new guiding relationship networks and the decrease or increase in the proportion of some old guiding relationship networks), and use them for the next recommendation.

[0123] Obviously, Figure 4Each module in it not only works independently but also is closely connected in the data stream, jointly constituting a comprehensive recommendation system framework that combines the client and the server. The outputs of these modules are integrated by carefully designed algorithms to produce a list of recommendation results that takes into account both accuracy and novelty. This ensures that the system can not only reflect the current behaviors and preferences of users but also predict and promote their exploration of potential interests.

[0124] In addition, the reduced user data exposure also reduces the risk of user privacy data leakage. Under the Figure 5 framework attached, a movie recommendation method combining deep relationships in social networks proposed by the present invention is explained.

[0125] Users' behaviors in social networks are not isolated but exist as interconnected nodes in multi-level social circles. Users' preferences and choices are the result of the combined effects of direct and indirect influencing factors in their social environment. Traditional recommendation systems often fail to fully utilize these social elements, resulting in limitations in the accuracy and diversity of recommended content. The innovation of the technology of the present invention lies in integrating users' social dimension information and enhancing the performance of personalized recommendation systems through precise social network analysis.

[0126] The starting module of the system is the data collector, which is responsible for collecting users' behavioral data on multiple social platforms, such as browsing history, purchase records, evaluation feedback, and social interaction data, etc. The obtained data stream is screened by the preprocessing module, which includes steps of data cleaning, formatting, and standardization to ensure the quality and efficiency of subsequent module processing.

[0127] After preprocessing, the data is sent to the core of the recommendation system - the recommendation engine. This engine combines cutting-edge data mining algorithms, including but not limited to machine learning and deep learning technologies, to finely analyze the neighborhood information of individual users. The analysis of neighborhood information involves the relative position of users in their social networks and the strength of relationships with other nodes, thus determining the degree of personalization of recommendations.

[0128] Subsequently, the system enters the social circle integration stage, and a specially designed algorithm module analyzes and reveals users' behavioral patterns and roles in different social environments. This module not only analyzes the interactions between users and their direct contacts, such as relationships between friends, family members, or colleagues, but also extends to the position of users in a broader social graph, such as the connections between users and others with the same concerned topics, and the indirect connections established through third-party introductions.

[0129] All these social network analysis results from micro to macro are integrated and used in the recommendation generation module. This module utilizes complex algorithms, combines the user's personal preferences and their position and role in the social network, and outputs a customized recommendation list. This list is optimized by algorithms, not only reflecting the user's immediate interests but also encouraging the user to explore novel content, increasing the possibility for the user to discover unknown areas.

[0130] Finally, the system will update the database again based on the recommendation effect to continuously improve the practicality of the recommendation.

[0131] During the entire recommendation process, the user experience module monitors and analyzes the user's interaction with the recommended content, adjusts the recommendation strategy through this feedback, and ensures the relevance and attractiveness of the content. The system includes an adaptive learning mechanism that will adjust the recommendation algorithm in real time according to the evolution of the user's behavior to ensure the continuous adaptability of the recommended content.

[0132] Refer to Figure 6 , the present invention can generate a large number of guiding relationships based on the relationships (purchase relationships) of a small number of existing users (User 1 - User 8) and commodities (Item 1 - Item 12) to achieve associated data mining, and then complete the commodity recommendation for new users having strong connection relationships with User 4 and User 5.

[0133] Refer to Figure 7 , when User 1 obtains the recommendation of Item 1 through the guiding relationship composed of User 2, User 3, and User 4, the present invention can clearly know the reason why Item 1 is recommended to User 1 (because there is a valuable social network between User 4 and User 1, see Figure 3 ). Figure 7 The purchase relationship in refers to the purchase behavior of User 4 for Item 1, providing a direct interest signal for the subsequent recommendation results. The recommendation path represents the information dissemination path of the system based on the guiding relationship between users, transmitting the information of Item 1 from the purchaser (User 4) via User 3 and User 2 to the user (User 1) who is to receive the recommendation step by step. The traceability path represents the reverse recommendation path that traces back along the recommendation path, and thus can trace back from the recommendation result of User 1 to the initial purchase behavior (User 4's purchase of Item 1) and the social relationship behind it.

