Recommendation method and device based on enhanced graph neural network, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to service system platforms of medical health, financial science and technology and the like, and discloses a recommendation method based on an enhanced graph neural network. Calculating node attributes and associated edge weights according to the processed user multi-source information, and constructing a graph structure through the node attributes and the associated edge weights; generating a node importance evaluation result according to the graph structure and the processed user multi-source information; dynamically adjusting an information propagation path and a node propagation weight according to a node importance evaluation result and feature correlation; and carrying out cooperative processing on the target recommendation task according to the information propagation path and the node propagation weight through an attention mechanism and a graph neural network, and outputting a recommendation result. According to the method, user behavior changes are captured in real time through a dynamic graph structure, invalid information is filtered in combination with a self-adaptive propagation mechanism, and accurate personalized recommendation is realized through cooperation of an attention mechanism and a graph neural network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a recommendation method and device based on an enhanced graph neural network, equipment and a storage medium. BACKGROUND

[0002] In the current field of recommendation systems, collaborative filtering and content-based recommendation are two mainstream methods. Collaborative filtering relies on the similarity of user behavior for recommendation, while content-based recommendation focuses on the matching degree of item features and user historical preferences. However, these traditional methods have certain limitations in information utilization and propagation mechanisms. Traditional collaborative filtering and content-based recommendation are difficult to fully capture the complex high-order relationships between users and items. For example, when analyzing the recommendations of user A, these methods cannot effectively utilize the association information between the friends of user A and the items, which may result in the recommended results ignoring potential user interest points. With the rapid increase in the number of users and the variety of items, the problem of data sparsity becomes increasingly serious. In the high-dimensional sparse data environment, traditional methods are difficult to generate effective user and item representations, which further reduces the recommendation effect. For example, on large e-commerce platforms, a large number of users may only have interacted with a small number of goods, which makes it difficult for traditional algorithms to accurately grasp user preferences.

[0003] In recent years, recommendation systems based on graph neural networks have begun to emerge. These systems construct users and items into a graph structure and capture the interaction patterns between users and items through the connection relationships of the graph. However, existing recommendation systems based on graph neural networks also have some problems. On the one hand, they have certain difficulties in combining with other recommendation models, and cannot fully utilize the advantages of different models to further improve the recommendation effect. On the other hand, these models lack adaptability to changing user needs and data characteristics. These problems limit the performance and application range of recommendation systems in complex scenarios.

[0004] In the field of financial technology, traditional collaborative filtering cannot handle high-order relationships between users and financial products (such as the transmission of product preferences in user social circles), resulting in a recommendation accuracy of less than 60% in data sparse scenarios; existing graph neural network recommendations are static in graph structure, and when a user adds a transaction record, the recommendation update delay is more than 1 hour, which cannot meet the real-time requirements. These methods are difficult to fully capture the complex high-order relationships between users and financial products, for example, they cannot effectively utilize the association information between the friends of user A and the financial products, which may result in the recommended results ignoring potential user interest points. With the rapid development of financial technology business, the number of users and the variety of financial products are increasing rapidly, and the problem of data sparsity is becoming increasingly serious. In the high-dimensional sparse data environment, traditional methods are difficult to generate effective user and financial product representations, which further reduces the recommendation effect.

[0005] In the medical health business field, traditional collaborative filtering and content-based recommendation methods also face challenges. These methods are difficult to fully capture the complex high-order relationship between patients and medical products, medical services. For example, when analyzing the recommendations of patient A, it is difficult to effectively utilize the association between the health information of patient A's family or friends and medical products, which may result in the recommended results ignoring potential patient needs. With the digital transformation of medical health business, the number of patients and the types of medical products and services are increasing, and the problem of data sparsity is becoming more and more serious. In the high-dimensional sparse data environment, traditional methods are difficult to generate effective patient and medical product / service representations, which further reduces the recommendation effect. In addition, the existing graph neural network-based recommendation system has certain difficulties in combination with other recommendation models in the medical health field, and cannot fully exert the advantages of different models to further improve the recommendation effect. SUMMARY

[0006] The main purpose of the present application is to provide a recommendation method, device and equipment based on enhanced graph neural network and storage medium, aiming at solving the technical problems that the existing recommendation system has limitations in capturing relationships and cannot accurately personalized recommend.

