An item recommendation method and system based on social-oriented relationships

Abstract

The present invention discloses a method and system for recommending items based on social-oriented relationships, which relates to the technical field of item recommendation and includes the following steps: embedding a knowledge graph into a graph convolutional network, obtaining a social-oriented relationship between any two first nodes, and obtaining a similarity score between the two first nodes based on the social-oriented relationship; assigning oriented relationship weights to different first nodes through an attention mechanism to obtain oriented relationship embeddings, combining the similarity scores with the oriented relationship embeddings to obtain oriented relationship values ​​between the two first nodes; establishing a oriented relationship network based on the oriented relationship values ​​between any two first nodes; in the oriented relationship network, aggregating node information of second nodes connected to a target node to the target node, and recommending items to the recommended nodes through multiple connection channels. The present invention significantly improves the accuracy and novelty of recommendations, avoids the problem of information cocoons, and broadens the user's horizons.

Description

Technical Field

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

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

[0003] The effectiveness of these recommendation systems depends largely on their ability to process and interpret complex user networks, which often exist in the form of higher-order relationship networks in modern society. Knowledge graphs, as a structured data form, with their rich relationships and attributes between entities, show great potential for improving the accuracy, interpretability, and personalization of recommendation systems. Therefore, how to fully utilize this rich information resource, especially how to reveal and utilize the indirect or implicit social relationships in knowledge graphs, has always been a difficult problem in recommender system research.

[0004] Traditional item recommendation systems, such as those based on collaborative filtering and content recommendation, often struggle to leverage and understand this high-order relational network, especially when considering the rich semantic information in knowledge graphs. The current recommendation system process generally includes data collection and preprocessing, recall, sorting, re-ranking and filtering, and recommendation result generation and evaluation. However, a significant amount of computation is consumed in traditional recommendation systems during the user data collection, sorting, re-ranking, and filtering stages. The resulting recommendations are often overly correlated with the user, easily creating information cocoons. Summary of the Invention

[0005] Based on the defects of the above-mentioned existing technologies, the present invention provides an item recommendation method and system based on social-oriented relationships, which solves the problem that a large amount of computing power is consumed in the traditional recommendation system in the stages of collecting user data, sorting, rearranging and filtering. The recommendation results formed in this way are usually too strongly correlated with the user itself, which can easily form the problem of information cocoon.

[0006] The present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a method for recommending items based on social-oriented relationships, comprising the following steps:

[0008] Collect item data and user data from multiple users, and build a knowledge graph based on the item data and user data;

[0009] Embed the knowledge graph into a graph convolutional network, where multiple users serve as first nodes and multiple items serve as second nodes.

[0010] Based on the social orientation relationship between any two users, a similarity score between the corresponding two first nodes is obtained; through the attention mechanism, orientation relationship weights are assigned to different first nodes, and the orientation relationship value between the two first nodes is obtained based on the similarity score and the orientation relationship weight; based on the orientation relationship value between any two first nodes, an orientation relationship network is established;

[0011] In a guiding relationship network, any first node is used as a target node, and another first node that has no guiding relationship with the target node is used as a recommended node. Multiple connection channels are established between the target node and the recommended node, each connection channel being formed by connecting multiple first nodes with all guiding relationships enabled. When the guiding relationship value between the two first nodes is greater than the corresponding guiding relationship threshold, the guiding relationship is enabled.

[0012] The node information of the second node connected to the target node is aggregated to the target node, and items are recommended to the recommendation node through multiple connection channels.

[0013] Preferably, the similarity score is specifically as follows:

[0014] ;

[0015] Where, The first node u and v The similarity score between are trainable weights, is the social-oriented relationship between two first nodes, The first node v The node characteristics of T is the matrix transpose operation.

[0016] Preferably, the method of assigning guiding relationship weights to different first nodes through an attention mechanism and obtaining a guiding relationship value between two first nodes based on a similarity score and the guiding relationship weight specifically includes the following steps:

[0017] The attention mechanism is used to assign guidance relation weights to different first nodes to obtain guidance relation embeddings;

[0018] The similarity score is combined with the guiding relation embedding to obtain the guiding relation value between the two first nodes.

[0019] Preferably, the guiding relationship is embedded as follows:

[0020] ;

[0021] Where, To guide the relationship embedding, For the relationship The embedding vector of is the attention score, R All relationships r A collection of .

[0022] Preferably, the guiding relationship value is specifically as follows:

[0023] ;

[0024] Where, is the oriented relationship value, is the trainable weight parameter, Oriented relational embedding.

[0025] Preferably, the step of recommending items to the recommendation nodes through multiple connection channels specifically includes the following steps:

[0026] Training the guided relationship network to obtain a plurality of trained trainable weights, and generating a plurality of first connection channels according to the plurality of trained trainable weights;

[0027] A first channel length threshold is set, multiple first connection channels with lengths shorter than the first channel length threshold are retained to obtain multiple second connection channels, and items are recommended to the recommendation node through the multiple second connection channels;

[0028] A second channel length threshold is set, and multiple second connection channels that are smaller than the second channel length threshold are retained to obtain multiple third connection channels, and items are recommended to the recommendation nodes through the multiple third connection channels; wherein the first channel length threshold is smaller than the second channel length threshold.

[0029] Preferably, the user data includes browsing time, user evaluation and questionnaire feedback, and the item data includes item or item click volume, transaction volume, item category and item attributes.

[0030] In a second aspect, the present invention provides an item recommendation system based on social-oriented relationships, comprising:

[0031] A collection unit, configured to collect item data and user data of multiple users and construct a knowledge graph based on the item data and user data;

[0032] An embedding unit, configured to embed the knowledge graph into a graph convolutional network, wherein a plurality of users serve as first nodes and a plurality of items serve as second nodes;

[0033] A construction unit is configured to obtain a similarity score between two corresponding first nodes based on a social guiding relationship between any two users; assign guiding relationship weights to different first nodes through an attention mechanism, and obtain a guiding relationship value between the two first nodes based on the similarity score and the guiding relationship weight; and establish a guiding relationship network based on the guiding relationship value between any two first nodes;

[0034] an establishing unit configured to establish, in a guiding relationship network, a plurality of connection channels between the target node and the recommended node, with any first node as the target node and another first node that has no guiding relationship with the target node as the recommended node, each connection channel being formed by connecting a plurality of first nodes with all guiding relationships enabled; and when a guiding relationship value between two first nodes is greater than a corresponding guiding relationship threshold, the guiding relationship is enabled;

[0035] The recommendation unit is used to aggregate node information of the second node connected to the target node to the target node, and recommend items to the recommendation node through multiple connection channels.