[0134] Refer to Figure 8 , since the interests of individuals in the same social circle (Social Circle 1 or Social Circle 2) are more likely to be similar, accurate recommendation tends to recommend commodities that other individuals in the same social circle are interested in to the target user. Since the interests of individuals in different social circles are more likely to be dissimilar, novel recommendation tends to recommend commodities that other individuals in different social circles are interested in to the target user. The present invention can take into account both the effects of such accurate and novel recommendations.

[0135] Refer toFigure 9 and Figure 10 where C refers to a recommendation mode that only focuses on accuracy, concentrating on recommending a small range of content, which is prone to forming an information cocoon. D refers to a recommendation mode that takes both accuracy and novelty into account, creating new interest points while protecting the original interests. The present invention can take into account both the accuracy and novelty of the recommendation effect. The recommendation accuracy referred to in the present invention: The recommendation model recommends products that the target user is interested in by focusing on the high-interest areas (with a high interest index, corresponding to a high weight of the guiding channel, which are the main interest peaks) in specific categories (such as books or movies, etc.). For example, Figure 10 the high-interest area corresponding to the "movie" category in

[0136] shows that the target user has a relatively high acceptance of such products, meeting the accuracy standard of the recommendation. Figure 10 The novelty of the recommendation referred to in the present invention: The recommendation model can identify products that the target user has not been exposed to but has potential interest in. Although these products are not within the scope of the user's known interests, they are new interest points mined based on the guiding relationship (such as

[0137] the interest peak of the "documentary" category in

[0138] Examples Forty volunteers (20 teenagers and 20 middle-aged and elderly people, with a male-to-female ratio of 1:1) were invited to conduct an AB test on the recommendation system (System A) proposed based on the present invention and another simple recommendation system (System B) constructed based on CGAT (Contextualized Graph Attention Network, a new type of recommendation algorithm). Among them, the Movie-lens-20M dataset was used for model training.

[0139] During the test, each volunteer would perform 10 recommendation operations (randomly 5 times using System A for recommendation and the other 5 times using System B for recommendation, and the corresponding recommendation results were generated based on the simulated behaviors of the volunteers such as purchasing or reading books and watching movies). After each recommendation operation, the volunteer needed to fill in an evaluation of 0 - 5 points for the current recommendation result. Finally, by performing an independent score normalization operation on each volunteer group, it was determined which recommendation system the volunteers generally preferred.

[0140] Table 1 Comparison of AB test results

[0141] The original scoring data in Table 1 shows that male and female volunteers of teenagers and middle-aged and elderly people always rate the recommendation results of System A higher than those of System B. The average score increase per single recommendation is as high as 8.78%, which is equivalent to 0.5 points on a 0-5 point scale, and the improvement effect is relatively obvious.

[0142] System A has the most significant scoring improvement effect among middle-aged and elderly female volunteers, reaching 12.02%, exceeding 1 point on a 0-5 point scale, but the scoring improvement effect among teenage male volunteers is the weakest, only 6.99%. However, the 12.02% scoring improvement effect is significantly larger compared to the other increases of 6.99%, 8.33% and 7.78%, and there may be certain accidental situations.

[0143] In addition, the original scoring data in Table 1 also shows that male volunteers generally rate the recommendation results higher than female volunteers, and there are two phenomena: (1)Female volunteers of the same age group rate the recommendation results of System A close to the recommendation results of System B by male volunteers.

[0144] (2)Middle-aged and elderly volunteers always rate the recommendation results of System A and B significantly higher than those of teenage volunteers for System A and B.

[0145] To minimize the differences in scoring criteria among different volunteers as much as possible, in this embodiment, the original scoring data is normalized in groups.

[0146] For the normalized scoring data, the normalization weakens the influence of high scores on the total score, and ultimately results in the total score being lower than the original total score.