[0007] To achieve the above-mentioned purpose, the present application provides a recommendation method based on enhanced graph neural network, comprising: obtaining user multi-source information and preprocessing the user multi-source information; According to the preprocessed user multi-source information, the node attribute and the associated edge weight are calculated, and the graph structure is constructed through the node attribute and the associated edge weight; According to the graph structure and the preprocessed user multi-source information, a node importance evaluation result is generated; According to the node importance evaluation result and the feature correlation dynamic adjustment information propagation path and node propagation weight; Through the attention mechanism and the graph neural network, the target recommendation task is cooperatively processed according to the information propagation path and the node propagation weight, and the recommendation result is output.

[0008] Further, to achieve the above-mentioned purpose, the present application provides a recommendation device based on enhanced graph neural network, comprising: Data acquisition and processing module, for obtaining user multi-source information and preprocessing the user multi-source information; Graph structure construction module, for calculating node attribute and associated edge weight according to preprocessed user multi-source information, and constructing graph structure through node attribute and associated edge weight; Node importance evaluation module, for generating node importance evaluation result according to the graph structure and the preprocessed user multi-source information; The propagation dynamic adjustment module is configured to dynamically adjust an information propagation path and a node propagation weight according to the node importance evaluation result and feature correlation dynamic adjustment information. The hybrid model prediction module is configured to perform collaborative processing on a target recommendation task according to the information propagation path and the node propagation weight by means of an attention mechanism and a graph neural network, and output a recommendation result.

[0009] Further, to achieve the above object, the present application also provides a computer device, which comprises a memory, a processor, and a recommendation program based on an enhanced graph neural network stored in the memory and executable on the processor, and the recommendation program based on the enhanced graph neural network implements the steps of the recommendation method based on the enhanced graph neural network when executed by the processor.

[0010] Further, to achieve the above object, the present application also provides a computer readable storage medium, which stores a recommendation program based on an enhanced graph neural network, and the recommendation program based on the enhanced graph neural network implements the steps of the recommendation method based on the enhanced graph neural network when executed by a processor.

[0011] Beneficial effects: The present application relates to the field of artificial intelligence technology, and can be applied to business system platforms such as medical health and financial technology. A recommendation method based on an enhanced graph neural network is disclosed, which comprises: obtaining user multi-source information, and preprocessing the user multi-source information; calculating node attributes and associated edge weights according to the preprocessed user multi-source information, and constructing a graph structure through the node attributes and the associated edge weights; generating a node importance evaluation result according to the graph structure and the preprocessed user multi-source information; dynamically adjusting an information propagation path and a node propagation weight according to the node importance evaluation result and feature correlation dynamic adjustment information; and performing collaborative processing on a target recommendation task according to the information propagation path and the node propagation weight by means of an attention mechanism and a graph neural network, and outputting a recommendation result. The present application can capture user behavior changes in real time through a dynamic graph structure, filter invalid information through an adaptive propagation mechanism, and realize precise personalized recommendation through the cooperation of the attention mechanism and the graph neural network. BRIEF DESCRIPTION OF DRAWINGS

[0012] The present application will be further described below in combination with the drawings and embodiments, wherein: Figure 1 An application environment diagram of the recommendation method based on the enhanced graph neural network in an embodiment of the present application; Figure 2 A flow diagram of the recommendation method based on the enhanced graph neural network in an embodiment of the present application; Figure 3 A functional module diagram of the recommendation device based on the enhanced graph neural network in a preferred embodiment of the present application; Figure 4 Fig. 1 is a structural schematic diagram of a computer device in an embodiment of the present application; Figure 5 Fig. 2 is another structural schematic diagram of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0013] It should be understood that the specific embodiments described herein are merely exemplary and are not intended to limit the present application.