[0036] Compared with the prior art, the at least one technical solution adopted by the present invention can achieve the following beneficial effects:

[0037] 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 guidance relationship between any two first nodes is obtained, and the interactive relationship between users in the knowledge graph is deeply mined to obtain the corresponding similarity score. At the same time, the attention mechanism is used to assign guidance relationship weights to different first nodes, and then the guidance relationship embedding is obtained. Based on the attention mechanism, relationship weights are assigned to different users, and the similarity score is combined with the guidance relationship embedding to obtain the guidance relationship value between the two first nodes; a guidance relationship network is established based on the guidance 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 to establish a guidance relationship network for users.

[0038] In the context of a guiding relationship network, we propose the concept of a guiding relationship threshold, which controls the priority of using that relationship. Compared to existing recommendation methods, this technology offers both breadth and depth. By enabling guiding relationships, we establish multiple connection channels and build a complete recommendation path, significantly improving the accuracy and novelty of recommendations, avoiding the information cocoon problem, and broadening users' horizons. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 A flowchart of a method for recommending items based on social-oriented relationships according to the present invention;

[0041] Figure 2 This is a schematic diagram of the structure of the recommendation system involved in the present invention;

[0042] Figure 3 Schematic diagram of various guide relationships modeled by the present invention;

[0043] Figure 4 This is a schematic diagram of an embodiment of the recommendation system involved in the present invention;

[0044] Figure 5 This is a movie recommendation flow chart involved in the present invention;

[0045] Figure 6 A schematic diagram of the relationship network constructed by the present invention for new users with scarce data;

[0046] Figure 7 A schematic diagram of the complete traceable recommended path constructed by the present invention;

[0047] Figure 8 A schematic diagram of the recommended principles for the accuracy and novelty of the present invention;

[0048] in, Figure 8 (a): Recommended schematic diagram focusing on accuracy, Figure 8 (b): Schematic diagram of novelty-focused recommendation. Figure 8 (c): Schematic diagram of the recommendation focusing on accuracy and novelty;

[0049] Figure 9 Interest index graph for recommendation patterns of existing recommendation systems;

[0050] Figure 10 An interest index graph for the recommendation model that takes both accuracy and novelty into account, adopted by the present invention;

[0051] Figure 11 This is a schematic diagram showing the effect of reducing the amount of system calculation by setting a threshold value in the present invention;

[0052] Figure 12 Schematic diagram of the principle of selecting recommended items by the recommendation algorithm of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] Explanation of relevant terms involved in this invention:

[0055] Neighborhood information enhancement: An image processing method that improves image quality, particularly in terms of detail and contrast, by analyzing and adjusting the relationship between each pixel and its neighbors. This technique is widely used in image restoration, medical imaging, and satellite imagery to enhance image visibility and key information.

[0056] Social sphere integration: A concept in social network analysis that involves integrating and modeling the strong and weak interpersonal relationships within different social spheres or communities. This integration provides a deeper understanding of the structure and dynamics of social networks, which is crucial for identifying key individuals within social spheres, predicting social trends, and developing effective social media marketing strategies.

[0057] Knowledge graph: A complex data structure consisting of "head entity-relationship-tail entity" structured data that can be used to organize and express complex relationships between entities. In recommendation systems, knowledge graphs can improve the accuracy and personalization of recommendations by connecting related entities such as users, products, and services, thereby enhancing the user experience.

[0058] Strong and Weak Tie Theory: Strong ties, as defined in sociology, refer to close interpersonal relationships, often found between close friends, family members, and closely collaborating colleagues. These relationships play a key role in social network analysis, demonstrating their unique value in information transfer, trust building, and group cohesion analysis. Unlike strong ties, weak ties refer to less frequent and less intensive social relationships, yet they demonstrate their unique value in facilitating information flow across social spheres, stimulating innovation, and building social capital. This theory has been widely applied in social network analysis, marketing strategy, and organizational behavior research.

[0059] Attention mechanism: A feature of computational models that dynamically adjusts the model's attention to different parts of the input data. By assigning different levels of attention to different parts of the model, this mechanism helps the model more effectively identify and process more critical information. In complex data environments, such as natural language processing or image recognition, attention mechanisms can significantly improve model performance and accuracy by capturing key information.

[0060] Graph Convolutional Network (GCN): A neural network architecture specialized for processing graph data. By learning the characteristics of nodes and their connections within a graph, GCNs can effectively process and interpret complex network structures. They are commonly used in fields such as image processing and social network analysis.

[0061] Guided Fusion Graph Convolutional Network: This paper proposes a neural network structure that combines the strengths of sociological strong and weak connection theory, attention mechanisms, and graph convolutional networks to improve the model's ability to process complex user interaction data. This network structure embeds user neighbor relationship information into entities through convolutional aggregation operations, modeling the combination of strong and weak connections between people in different social networks or social circles. It uses user groups with similar interests as the basic unit of recommendation, recommending high-quality content within the same social circle and novel and surprising content from different social circles, achieving item recommendations that are more similar to the structure of modern social relationship networks.

[0062] In order to solve the problems existing in the existing technology, this paper proposes an item recommendation method based on social-oriented relationships. It adopts an innovative method that integrates the neighborhood information enhancement technology based on knowledge graph and the social circle fusion technology based on attention mechanism to explore the current and future development of users' interests and achieve recommendations of various items that take into account both accuracy and novelty. Figure 1 , including the following steps:

[0063] S1: Collect user behavior data on different platforms, that is, raw data, which includes item data and user data.

[0064] Item data includes clicks, transaction volume, item category, and item attributes. User data includes browsing time, user reviews, and customer service / questionnaire feedback. Raw data is cleansed, formatted, and standardized to ensure the quality and efficiency of subsequent processing. Items include commodities, movies, books, and videos.

[0065] S2: Knowledge graph construction: refers to the process of converting raw data into a knowledge graph according to the "head entity-relationship-tail entity" triple pattern.

[0066] In this example, the head entity refers to the user, the relationship refers to the connection between users and users or items, such as friends (user<->user), evaluation (user->item) and feedback information (user->item), and the tail entity refers to the user or item. And define the knowledge graph ,in and Represent the head entity and tail entity sets respectively, Represents a set of relationship edges.

[0067] First, the knowledge graph Embedded in the recommendation model, multiple users are used as the first nodes and multiple items are used as the second nodes.

[0068] S3: Oriented relationship modeling.

[0069] Before modeling guided relationships, convolutional aggregation is performed on multiple nodes. This operation aggregates features of neighboring nodes through convolution. In graph data, nodes are connected by edges. Convolution aggregates information within a local neighborhood (i.e., nodes connected to the current node) to obtain a higher-level representation of the node.

[0070] Local neighborhood information refers to the information carried by a node's immediate neighbors. In a graph, a node's representation relies not only on its own information but also on information about its neighbors. By aggregating this neighbor information, a node can obtain richer contextual information.