[0147] In addition, normalization also weakens the differences in low scores, making the total score attenuation amplitude of groups with fewer high-score ratings larger, especially the situation where the recommendation results of System A by teenage female volunteers are lower than those of System B. At the same time, normalization retains some special phenomena, such as the ratings of System A by female volunteers of the same age group being close to the ratings of System B by male volunteers, and the ratings of System A and B by middle-aged and elderly volunteers being significantly higher than those of teenage volunteers.

[0148] To sum up, in the AB test, System A has generally improved the user feedback scoring compared to System B, showing a relatively obvious optimization effect of the recommendation results. Therefore, it can be considered that the actual recommendation effect of System A is better than that of System B, and the practical application superiority of the present invention is verified.

[0149] The present invention establishes a recommendation system that can adapt to new users and dynamic changes by deeply mining users' complex social relationships. This system can not only alleviate the sparsity and cold start problems in traditional recommendation systems, but also enhance the system's dynamic adaptability, providing accurate and novel recommended items. By constructing a complete recommendation explanation path, the interpretability of the recommendation system is enhanced, and users' trust is improved. At the same time, the system can balance the accuracy and novelty of recommended items, avoid the "information cocoon" problem, and broaden users' horizons. In addition, by deeply mining the relationship value between individuals and social groups, the information utilization rate is improved, and the "long-tail effect" of information is alleviated. The system can also ensure the depth, breadth, and timeliness of recommendations through dynamic social relationship modeling, improving user stickiness and retention rate. Finally, the present invention realizes a virtuous cycle of "making the best use of things" and "putting people to good use" based on a complex relationship network, improving the overall performance of the recommendation system and the user experience.

[0150] In the recommendation process, the present invention mainly models the guiding relationship instead of directly using users' original information, thereby reducing the frequency of using user data and the links for the outside world to access user data, and significantly reducing the risk of leakage of users' private data. By artificially regulating the model calculation amount, the actual information utilization rate, and the balance between the novelty and accuracy of recommended content, the present invention can achieve various specific recommendation purposes while protecting users' privacy, enhancing the versatility and practicality of the system.

[0151] The system effectively processes data sparsity, alleviates the cold start problem, and enhances the dynamic adaptability through graph embedding technology and guiding relationship network. Through a multi-level social network model, it deeply mines users' behavior patterns and influence, identifies potential interests, and improves the exposure rate of long-tail content. At the same time, the system adopts a real-time data stream recommendation algorithm and a hybrid recommendation model to dynamically adjust the guiding relationship network, combining users' personal portraits, historical behavior data, and social network data, as well as item content features, category labels, and user feedback, to generate a comprehensive recommendation score, ensuring the depth, breadth, and timeliness of recommendations. For new users, the system can generate potential interest points based on a small number of interaction records and construct a refined user portrait; for old users, it controls the number of guiding relationships to avoid an explosive growth in the number of relationships. In addition, the system can also provide a complete recommendation explanation path, enhancing user trust, and improving user stickiness through a novelty-accuracy trade-off mechanism.

[0152] The system provides a parameter adjustment mechanism that allows operators to dynamically adjust the model's computational volume and information utilization rate according to performance requirements, resource limitations, marketing strategies, etc., to adapt to different application scenarios and user needs. Moreover, the system allows operators to dynamically adjust the novelty and accuracy of the recommended content according to the characteristics of the user group and market trends. Since the system mainly realizes recommendations by modeling the guiding relationship, the original information of users is only used for modeling and does not directly participate in the recommendation process, which greatly reduces the risk of leakage of user privacy data.

[0153] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concept. Therefore, the appended claims are intended to be construed as including the preferred embodiments as well as all changes and modifications that fall within the scope of the present invention.

[0154] Obviously, those skilled in the art can make various changes and deformations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and deformations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and deformations.