[0014] The recommendation method based on the enhanced graph neural network provided by the embodiments of the present application can be applied in an application environment such as Figure 1 , wherein a user end communicates with a service end through a network. The service end can obtain user multi-source information through the user end, and pre-process the user multi-source information; calculate node attributes and associated edge weights according to the pre-processed user multi-source information, and construct a graph structure through the node attributes and the associated edge weights; generate a node importance evaluation result according to the graph structure and the pre-processed user multi-source information; dynamically adjust an information propagation path and a node propagation weight according to the node importance evaluation result and a feature correlation; and perform collaborative processing on the target recommendation task through an attention mechanism and a graph neural network according to the information propagation path and the node propagation weight, and output a recommendation result. The present application can capture user behavior changes in real time through a dynamic graph structure, filter invalid information through an adaptive propagation mechanism, and realize precise personalized recommendation through the cooperation of the attention mechanism and the graph neural network. The user end can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices. The service end can be realized by an independent server or a server cluster composed of multiple servers. The present application will be described in detail through specific embodiments.

[0015] Please refer to Figure 2 , Figure 2 Fig. 3 is a flowchart of an embodiment of the recommendation method based on the enhanced graph neural network provided by the present application. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described can be performed in an order different from that described herein.

[0016] As shown in Figure 2 , the recommendation method based on the enhanced graph neural network provided by the present application includes the following steps: S100, obtaining user multi-source information, and pre-processing the user multi-source information; S200, calculating node attributes and associated edge weights according to the pre-processed user multi-source information, and constructing a graph structure through the node attributes and the associated edge weights; S300, generating a node importance evaluation result according to the graph structure and the preprocessed user multi-source information; S400, dynamically adjusting an information propagation path and a node propagation weight according to the node importance evaluation result and feature correlation; S500, performing collaborative processing on a target recommendation task by an attention mechanism and a graph neural network according to the information propagation path and the node propagation weight, and outputting a recommendation result.

[0017] The embodiment relates to a recommendation system method based on a graph neural network and an attention mechanism. First, the system obtains multi-source information of a user, including user basic information, a user behavior sequence, social relationship information and item information, and pre-processes the information. The pre-processed user multi-source information is used to calculate node attributes and associated edge weights, and then a graph structure is constructed. In the graph structure, the pre-processed basic information and behavior sequence features of the user are converted into quantifiable features and integrated into a user node; in addition to basic attributes, the extracted latent features of the item are used as supplementary attributes. The interaction frequency between users is used to calculate the social relationship edge weight. For items, similarity analysis of similar items and co-occurrence frequency calculation of different types of items with correlations are used to construct the correlation edges and weights between items. Finally, taking the user and the item as nodes, according to the calculated node attributes and associated edge weights, a complex network containing user nodes, item nodes and associated edges between them is constructed.

[0018] In the network, the activity of the user node is measured by the number of behaviors (such as clicks, purchases, etc.) within a certain period of time, and the higher the activity, the higher the importance of the user node. The activity of the item node is evaluated by the number of times it is browsed, purchased and the trend of its popularity (such as whether it has shown an upward trend recently). In addition, the system also calculates the correlation between the node features and the target recommendation task, and uses the centrality (such as degree centrality, betweenness centrality) of the node to evaluate its position importance in the graph structure. The node that is in the center of the network and connects more key nodes has higher importance in information propagation. By integrating the activity, feature correlation and centrality of the node, an importance score is assigned to each node, and a node importance evaluation result is generated.

[0019] Based on the node importance evaluation results, the system assigns an importance score to each node and makes the information propagated by high importance nodes have higher weight in the information propagation process. Through the path analysis algorithm, the propagation paths that are more valuable for recommending to the target user are identified, such as the paths that pass through multiple high-relevance item nodes, and the information propagation strength of these paths is increased. Conversely, for paths with lower value, the propagation is reduced or directly truncated. At the same time, an attention mechanism is introduced to preprocess the input node features. For user nodes, the attention mechanism focuses on historical behavior features that are highly relevant to the user's recent interests; for item nodes, it focuses on item attribute features that have a greater impact on user decision-making (such as price, brand, etc.). The features processed by the attention mechanism are then input into the graph neural network for propagation and learning, enabling the graph neural network to more efficiently utilize key information.