[0071] Guidance relationship modeling is the core part of this invention, which involves social guidance relationship The definition of the relationship is based on any two users Embedded relationships The distance correlation is:

[0072] (1);

[0073] Where, express and The distance correlation between is a user pair, where the subscript Is the sequence number of the relationship.

[0074] Formula (2) is used to calculate the effective guiding relationship value between any two users. Then, by introducing the attention mechanism, the guiding relationship value is refined to obtain the final formula (7). Through guiding relationship modeling, this paper proposes an evaluation standard that can quantitatively and intuitively describe the connection between users and nodes, as shown in Formula (4).

[0075] The guided relationship mining process can be defined as calculating the first node u and v The similarity score between :

[0076] (2);

[0077] Where, The first nodeu and v The similarity score between are trainable weights, is the social orientation relationship between the two first nodes, The first node v The node characteristics of T is the matrix transpose operation. It can be expressed as:

[0078] (3);

[0079] Where, is the attention weight parameter, is the node feature output by the network, L It is a specific i value, Represents a splicing operation, is the output of the activation function.

[0080] Rating prediction and attention mechanism:

[0081] Use Sigmoid function as activation function, score It can be expressed as:

[0082] (4);

[0083] In fact, when getting the first node u and v After the similarity score is obtained, the influence of time sequence on similarity can be improved by the attention mechanism. The present invention uses the attention mechanism to assign relationship weights to different users and uses Representation attention mechanism creates guided relation embedding:

[0084] (5);

[0085] In the above formula, It's a relationship The embedding vector is assigned an attention score Expressing its importance, R All relationships r A collection of

[0086] (6);

[0087] Here, It's a relationship Relationship with guidance The trainable weights of yes R Except r Other relationships besides.

[0088] Single-path relationship: target user-guiding user-item to be recommended. The guiding user is used to expand the target user's social circle to effectively collect items to be recommended.

[0089] Furthermore, we introduce the embedding of attention weights and guidance relationships, namely:

[0090] (7);

[0091] In the above formula, is the oriented relationship value, is a trainable weight parameter.

[0092] Again, the Sigmoid function is used as the activation function, and the score It can be expressed as:

[0093] (8);

[0094] Finally, a simple binary cross entropy loss function is defined to measure the label and assessment score The error between:

[0095] (9).

[0096] By optimizing the loss function, the model can obtain better recommendation results in the targeted user group.

[0097] 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 through graph convolutional networks based on the attention mechanism, ultimately achieving solutions to the problems of cold start of existing recommendation systems, weak dynamic adaptability, poor interpretability, low user information utilization, and easy formation of user information cocoons.

[0098] S4: Guided relationship modeling threshold setting.

[0099] In the process of guiding relationship modeling, only when the threshold of a certain type of guiding relationship is less than 1, the guiding relationship will be generated and accessed. The smaller the threshold, the higher the priority of the guiding relationship relative to other guiding relationships, that is, the model will give priority to using this type of guiding relationship to form recommendation results. The threshold here is the control value of the priority of the guiding relationship. It is effective when the threshold is set to less than 1, that is, when the value calculated by formula (7) is When the threshold is exceeded, the guided relationship is enabled. Even if no guided relationship can be generated or accessed, the algorithm can still generate recommendations based on the strong and weak connection channels in the original data. The specific threshold parameters are automatically generated during the training of the guided fusion graph convolutional network.

[0100] Taking the Graph Attention Network (GAT) integrated with the attention mechanism as an example, its training process can be divided into the following key steps:

[0101] Data selection: Choose datasets suitable for graph structures, such as social networks, molecular structures, or knowledge graphs.

[0102] Data preprocessing: Constructing the graph adjacency matrix and node feature matrix. For graph data, it is usually necessary to calculate the node's neighbor information and normalize the features.

[0103] Model architecture: GAT introduces an attention mechanism to weight the features of aggregated neighbor nodes.

[0104] Multi-head attention: To enhance the expressiveness of the model, GAT usually adopts a multi-head attention mechanism to concatenate or average the attention results of different heads.

[0105] Loss function: The cross-entropy loss function is usually used to measure the difference between the model's prediction value and the true label.

[0106] Optimizer: Commonly used optimizers include Adam, whose hyperparameters such as learning rate and weight decay need to be adjusted according to the task.

[0107] Forward propagation: Input node features and graph structure information, and calculate the representation of each node through the graph attention layer.

[0108] Backpropagation: Calculate the gradient of the loss function and update the model parameters through backpropagation.

[0109] Training loop: Repeat the forward propagation and backpropagation process until the model converges or reaches the preset number of training epochs.

[0110] Validation set evaluation: Use the validation set to evaluate model performance. Common indicators include accuracy, recall, etc.

[0111] Early stopping mechanism: If the performance of the validation set no longer improves, training is stopped early to avoid overfitting.

[0112] Test set evaluation: Evaluate the final performance of the model on the test set to ensure that the model has good generalization ability.

[0113] Model saving: Save the trained model as a file for subsequent use.

[0114] S5: Setting the threshold of strong and weak connection channel length.

[0115] In the process of modeling the guided relationship, the process of accessing the strong and weak connection channel length threshold in the knowledge graph composed of raw data can be used. A channel refers to the path between two specific nodes (target node and recommended node) connected by a specific relationship. There can be multiple channels between two specific nodes (that is, connecting with nodes through different relationships). The channel length refers to the number of relationships contained in the path corresponding to a specific channel. The channel length threshold is a hyperparameter. When the channel length exceeds the threshold, the guided relationship becomes invalid (the weight is assigned to 0). When the channel length is less than the threshold, the guided relationship is valid and the weight is valid. See Figure 11 Schematic diagram of the effectiveness / ineffectiveness of the described guiding relationship.

[0116] The length of a strong and weak connection channel is proportional to the number of entities contained in that channel. Subsequent recommendation algorithms can access strong and weak connection channels only when this threshold is greater than 0. If the threshold is equal to 0, all relationships between all entities will be inaccessible, and no recommendations can be generated. This threshold parameter is automatically generated during network training.

[0117] S6: Recall.

[0118] Recall refers to the process of using the established directional relationship network to obtain as many items as possible that are valuable to the user. This step will automatically find the optimal user relationship network route and generate trainable weights during the network training process. Here, it refers to the channel weight parameter, which is a result value obtained by summing the weights of all relationships in a channel. It is calculated and generated in the process of finding the optimal user relationship network route. It is sorted according to the sum of the weights of each channel. If several channels are to be screened, the channels with the largest weights are selected. The channel with the largest channel weight is the optimal user relationship network route. See Figure 11 .

[0119] S7: Rough arrangement.

[0120] Rough sorting involves leveraging a small number of established directional relationship networks to obtain as many items as possible that are valuable to the user, sorting them by their predicted value, and removing duplicate items. This step uses the relationship network established in the previous step to identify duplicate items recommended to the same user and prune the recommended route to avoid duplication.