Claims

1. An item recommendation method based on social-oriented relationships, characterized in that It includes the following steps: Collect the item data and user data of multiple users, and construct a knowledge graph based on the item data and user data; Embed the knowledge graph into a graph convolutional network, where multiple users serve as the first nodes and multiple items serve as the second nodes; Obtain the similarity score between two corresponding first nodes based on the social orientation relationship between any two users; assign the orientation relationship weight to different first nodes through the attention mechanism, and obtain the orientation relationship value between two first nodes based on the similarity score and the orientation relationship weight; establish an orientation relationship network based on the orientation relationship values between any two first nodes; In the orientation relationship network, take any one first node as the target node and another first node without an orientation relationship with the target node as the recommended node, and establish multiple connection channels between the target node and the recommended node. Each connection channel is obtained by connecting multiple first nodes with all enabled orientation relationships; when the orientation relationship value between two first nodes is greater than the corresponding orientation relationship threshold, the orientation relationship is enabled; Aggregate the node information of the second nodes connected to the target node to the target node, and recommend items to the recommended node through multiple connection channels.

2. The method for recommending items based on social orientation relationships according to claim 1, wherein The similarity score is specifically as follows: ； wherein, is the similarity score between the first nodes u and v ; is the trainable weight is the social orientation relationship between two first nodes is the node feature of the first node v ; T is the matrix transpose operation 3. The item recommendation method based on social orientation relationship according to claim 2, characterized in that The step of assigning the orientation relationship weight to different first nodes through the attention mechanism and obtaining the orientation relationship value between two first nodes based on the similarity score and the orientation relationship weight specifically includes the following steps: Assign the orientation relationship weight to different first nodes through the attention mechanism to obtain the orientation relationship embedding; Combine the similarity score with the orientation relationship embedding to obtain the orientation relationship value between two first nodes.

4. The article recommendation method based on social orientation relationship according to claim 3, characterized in that The orientation relationship embedding is specifically as follows: ； Wherein, is the guidance relationship embedding, is the embedding vector of the relationship , is the attention score, R is the set of all relationships r .

5. The article recommendation method based on social orientation relationship according to claim 4, wherein The orientation relationship value is specifically as follows: ； In the formula, is the guiding relationship value, is the trainable weight parameter.

6. The method for recommending items based on social orientation relationships according to claim 1, characterized in that The step of recommending items to the recommended node through multiple connection channels specifically includes the following steps: Train the orientation relationship network to obtain multiple trained trainable weights, and generate multiple first connection channels according to the multiple trained trainable weights; Set the first channel length threshold, retain multiple first connection channels smaller than the first channel length threshold to obtain multiple second connection channels, and recommend items to the recommended node through the multiple second connection channels; Set the second channel length threshold, retain multiple second connection channels smaller than the second channel length threshold to obtain multiple third connection channels, and recommend items to the recommended node through the multiple third connection channels; where the first channel length threshold is smaller than the second channel length threshold.

7. The item recommendation method based on social orientation relationship as described in claim 1, characterized in that The user data includes browsing duration, user evaluation, and questionnaire feedback, and the item data includes the click volume, transaction volume, item category, and item attributes of the item.

8. An item recommendation system based on social-oriented relationships, characterized in that It includes: A collection unit for collecting the item data and user data of multiple users and constructing a knowledge graph based on the item data and user data; An embedding unit for embedding the knowledge graph into a graph convolutional network, where multiple users serve as the first nodes and multiple items serve as the second nodes; A building unit is used to obtain the similarity score between two corresponding first nodes based on the social orientation relationship between any two users; allocate the orientation relationship weights to different first nodes through the attention mechanism, and obtain the orientation relationship value between the two first nodes based on the similarity score and the orientation relationship weights; establish an orientation relationship network based on the orientation relationship values between any two first nodes. An establishment unit is used to, in the orientation relationship network, take any one first node as the target node and another first node that has no orientation relationship with the target node as the recommended node, and establish multiple connection channels between the target node and the recommended node. Each connection channel is obtained by connecting multiple first nodes with all enabled orientation relationships; when the orientation relationship value between two first nodes is greater than the corresponding orientation relationship threshold, the orientation relationship is enabled. A recommendation unit is used to aggregate the node information of the second nodes connected to the target node to the target node, and perform item recommendations for the recommended nodes through multiple connection channels.

Images