[0020] In addition, the system establishes an information sharing module, enabling key features processed by the attention mechanism to be transmitted to the graph neural network in real time, and new feature association information discovered by the graph neural network during propagation can also be fed back to the attention mechanism, dynamically adjusting its focus. For example, if the graph neural network discovers in the propagation process that a new attribute of a certain type of item is highly relevant to user preferences, it feeds this information back to the attention mechanism, causing the attention mechanism to pay more attention to this attribute feature in subsequent processing. The graph neural network propagates and learns information based on the dynamically adjusted information propagation path and node propagation weight, and finally analyzes the target recommendation task to output the recommendation result.

[0021] For example, in the medical health field, a personalized treatment plan recommendation system for tumor patients is based on an enhanced graph neural network, which integrates multiple sources of information, including patient genetic data, tumor staging, medical history, doctor expertise, and past cases. In the graph structure, nodes cover patients, tumor types, treatment plans, and doctors, and edge weights are calculated based on pathological matching degree, plan remission rate, and doctor success case proportion. The system has a dynamic updating mechanism that can adjust the associated edges in real time when new genetic test results are added for patients. The attention mechanism focuses on key features such as gene mutation types, while the graph neural network propagates high-order relationships such as "response rate of patients with the same mutation type to a certain targeted drug." The model training uses a hybrid loss function that takes into account plan matching accuracy, safety (avoiding contraindications), and novelty (including new therapies), and optimizes parameters through the Adam algorithm. Ultimately, the system can recommend precise and safe personalized treatment plans for patients, effectively improving treatment efficiency.

[0022] In the field of financial technology, personal credit risk assessment and product recommendation system integrates user basic information, credit record, transaction flow, social association and credit product characteristics to construct a graph structure. The nodes include users, credit products and credit nodes, and the edge weights are calculated based on income stability, overdue records and product pass rate. The system has a dynamic adjustment mechanism, which updates the weights in real time when new transaction records are added. The attention mechanism focuses on recent repayment ability characteristics, and the graph neural network mines associations such as "product selection preference of high-credit people in user social circle". The model training uses a hybrid loss function to balance evaluation accuracy, risk control (reduce bad debts) and product diversity. After grid search and Bayesian optimization parameters, it can accurately assess user credit risk and recommend suitable credit products, effectively improving the pass rate while reducing the default rate.

[0023] In one embodiment, the step S100 comprises: S101, acquiring user basic information and user behavior sequence, and quantitatively processing the user behavior sequence to generate user behavior characteristics; S102, acquiring item basic attributes and potential characteristics, analyzing the item basic attributes and potential characteristics, and extracting item key characteristics; S103, acquiring user social information, and counting the user interaction frequency in the user social information; S104, acquiring item association information, distributing the item association information, and generating same-type item attribute information and non-same-type item co-occurrence records.

[0024] In this embodiment, through multi-dimensional data collection and processing, the foundation is laid for constructing an enhanced graph neural network graph structure to improve the capture ability of the recommendation system for user preferences.

[0025] The basic attributes provided by the user during registration, such as age, gender, region, occupation, etc. structured information, are collected as the basic attributes of the user node. At the same time, the full interaction behavior of the user in the platform is collected, including browsing, clicking, collecting, adding to the shopping cart, purchasing, commenting, sharing, etc. behavior, and recording the time, duration, frequency, etc. auxiliary information of behavior occurrence. By converting unstructured behavior sequence into quantifiable features, user node attributes are integrated. For example, calculate the proportion of a single behavior in total behavior, such as "browsing time proportion" "purchase conversion rate"; compare a certain type of behavior with other behaviors, such as "browsing frequency ratio" "high-frequency behavior proportion"; combine the time of behavior occurrence, such as "recent behavior weight" "behavior interval duration". Taking an e-commerce platform user as an example, if he / she has repeatedly browsed "laptop" but has not purchased, it can be quantified as "laptop browsing time proportion = 40%" "laptop browsing frequency ratio = 3:1" "laptop purchase conversion rate = 0" etc. features as the core attributes of the user node.