[0121] In this embodiment, by setting a smaller channel length threshold, only a small number of channels are activated, thereby reducing the computational effort when filtering items and ensuring, to a certain extent, that the items filtered are the ones the user is interested in. The nodes and relationships filtered out by the rough sorting will eventually be applied to the fine sorting stage.

[0122] S8: Fine arrangement.

[0123] This process leverages the majority of the established network of guiding relationships to identify a small number of items that are most valuable to the user, ranking them by their predicted value, and removing duplicate items. This step amplifies the contribution of certain key guiding relationships to the overall recommendation score (increasing their weight). This process differs from the previous one in that it prioritizes recommended items that are likely to be of greater interest to the user based on their value (the sum of the weights of each item's guiding relationship network) after the network has been established. It also further removes duplicate items.

[0124] In this embodiment, by setting a larger channel length threshold, most channels are activated, which may lead to the discovery of an optimal path with a larger channel weight than that in the coarse sorting stage (i.e., a channel with a channel length greater than the channel length threshold in the coarse sorting stage but a larger channel weight). Since the nodes and relationships used in the fine sorting stage are all from the coarse sorting stage, the complexity of the guided relationship network when subjected to a larger channel length threshold is also within an acceptable range.

[0125] S9: Re-ranking: refers to the process of inserting other advertisements and other content according to the marketing strategy, re-weighing the accuracy and novelty of the recommended items, and carefully adjusting the ranking of the recommended results.

[0126] S10: Recommendation result: refers to the process of generating the final recommendation result using the rearranged results.

[0127] S11: User operations: refers to the user’s actions such as liking, collecting and commenting on the recommendation results.

[0128] S12: Operation data: refers to the processing process of generating user operation behavior data based on user operation behavior.

[0129] S13: Update: refers to the process of adding new relationship networks that cause effective user operation behaviors to the original data.

[0130] The 13 steps above collectively constitute the recommendation process of this invention. Specifically, the thresholds for modeling user-directed relationships and setting thresholds for strong and weak connection channel lengths offer a recommendation technique that combines both breadth and depth compared to existing recommendation system algorithms (such as user-based collaborative filtering). This technique, based on user relationships, forms a directive network between users, creating a more comprehensive user profile and enabling in-depth recommendations based on browsing information from users who share their preferences.

[0131] Reference Figure 12 First, a user 1 to be recommended is selected, and the guidance relationship is modeled by the recommendation algorithm proposed by the present invention, and the sum of the guidance relationship weights is generated. NThen, according to the product purchase relationship of the users pointed by these guiding relationship channels, the most n Recommended objects, and then n The system then conducts trial recommendations for each product. If user feedback is unsatisfactory (e.g., a low click-through rate or blocked recommendation), the guidance relationship is regenerated. If user feedback is positive (e.g., a high click-through rate or high view rate), advertising and other information are inserted based on these products according to the marketing strategy to generate the final recommendation results. Finally, users take actions on the recommended products (e.g., click, purchase, review, favorite, block, etc.), generating new action data that is used to update the database and execute future recommendations.

[0132] The present invention relates to guided relationship modeling, referring to Figure 3 The various directional relationships defined are as follows (it is assumed that individuals in the same social circle already have relatively frequent interactions, and bidirectional arrows indicate strong connections):

[0133] ① A strong connection-oriented relationship within the same social circle is established through interest relations in the same social circle: if there is at least one strong connection relationship channel between an individual and other individuals in the same social circle that does not cross the social circle, then there is a probability that the two parties will form this oriented relationship.

[0134] ② A weak connection guiding relationship within a layer established through interest relations across social circles: if there is at least one strong connection channel across social circles between an individual and other individuals in the same social circle, there is a probability that the two parties will form this guiding relationship.

[0135] ③ A strong inter-layer connection-oriented relationship established through interest relationships across social circles: If there is at least one strong connection relationship channel across social circles between an individual and other individuals in different social circles, there is a probability that the two parties will form this guiding relationship.

[0136] ④ is the inter-layer weak connection guiding relationship established through the interest relationship of the same social circle: if there is no or only a small number of strong connection relationships between an individual and other individuals in the same social circle (usually a cold start problem), but there are strong connection relationship channels across social circles between other individuals and a large number of individuals in another social circle, then there is a probability that this guiding relationship will be formed between this individual and the individuals in the other social circle.

[0137] ⑤ A weak connection-oriented relationship within the same social circle is established through interest relations in the same social circle: if there is no or only a small number of strong connections between an individual and other individuals in the same social circle (usually a cold start problem), then there is a probability that this guiding relationship will be formed between both parties.

[0138] ⑥ is a strong connection guiding relationship within the layer established through interest relations across social circles: if there is at least one strong connection relationship channel that does not cross social circles between an individual and other individuals in the same social circle, and both parties have a strong connection relationship with the same individual in another social circle or between two individuals connected by at least one strong connection relationship channel that does not cross social circles, then there is a probability that this guiding relationship will be formed between the two parties.

[0139] ⑦ is the inter-layer weak connection guiding relationship established through interest relations across social circles: if there is at least one strong connection relationship channel across social circles between an individual and other individuals in different social circles that needs to be connected by other strong connection guiding relationships, then there is a probability that this guiding relationship will be formed between the two parties.

[0140] ⑧ is a strong connection guiding relationship between layers established through interest relations in the same social circle: if there are a large number of individuals in a social circle who have strong connection channels with a large number of individuals in another social circle, then there is a probability that the individuals in these two social circles will form this guiding relationship.

[0141] Different social circles affect user interests, and the strength of connections affects the degree of freshness of recommended content to users.

[0142] Among them, the "probability" mentioned means that when users have a certain interest relationship, if only one of the proposed conditions for establishing a guiding relationship is met, the probability of forming this guiding relationship is 100%; when there is more than one interest relationship, the weight of each guiding relationship is allocated according to the actual model training situation.

[0143] The results of guided relationship modeling are calculated based on importance (weight) using an attention mechanism. Each relationship occupies a certain proportion and plays a corresponding role. This guided relationship modeling allows content of interest to be recommended to another user based on accuracy or novelty through guided relationships. This effect is achieved regardless of whether the two users are acquaintances in the same social circle with many shared or similar interests or complete strangers in two completely different social circles, achieving item recommendations that balance accuracy and novelty.

[0144] The following is a user emotion judgment standard that can be used as a reference (similar effects can be achieved by modifying the thresholds):

[0145] Highly Interested: If a user views a certain type of content in 70 or more of their most recent 100 views, they are considered highly interested in that type of content. For users with fewer than 100 but more than 30 views, the user is considered highly interested in content of that type based on the percentage of views. For users with fewer than 30 views, content of that type that accounts for more than 70% of the views is considered highly interested. This is not considered highly interested in content of this type for users with fewer than 30 views.

[0146] Relatively Interested: If a user views a certain type of content in at least 30 of their most recent 100 views, they are considered relatively interested in that type of content. For users with less than 100 but more than 10 views, the user is considered relatively interested in that type of content based on the percentage of views. For users with less than 10 views, content representing at least 30% of the views is considered relatively interesting. This determination is not made for users with less than 10 views.

[0147] Relatively Novel: If a user has viewed fewer than 20 items of a particular type of content in their most recent 100 browsing history, but similar users in their social circle have shown a relatively high level of interest in that type of content, or have shown a relatively high level of sentimental judgment, then the user is considered relatively novel to that type of content. For users with fewer than 100 browsing history, if the user has viewed less than 20% of that type of content in their total browsing history, and that user has viewed that type of content in their most recent 5 browsing history, then the user is considered relatively novel to that type of content. This judgment does not apply to users with fewer than 10 browsing history.

[0148] Very Novel: If a user browses a certain type of content for at least 3 out of their most recent 10 views, but for less than 30 out of their most recent 100 views, the user is considered very novel about that type of content. This determination does not apply to users with fewer than 100 views.

[0149] It is worth noting that the above sentiment judgment is not an independent relationship and is only used to explain the effect of the guidance relationship proposed by the present invention and has nothing to do with the recommendation process.

[0150] This paper meticulously utilizes the self-attention mechanism to quantify the importance of combinations of strong and weak interpersonal connections within different social circles and models these social relationships. This process reveals the changing trends of user interests across the nature of multi-layered social circles. Based on the rich semantic relationships of the knowledge graph, this paper embeds local neighbor information into entities through advanced convolutional aggregation operations, achieving the transition from individual users to user groups. It also implements neighborhood selection based on attention weights, effectively avoiding the increased model complexity caused by excessively large neighborhoods. This expands the space for specific parameters of the artificial control system (such as the amount of computation required to control the model, the actual utilization rate of information, and the balance between the novelty and accuracy of recommended content), effectively improving the practicality and versatility of the model.

[0151] Based on the same concept, the present invention also provides an item recommendation system based on social-oriented relationships, including a collection unit, an embedding unit, a construction unit, an establishment unit and a recommendation unit.

[0152] The collection unit is used to collect item data and user data of multiple users, and build a knowledge graph based on the item data and user data.

[0153] The embedding unit is used to embed the knowledge graph into the graph convolutional network, where multiple users serve as first nodes and multiple items serve as second nodes.

[0154] The construction unit is used to obtain the similarity score between the corresponding two first nodes based on the social guidance relationship between any two users; assign guidance relationship weights to different first nodes through the attention mechanism, and obtain the guidance relationship value between the two first nodes based on the similarity score and the guidance relationship weight; and establish a guidance relationship network based on the guidance relationship value between any two first nodes.

[0155] The establishing unit is used to establish multiple connection channels between the target node and the recommended node in a guidance relationship network, with any first node as the target node and another first node that has no guidance relationship with the target node as the recommended node. Each connection channel is obtained by connecting multiple first nodes with all guidance relationships enabled; when the guidance relationship value between the two first nodes is greater than the corresponding guidance relationship threshold, the guidance relationship is enabled.

[0156] The recommendation unit is used to aggregate node information of second nodes connected to the target node to the target node, and recommend items to the recommendation node through multiple connection channels.

[0157] Based on the same concept, Figure 2 and Figure 4 The recommendation system of the present invention also includes a client and a server, wherein 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, and 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.

[0158] The client-operation module supports user actions such as likes, favorites, comments, and searches. Compared to the MovieLens and Jester datasets, which provide user ratings of specific books or films, this module utilizes richer behavioral data, better reflects user viewing habits, and can mine more user relationships, thus improving the utilization of user behavioral data.

[0159] The client-processing module is used to perform three data processing stages of extraction, conversion and loading of user operation behaviors, and package the user's operation behaviors in the operation module into text data that can be accepted and used by the server.

[0160] The client-sending module is used to send the user operation behavior data processed by the processing module to the server through the network.

[0161] The server-side receiving module is used by the server to receive the packaged user operation behavior text dataset through the network.

[0162] The server-side settings module is used to switch or modify recommendation model parameters (including thresholds for guided relationship modeling and strong and weak link channel length thresholds) based on changes in user behavior, item lists, and other data, or the business objectives of system users. This module is only enabled during system updates. Similar to the online server-side processing technology mentioned above, this module uses online data processing to ensure the relevance and timeliness of the recommendation model.

[0163] The server-side verification module is used to check whether there are any errors in the recommendation model after switching and modifying. It is only enabled when the system is updated. This module is the verification module of the setting module. It can verify the recommendation 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 watched by the user in a short period of time (such as three times in the past 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 of time (such as ten times in the past week) has been included in the blacklist of the recommendation model for the next recommendation. That is, the above-mentioned content will no longer be recommended, and all generated recommendation results are composed of relationship weights that meet the requirements of the setting module. This module serves as a guarantee of the correctness and rationality of the recommendation model before it goes online, and can avoid errors in the updated model.

[0164] The server-side recommendation module is used to call upon the recommendation model to complete personalized recommendations for information or items. After receiving user behavior data on the server, and unless the system requires an update, the data is directly transmitted to this module for online processing. The server then selects several items that the user may be most interested in and generates recommendation reasons for each item based on the composition of the relationship weights.

[0165] The server-side update module processes the recommended items and reasons selected by the recommendation module, packaging them into data that can be accepted and used by the client, and updates user relationship information in the database. After a valid recommendation is made, the recommendation module generates new relationship network information. The user's actions on the previous recommendation result can be used to determine whether the previous recommendation was valid. If so, the relationship network information generated by the previous recommendation can be updated.

[0166] The server-side sending module is used to send the packaged recommended items and recommendation reasons and other recommendation result data to the client through the network.

[0167] The client-receiving module is used to receive the packaged recommended items and recommendation reasons and other recommended content through the network.

[0168] The client-side validation module uses a fixed format to determine whether recommendation result data contains packet loss or formatting errors. It then tests partial recommendation results and uses user behavior data such as clicks and browsing feedback to verify that the recommendations meet requirements. It then determines whether additional advertising should be inserted based on marketing strategies, and generates the final, complete recommendation results. Existing recommendation systems typically lack this "test recommendation" process.

[0169] The client-display module is used to display the final recommendation results to the user and allow the user to perform subsequent operations such as liking and watching.

[0170] This paper proposes a graph convolutional neural network item recommendation algorithm based on knowledge graph and sociological 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, and extract high-order guiding relationship features of users in different social circles at the weak connection level, and provide users with accurate recommendations and novel recommendations based on strong connections and weak connections respectively, taking into account users' emotional preferences and the diversity of recommendations.