[0026] Collect the direct social relationship of users, such as friend list, follow / followed relationship, and interaction behavior data between users, including chat record times, joint activity times, mutual evaluation / mutual like times, and common browsing / purchasing same item times, etc. Count the total number of interactions between users within a unit time, such as "the chat times between user A and user B = 20 times" and "the joint activity times = 5 times". Add up the number of times of all interaction behaviors to obtain the total interaction times I ij (user i and user j's interaction times). Assign a social edge weight based on the interaction frequency, and the formula is: , wherein represents the edge weight between user i and user j, represents the interaction times of user i and user j, and the denominator is the total interaction times of user i with all other users. For example, user A interacts with user B 15 times (chat 10 times + joint order 5 times) in the past 30 days, and the total interaction times of A with other users is 60 times, then the social edge weight w AB between A and B is 15 / 60 = 0.25, and if the interaction times between A and C is 5 times, then w AC = 5 / 60 ≈ 0.08, indicating that the social relationship between A and B is closer.

[0027] Collect the inherent structured attributes of the items, such as the brand, price, material, size of a commodity, the author, publisher, and publication time of a book, and the director, actor, and release time of a movie. At the same time, collect unstructured information of the items, such as commodity detail page description, user reviews, book synopsis, and movie plot introduction, etc., and mine hidden attributes through technical means. Extract core features that have a significant impact on user decision-making from basic attributes, such as "brand (whether it is a well-known brand)" "price range (high / medium / low)" "material (such as cotton / ramie / chemical fiber)" of e-commerce commodities, "genre (novel / popular science / history)" "page range" of books, etc. Analyze unstructured information through natural language processing (NLP) technology to extract abstract features, such as "style (serious / relaxed)" "theme tendency (science fiction / realism)" "reader evaluation keywords (such as 'logical rigor' 'dragging plot')" of books, and "narrative style (suspense / comedy)" "emotional tendency (healing / depression)" "core theme (family / romance)" of movies. For example, in book recommendation, through NLP analysis of the synopsis and reviews of "The Three-Body Problem", the potential features "hard science fiction style" "macroscopic world view" "reader evaluation mainly with'shocking' 'brain-burning'" are extracted, which together with the basic attributes "author = ***" "publication time = 2008" are used as the key features of the book node.

[0028] The correlation data between the collected items includes the attribute similarity of the same type of items and the co-occurrence behavior of different types of items. For the same type of items, the attribute similarity is calculated by methods such as cosine similarity, Euclidean distance, etc. For example, the similarity of parameters such as "memory", "processor", "price" of a mobile phone, and the similarity of feature vectors (such as the "style tag" vector similarity of two dresses) are calculated. Finally, a "same type of item similarity matrix" is generated. For different types of items but with correlation, the co-occurrence frequency is counted, the interaction behavior of all users to the items is recorded, the item pairs that are "browsed at the same time", "added to the shopping cart at the same time", and "purchased at the same time" are screened, the co-occurrence times are calculated, and the co-occurrence weight is calculated combined with the total number of occurrences of the items: . Wherein represents the edge weight between item A and item B, represents the co-occurrence times of item A and item B, represents the total number of occurrences of item A, represents the total number of occurrences of item B. For example, the co-occurrence times of item a (mobile phone) and item b (mobile phone case) = 500 times (purchased at the same time), C a = 2000 (total number of times the mobile phone is purchased), C b = 1000 (total number of times the mobile phone case is purchased), then the co-occurrence weight , indicating that the two are closely related.