[0171] The system collects and categorizes users' browsing preferences, forming a social graph in the background. This is then further differentiated based on the preferences of each user's social graph. After differentiation, users form connections of varying strengths based on their preferences. New relationships are then embedded based on the weight of connections between users in different circles.

[0172] Through the guided relationship network, the background will push relevant content to users, and then continuously refine the user portrait based on the user's click trends, enrich the user's social circle, deeply cross the social connection relationships between users, make in-depth recommendations, and increase the diversity of recommended content.

[0173] In the way that information is differentiated to form user social circles, user preset operations include one or more of the following operations: multiple browsing, repeated browsing of the same type of web pages, comment operations, sharing operations, and follow-up lists, etc.

[0174] The algorithms provided herein are not inherently tied to any particular computer, virtual system, or other device. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the present invention described herein can be implemented using a variety of programming languages, and the descriptions of specific languages ​​above are provided to better illustrate the functionality of the system.

[0175] Attachment Figure 2This diagram depicts the architecture of a recommendation system, including hardware such as a power supply component. The core function of the communication interface is to collect user behavioral data from various platforms through input and output devices, including clickstreams, browsing time, transaction data, reviews, and feedback, and to deliver the recommendation results generated by the recommendation system to the user's client. This behavioral data also includes social interaction data, such as user posts, interaction frequency (likes, comments, or favorites), discussion topics, and structural dynamics within the social network (friending, blacklisting, etc.). The power supply component monitors the usage patterns and power consumption of other hardware to ensure proper system operation. The data preprocessing submodule, after data collection, performs data cleansing. Standardization converts data into a format more suitable for analysis. Formatting further ensures data consistency across different sources. For example, data representation across different social media platforms is standardized to facilitate cross-platform analysis.

[0176] Attachment Figure 1 This paper details the recommendation process in a recommendation system. This process builds upon the guiding relationships of a knowledge graph and employs graph analysis techniques to construct a user's social network graph, establishing the complex connections between the user and other individuals. A user's position in the graph is determined by calculating network metrics such as node degree, betweenness, and closeness. These metrics reflect the user's importance and influence within their social network. Subsequently, all processes involving social network analysis (such as recall, coarse ranking, fine ranking, and re-ranking) identify the components of a user's complex guiding relationship network, including the definition of the user's primary social circle, secondary social circles, and various marginal connections. By deeply analyzing an individual's social graph, this process also identifies and quantifies the user's social influence, which helps the system understand what content the user might be most sensitive to or interested in, and predicts what content they are currently seeking.

[0177] Obviously, Figure 1 belong Figure 4 For example, Figure 1 "Guided relationship modeling threshold setting" and "Strong and weak connection channel length threshold setting" are both Figure 4 "Server-Settings Module" in the Figure 1 The "update" operation in Figure 4 "Server-Update Module" in Figure 1 "Knowledge graph construction", "guided relationship modeling", "recall", "rough sorting", "fine sorting" and "re-sorting" are all included in Figure 4 "Server-Recommendation Module" in Figure 1 The "user action" in Figure 4 "Client-Operation Module" in Figure 1 The "operational data" in Figure 4 "Client-Processing Module" in. Figure 1 The "recommended results" in Figure 4 Data sent from the server to the client.

[0178] At the same time, the various server-side processing processes involved above are implemented in Figure 2 The "program" and "processor" in the Figure 2 The "storage" in the server and client are all transmitted through Figure 2 The "communication interface" in the implementation; the order of the work of each module on the server side is determined by Figure 2 The "controller" in the implementation; all data communication within the server is based on Figure 2 All server-side hardware operations are based on the "communication bus" in Figure 2 Support for "Power Components" in .

[0179] Due to the high uncertainty of the user equipment configuration of the client hardware, the present invention does not need to make special restrictions on it.

[0180] Specifically, Figure 4 The focus is on the client-side and server-side operational mechanisms, which involve integrating users' behaviors and roles within different social circles into the recommendation system. After generating recommendations and feeding them back to the user, the system updates the database with new social relationship information for the next recommendation. This operational mechanism analyzes the types and frequency of user interactions within these circles. This process specifically incorporates social identity theory, identifying user behaviors within different circles as specific social identities and analyzing how these identities interact and influence users' preferences and decisions. Furthermore, the social circle integration process assesses users' engagement and influence within their social circles, including quantifying their responses to events within and outside their social circles, such as commenting on news events or adopting popular trends. Through in-depth analysis of these behaviors, the system can identify new content that users are likely to accept. After a recommendation is completed, the system assesses the effectiveness of the previous recommendation (e.g., whether it was liked, viewed, or saved by users). It selects valid data to update the database and update the guiding relationships (involving the creation of new guiding relationship networks and the reduction or increase in the weight of existing guiding relationship networks) for the next recommendation.

[0181] Obviously, Figure 4Each module in the system not only operates independently but is also tightly connected in the data flow, forming a comprehensive recommendation system framework that combines client and server-side functions. The outputs of these modules are integrated by a carefully designed algorithm to produce a list of recommendations that balances accuracy and novelty. This ensures that the system not only reflects users' current behavior and preferences but also predicts and promotes their exploration of potential interests.

[0182] In addition, the reduced exposure of user data also reduces the risk of user privacy data leakage. Figure 5 In the framework of , the present invention explains a movie recommendation method that combines deep relationships in social networks.

[0183] User behavior in social networks is not isolated; rather, it exists as interconnected nodes within multi-layered social circles. User preferences and choices are the result of a combination of direct and indirect factors influencing their social environment. Traditional recommendation systems often fail to fully leverage these social elements, resulting in limitations in the accuracy and diversity of recommended content. The innovation of this technology lies in integrating user social dimension information and enhancing the performance of personalized recommendation systems through precise social network analysis.

[0184] The system's initial module is the data collector, responsible for collecting user behavior data across multiple social platforms, such as browsing history, purchase history, review feedback, and social interaction data. The resulting data stream is filtered through a preprocessing module, which includes data cleaning, formatting, and standardization steps to ensure the quality and efficiency of subsequent processing modules.

[0185] After preprocessing, the data is fed into the recommendation engine, the core of the recommendation system. This engine incorporates cutting-edge data mining algorithms, including but not limited to machine learning and deep learning techniques, to perform a detailed analysis of individual users' neighborhood information. This neighborhood information analysis considers the user's relative position in their social network and the strength of their relationships with other nodes, which in turn determines the degree of personalized recommendations.

[0186] The system then enters the social circle integration phase, using a specially designed algorithmic module to analyze and reveal users' behavioral patterns and roles in different social environments. This module not only analyzes the user's interactions with direct contacts, such as friends, family, or colleagues, but also expands to the user's position in the broader social graph, such as the user's connections with others who share common interests, as well as indirect connections established through third-party introductions.