[0029] Through the above processing, the attributes of user nodes, item nodes, user social edges, and item correlation edges are quantified and structured, laying a foundation for subsequent construction of enhanced graph neural networks, ensuring that the graph contains comprehensive and accurate correlation information, thereby improving the ability of the recommendation system to capture user preferences.

[0030] In the medical and health field, for diabetic patients, precise quantitative features can be constructed by integrating multiple sources of information to provide comprehensive input for subsequent graph structure construction. Specifically, these multiple sources of information include the patient's personal basic information (such as age, weight, genetic history), behavior sequence (such as blood glucose monitoring frequency, insulin injection record, diet and exercise log), medical record (such as past complications, medication history, outpatient follow-up record), and social association (such as blood glucose control experience sharing with patients of the same disease, doctor-patient interaction frequency). Through analysis and transformation of these information, quantitative features such as "blood glucose fluctuation amplitude", "medication adherence score", and "patient experience similarity" can be generated, which can provide strong support for subsequent graph structure construction, similar to the logic of converting user behavior sequences into features.

[0031] In the field of financial technology, e-commerce platforms face massive amounts of goods and complex user behavior, with the problem of sparse user and product interaction data being particularly prominent. To address this issue, the present scheme constructs a comprehensive graph structure by integrating multiple user behaviors (such as browsing, purchasing, and collecting) and multi-dimensional information of goods (such as brand, category, price, and associated goods). Using an adaptive propagation mechanism, the system can accurately capture users' potential preferences for different goods. For example, for a user who recently browsed laptops, the system will not only recommend other laptops of the same price and brand, but also recommend appropriate accessories based on the strong association between laptops and accessories such as mice and computer bags, as well as the evaluations of related accessories by the user's friends. At the same time, the system will also take into account the recommendation of niche brand laptops that the user may be interested in, thereby improving the accuracy and diversity of recommendations.

[0032] In one embodiment, the step S200 comprises: S201, fusing the user basic information and user behavior characteristics to generate a user node feature vector; S202, fusing the item basic attributes and item key features to generate an item node feature vector; S203, constructing node attributes based on the user node feature vector and the item node feature vector; S204, calculating social edge weights according to the user interaction frequency; S205, calculating item association edge weights according to the same type of item attribute information and non-same type of item co-occurrence records; S206, constructing association edge weights from the social edge weights and the item association edge weights; S207, constructing a graph structure from the node attributes and the association edge weights; S208, continuously obtaining new user behavior information and updating the graph structure according to the new user behavior information.

[0033] In this embodiment, by constructing a dynamically updated graph structure, multi-source information is fused to provide accurate and real-time basic data support for information propagation of the enhanced graph neural network, thereby improving the accuracy and adaptability of the recommendation system. The construction of the user node feature vector integrates the user basic information and the quantized behavior characteristics. For structured basic information, numerical information (such as age) directly retains the original value and is mapped to the [0, 1] interval through Min-Max standardization processing to avoid the influence of magnitude difference on model learning; categorical information (such as gender, region) is converted into a binary vector using one-hot encoding, for example, gender "male" is encoded as [1, 0] and "female" is encoded as [0, 1], and the region "Beijing" corresponds to a vector with the "Beijing" dimension being 1 and the rest being 0. The quantized behavior characteristics (such as "browsing time proportion", "purchase conversion rate", "recent behavior weight", etc.) need to be processed consistently, and all behavior characteristics are adjusted to the same distribution through Z-Score standardization to avoid weight imbalance caused by magnitude difference between features. For time series features (such as "recent 7-day behavior proportion"), a higher priority is given, which can be enhanced in the feature vector by setting a weight coefficient (such as multiplying by 1.2). Finally, the preprocessed basic information vector and the integrated behavior characteristic vector are concatenated by dimension to form the final user node feature vector. For example, a user's basic information after encoding is [0.3 (age after standardization), 1 (male), 0 (non-Beijing), 1 (Shanghai)], and the behavior characteristic vector is [0.4 (browsing time proportion), 0.1 (purchase conversion rate), 0.7 (recent behavior weight)], then the fused user node feature vector is [0.3, 1, 0, 1, 0.4, 0.1, 0.7].