[0187] All of these social network analysis results, from micro to macro levels, are integrated and fed into the recommendation generation module. This module utilizes a sophisticated algorithm that combines the user's personal preferences with their position and role within their social network to produce a customized list of recommendations. This list is algorithmically optimized to not only reflect the user's immediate interests but also encourage exploration of novel content, increasing the likelihood of discovering previously undiscovered areas.

[0188] Finally, the system will update the database again based on the recommendation effect to continue to improve the practicality of the recommendation.

[0189] Throughout the recommendation process, the user experience module monitors and analyzes user interactions with recommended content, using this feedback to adjust recommendation strategies to ensure content remains relevant and engaging. The system's adaptive learning mechanism adjusts the recommendation algorithm in real time based on evolving user behavior, ensuring the continuous adaptability of recommended content.

[0190] Reference Figure 6 The present invention can generate a large number of guided relationships based on a small number of existing user (user 1-user 8) and product (item 1-item 12) relationships (purchase relationships) to realize associated data mining, and then complete product recommendations for new users who have strong connection relationships with user 4 and user 5.

[0191] Reference Figure 7 , when user 1 obtains a recommendation for item 1 through the directional relationship formed by user 2, user 3, and user 4, the present invention can clearly know the reason why item 1 is recommended to user 1 (because user 4 and user 1 have a valuable social network, see Figure 3 ). Figure 7 The purchase relationship in the example refers to User 4's purchase of Item 1, providing a direct interest signal for subsequent recommendation results. The recommendation path represents the information propagation path, based on the guiding relationships between users, from the purchaser (User 4) through User 3 and User 2, to User 1, the recipient of the recommendation. The traceability path represents the reverse recommendation path, tracing back from User 1's recommendation results to the initial purchase behavior (User 4's purchase of Item 1) and the underlying social relationships.

[0192] Reference Figure 8 Because individuals in the same social circle (social circle 1 or social circle 2) are more likely to have similar interests, accurate recommendations tend to recommend products that other individuals in the same social circle are interested in. Because individuals in different social circles are more likely to have dissimilar interests, novelty recommendations tend to recommend products that other individuals in different social circles are interested in. This invention achieves both accurate and novelty recommendation effects.

[0193] Reference Figure 9 and Figure 10 , C refers to a recommendation model that only focuses on accuracy, focusing on recommending a small range of content, and easily forming a knowledge cocoon. D refers to a recommendation model that takes into account both accuracy and novelty, creating new points of interest while protecting the original interests. The present invention can take into account this accuracy and novelty recommendation effect. The recommendation accuracy referred to in the present invention: the recommendation model recommends products of interest to target users by focusing on high-interest areas (high interest index, corresponding high guide channel weight, and 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 shows that the target users have a high acceptance of this type of products, which meets the recommendation accuracy standard.

[0194] The novelty of recommendation referred to in this invention is that 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 user's known interests, they are new points of interest mined based on the guiding relationship (such as Figure 10 The interest peak for the “documentary” category in the dataset is a smaller peak where the weight of the guiding channel is lower than the accuracy recommendation).

[0195] This method not only recommends familiar products to users by targeting primary interest peaks with high weights, but also guides users to new areas of interest by targeting secondary interest peaks with the next highest weights, broadening their scope of interests. This method can creatively recommend novel products while maintaining recommendation accuracy, effectively implementing recommendations that balance accuracy and novelty.

[0196] Example

[0197] Forty volunteers (20 adolescents and 20 middle-aged and elderly, with a 50:1 ratio of male to female) were invited to conduct an A / B test of the recommendation system proposed in this paper (System A) and a simpler recommendation system (System B) based on CGAT (Contextualized Graph Attention Network, a novel recommendation algorithm). The model training dataset used was Movie-lens-20M.

[0198] During the test, each volunteer performed 10 recommendation operations (5 of which were randomly selected using System A and 5 using System B, with recommendations generated based on simulated behaviors such as purchasing or reading books or watching movies). After each recommendation operation, the volunteer was asked to rate the result on a scale of 0 to 5. Finally, the scores were normalized independently for each group of volunteers to determine which recommendation system the volunteers generally preferred.

[0199] Table 1 AB test results comparison table

[0200]

[0201] The original scoring data in Table 1 show that adolescent and middle-aged male and female volunteers always give higher scores to the recommendation results of System A than to System B. The average score of a single recommendation increases by as much as 8.78%, which is equivalent to 0.5 points on a scale of 0 to 5 points. The improvement effect is quite obvious.

[0202] System A had the most significant effect on improving scores among middle-aged and elderly female volunteers, reaching 12.02%, exceeding 1 point on the 0-5 scale. However, it had the weakest effect on improving scores among adolescent male volunteers, at only 6.99%. However, this 12.02% score increase was significantly larger than the other increases of 6.99%, 8.33%, and 7.78%, respectively, suggesting a degree of chance.

[0203] In addition, the raw scoring data in Table 1 also show that male volunteers generally give higher scores to the recommendation results than female volunteers, and there are two phenomena:

[0204] (1) The recommendation results of system A given by female volunteers of the same age group are close to the recommendation results of system B given by male volunteers.

[0205] (2) The recommendation scores of middle-aged and elderly volunteers for systems A and B are always significantly higher than those of young volunteers.

[0206] In order to minimize the differences in scoring standards among different volunteers, this embodiment normalizes the original scoring data on a group basis.

[0207] For the normalized scoring data, the normalization process weakens the impact of high-level scores on the total score, which ultimately results in a total score that is lower than the original total score.

[0208] Furthermore, normalization weakens the differences in low-level scores, causing the total scores of groups with fewer high-level scores to decline significantly. This is particularly true for adolescent female volunteers, who rated System A lower than System B. At the same time, normalization preserves some unique phenomena, such as the fact that female volunteers of the same age group rated System A close to the ratings of male volunteers for System B, and that middle-aged and elderly volunteers rated Systems A and B significantly higher than adolescent volunteers.

[0209] To sum up, in the AB test, system A achieved a general improvement in user feedback scores compared to system B, showing a more obvious optimization effect of 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.

[0210] The present invention establishes a recommendation system that can adapt to new users and dynamic changes by deeply mining the complex social relationships of users. This system can not only alleviate the sparsity and cold start problems in traditional recommendation systems, but also enhance the system's dynamic adaptability and provide accurate and novel recommended items. By constructing a complete recommendation explanation path, the interpretability of the recommendation system is enhanced, and the user's sense of trust is improved. At the same time, the system can take into account the accuracy and novelty of recommended items, avoid the "information cocoon" problem, and broaden the user's horizons. In addition, by deeply mining the relationship value between individuals and social groups, the utilization rate of information 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, thereby improving user stickiness and retention rate. Ultimately, the present invention realizes a virtuous cycle of "making the best use of things" and "making the best use of people" based on a complex relationship network, improving the overall performance of the recommendation system and user experience.