[0034] The construction of the item node feature vector fuses basic attributes and extracted key features (including latent features). For structured basic attributes, numerical attributes (such as price) are converted into numerical features through normalization processing; categorical attributes (such as brand, material) are converted into low-dimensional vectors using one-hot encoding or embedding technology, for example, "brand = Apple" corresponds to the embedding vector [0.2, 0.5, 0.3]. The latent features of an item (such as the "style" and "theme tendency" of a book, the "narrative style" of a movie) are converted into vectors through natural language processing (NLP) technology: after tokenization and stop word removal on unstructured text (such as product descriptions, book reviews), a bag-of-words model (BoW) or a pre-trained language model (such as BERT) is used to generate a text vector; the extracted latent features (such as "style = serious" and "theme = science fiction") are converted into probability distribution vectors through topic modeling (such as LDA), for example, "science fiction theme probability = 0.8". Finally, the preprocessed basic attribute vector and the latent feature vector are concatenated to form the item node feature vector. For example, the basic attribute vector of a notebook computer is [0.6 (price after standardization), 1 (brand = Huawei), 0 (non- Lenovo), 0.8 (memory = 16G)], the latent feature vector is [0.7 (business style probability), 0.2 (game performance score)], and the fused item node feature vector is [0.6, 1, 0, 0.8, 0.7, 0.2].

[0035] Node attributes are the core information of "user nodes" and "item nodes" in the graph structure, which are directly defined by the feature vectors generated above. User node attributes include all dimensions of user basic information (after encoding) and behavior features (after quantization and standardization); item node attributes include all dimensions of item basic attributes (after encoding) and key features (after vectorization of latent features). Node attributes need to be stored in matrix form, with each row corresponding to a node's feature vector, ensuring that the graph neural network can be directly read and used for information propagation.

[0036] Social edges connect user nodes, and the weight is quantified from user interaction frequency. Count all interaction behavior times between users within a unit time (such as the last 30 days), including chatting, participating in activities together, mutual evaluation, mutual praise, etc., and accumulate to get the total interaction times interactij (the interaction times between user i and user j). The weight of the social edge is the proportion of the interaction times between user i and j to the total interaction times of user i. If user i has no social interaction, the weight of its social edge is set to 0 (i.e., it does not establish a social edge with other users).

[0037] The item association edge connects the item nodes, and the weight needs to distinguish "same type of items" and "non-same type of items" for separate calculation. The same type of items refers to items belonging to the same category (such as "smartphone" and "dress"), and the weight is calculated by attribute similarity. The final weight of the same type of item association edge is the weighted average of structured and unstructured similarity (the weight coefficient is set according to the importance of the attribute, such as structured attribute accounting for 0.6 and unstructured accounting for 0.4). Non-same type of items refers to items of different categories but with association (such as "mobile phone" and "mobile phone case"), and the weight is calculated by co-occurrence frequency.

[0038] The association edge is the core of connecting nodes in the graph structure, including three types of edges, and the weights are defined as follows: user-user edge, i.e. social edge weight, determined directly by the above social edge weight calculation formula; item-item edge, i.e. item association edge weight, same type of items use attribute similarity, and non-same type of items use co-occurrence weight; user-item edge, based on the user's interaction behavior (such as browsing, purchasing) to set the initial weight, the rule is: purchase behavior weight (such as 1.0) > collection / adding to shopping cart (such as 0.7) > browsing (such as 0.3), and dynamically adjusted according to the interaction frequency (such as multiple purchases weight increased to 1.2).

[0039] The graph structure is composed of "nodes" and "association edges", the user node set contains all users, and the attribute of each user node is its feature vector; the item node set contains all items, and the attribute of each item node is its feature vector. The user-user edge connects users with social relationship, and the edge weight is the social edge weight; the item-item edge connects associated items, and the edge weight is the item association edge weight; the user-item edge connects users and items with interaction behavior, and the edge weight is the initial weight corresponding to the interaction behavior. It is stored in the form of adjacency matrix or edge list, the row / column corresponds to the node, and the matrix value is the edge weight (0 if not connected), which is suitable for small-scale graphs; all edges are recorded in triple (source node ID, target node ID, weight), which is suitable for large-scale graphs (such as million-level nodes of e-commerce platform).