[0211] This invention primarily models the guiding relationships during the recommendation process, rather than directly using the user's original information. This reduces the frequency of user data usage and the number of external links that access user data, significantly reducing the risk of privacy data leakage. By manually adjusting the model's computational load, the actual utilization of information, and the balance between the novelty and accuracy of recommended content, this invention can achieve various specific recommendation objectives while protecting user privacy, enhancing the system's versatility and practicality.

[0212] The system effectively addresses data sparsity, alleviates cold start issues, and enhances dynamic adaptability through graph embedding technology and a guided relationship network. Through a multi-layered social network model, it deeply mines user behavior patterns and influence, identifies potential interests, and increases the exposure of long-tail content. Furthermore, the system employs a real-time data stream recommendation algorithm and a hybrid recommendation model to dynamically adjust the guided relationship network. This system combines user profiles, historical behavioral data, social network data, item content features, category tags, and user feedback to generate a comprehensive recommendation score, ensuring the depth, breadth, and timeliness of recommendations. For new users, the system generates potential points of interest based on a small number of interaction records, building a refined user profile. For returning users, the system limits the number of guided relationships to prevent explosive growth. Furthermore, the system provides a complete explanation path for recommendations, enhancing user trust and improving user stickiness through a novelty-accuracy trade-off.

[0213] The system provides a parameter adjustment mechanism, allowing operators to dynamically adjust model computational load and information utilization based on performance requirements, resource constraints, marketing strategies, and other factors to accommodate diverse application scenarios and user needs. Furthermore, the system allows operators to dynamically adjust the novelty and accuracy of recommended content based on user group characteristics and market trends. Because the system primarily implements recommendations by modeling guiding relationships, users' original information is only used for modeling and does not directly participate in the recommendation process. This significantly reduces the risk of user privacy data leakage.

[0214] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0215] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention is intended to include such modifications and variations.

Claims

1. A method for recommending items based on social-oriented relationships, characterized in that: The following steps are involved: Collect item data and user data from multiple users, and build 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 first nodes and multiple items serve as second nodes. Based on the social orientation relationship between any two users in different social circles, a similarity score between the corresponding two first nodes is obtained; through the attention mechanism, orientation relationship weights are assigned to different first nodes, and the orientation relationship value between the two first nodes is obtained based on the similarity score and the orientation relationship weight; based on the orientation relationship value between any two first nodes, an orientation relationship network is established; In a guiding relationship network, any first node is used as a target node, and another first node that has no guiding relationship with the target node is used as a recommended node. Multiple connection channels are established between the target node and the recommended node, each connection channel being formed by connecting multiple first nodes with all guiding relationships enabled. When the guiding relationship value between the two first nodes is greater than the corresponding guiding relationship threshold, the guiding relationship is enabled. Aggregate node information of the second node connected to the target node to the target node, and recommend items to the recommendation node through multiple connection channels; The similarity scores are specifically as follows: Where S uv is the similarity score between the first nodes u and v, α uv is the trainable weight, g ur is the social-oriented relationship between two first nodes, is the node feature of the first node v, T is the matrix transpose operation; The method of assigning guiding relationship weights to different first nodes through the attention mechanism and obtaining the guiding relationship value between two first nodes based on the similarity score and the guiding relationship weight specifically includes the following steps: The attention mechanism is used to assign guidance relation weights to different first nodes to obtain guidance relation embeddings; Combine the similarity score with the guiding relationship embedding to obtain the guiding relationship value between the two first nodes; The guiding relationship embedding is specifically as follows: e g N∑ r∈R δ(r,g ur )e r 100. Where, e g To guide the relationship embedding, e r is the embedding vector of relation r, δ(r,g ur ) is the attention score, R is the set of all relations r; The guiding relationship value is specifically as follows: Where S' uv is the guiding relationship value, and α is a trainable weight parameter.

2. The method for recommending items based on social-oriented relationships according to claim 1, wherein: The method of recommending items to the recommendation nodes through multiple connection channels specifically includes the following steps: Training the guided relationship network to obtain a plurality of trained trainable weights, and generating a plurality of first connection channels according to the plurality of trained trainable weights; A first channel length threshold is set, multiple first connection channels with lengths shorter than the first channel length threshold are retained to obtain multiple second connection channels, and items are recommended to the recommendation node through the multiple second connection channels; A second channel length threshold is set, and multiple second connection channels that are smaller than the second channel length threshold are retained to obtain multiple third connection channels, and items are recommended to the recommendation nodes through the multiple third connection channels; wherein the first channel length threshold is smaller than the second channel length threshold.

3. The method for recommending items based on social-oriented relationships according to claim 1, wherein: The user data includes browsing time, user evaluation and questionnaire feedback, and the item data includes item click volume, transaction volume, item category and item attributes.

4. An item recommendation system based on social-oriented relationships, characterized in that: include: A collection unit, configured to collect item data and user data of multiple users and construct a knowledge graph based on the item data and user data; An embedding unit, configured to embed the knowledge graph into a graph convolutional network, wherein a plurality of users serve as first nodes and a plurality of items serve as second nodes; A construction unit is configured to obtain a similarity score between two corresponding first nodes based on a social orientation relationship between any two users in different social circles; assign orientation relationship weights to different first nodes through an attention mechanism, and obtain an orientation relationship value between the two first nodes based on the similarity score and the orientation relationship weight; and establish an orientation relationship network based on the orientation relationship value between any two first nodes; an establishing unit configured to establish, in a guiding relationship network, a plurality of connection channels between the target node and the recommended node, with any first node as the target node and another first node that has no guiding relationship with the target node as the recommended node, each connection channel being formed by connecting a plurality of first nodes with all guiding relationships enabled; and when a guiding relationship value between two first nodes is greater than a corresponding guiding relationship threshold, the guiding relationship is enabled; A recommendation unit, configured to aggregate node information of a second node connected to a target node to the target node, and recommend items to the recommendation node through multiple connection channels; The similarity scores are specifically as follows: Where S uv is the similarity score between the first nodes u and v, α uv is the trainable weight, g ur is the social-oriented relationship between two first nodes, is the node feature of the first node v, T is the matrix transpose operation; The method of assigning guiding relationship weights to different first nodes through the attention mechanism and obtaining the guiding relationship value between two first nodes based on the similarity score and the guiding relationship weight specifically includes the following steps: The attention mechanism is used to assign guidance relation weights to different first nodes to obtain guidance relation embeddings; Combine the similarity score with the guiding relationship embedding to obtain the guiding relationship value between the two first nodes; The guiding relationship embedding is specifically as follows: e g N∑ r∈R δ(r,g ur )e r 100. Where, e g To guide the relationship embedding, e r is the embedding vector of relation r, δ(r,g ur ) is the attention score, R is the set of all relations r; The guiding relationship value is specifically as follows: Where S' uv is the guiding relationship value, and α is a trainable weight parameter.

Images