[0040] Real-time monitoring of new user behaviors, including user's new browsing, purchasing, collecting, and other interactive behaviors; new item information, such as new item attributes (such as price adjustment, specification update); new social relationships, such as user's new friends, and changes in interaction frequency. User-item edge update, if a user adds a new interactive behavior, immediately add or adjust the weight of the corresponding edge, for example, if a user newly purchases item A, the user-item A edge weight is set to 1.0, and if there is a browsing record before, the purchase weight is overwritten; item-item edge update, if new interaction leads to changes in co-occurrence frequency of items (such as a user simultaneously purchasing items A and B), the co-occurrence weight of A and B is recalculated and the item-item edge is updated; user-user edge update, if a user adds a new interaction (such as an increase in chat frequency), the social edge weight is recalculated and updated; node attribute update, when user behavior characteristics change (such as new purchases), the user node feature vector is regenerated; when the item information is updated, the item node feature vector is regenerated. The "real-time triggering + batch optimization" mode is adopted: single new behavior triggers immediate update, for example, immediately update the user-item edge after purchase, and perform batch calibration on the full edge weight every morning, such as recalculating the co-occurrence weight based on all new data on the day.

[0041] Through the above steps, the graph structure can realize comprehensive fusion of multi-source information and dynamically adjust with data, providing accurate and real-time basic data support for information propagation of subsequent enhanced graph neural networks, and ultimately improving the accuracy and adaptability of the recommendation system.

[0042] For example, in the medical health field, precise matching of treatment plans for tumor patients can be achieved by constructing a graph structure. Specifically, the graph structure includes user (patient) and item (tumor type, treatment plan, doctor) nodes, and associated edges between them. Node settings include patients (including genetic data, tumor stage), tumor types (such as lung adenocarcinoma, squamous carcinoma), treatment plans (chemotherapy, targeted drugs, immunotherapy), and doctors (specialty area, past cases). The calculation of edges and weights is based on multiple factors: the edge weight between the patient and the tumor type is based on the degree of pathological diagnosis matching; the edge weight between the tumor type and the treatment plan is based on the remission rate of the same type of tumor to the plan, for example, if a targeted drug has an 80% remission rate for EGFR-mutated lung adenocarcinoma, the edge weight is 0.8; the edge weight between the doctor and the treatment plan is based on the proportion of successful cases of the doctor using the plan. When a patient has new genetic test results (such as the discovery of ALK fusion mutation), the system will immediately update the "patient-tumor type" edge weight and add an associated edge with the "ALK inhibitor" treatment plan, similar to the mechanism of "new user behavior triggering graph structure update".

[0043] In a social platform, the social relationship between users has an important influence on content dissemination, and the interests of users are also diverse. Therefore, the scheme combines the user's social relationship (such as friends, attention, interaction frequency), the published content (such as text, pictures, videos), and the browsing and like records and other information. Through the propagation mechanism of the graph neural network, the system can efficiently deliver high-quality content that the user's friends like to the user. At the same time, combined with the attention mechanism, the system can focus on the content theme that the user recently pays attention to. For example, on a social sharing platform, if the user A's friend B recently shares an article about photography skills and gets more likes, the system will recommend the article to A according to the high interaction frequency between A and B. In addition, the system will also recommend content related to the photography equipment that A recently browsed, thereby improving the user's recognition of the recommended content.

[0044] In one embodiment, the step S300 comprises: S301, determining the activity of each node in the graph structure according to the preset measurement rule through the user multi-source information, and generating a node activity; S302, calculating the feature correlation through the user node feature vector and the item node feature vector; S303, evaluating the position of each node in the graph structure through the degree centrality, and generating a node graph structure position; S304, evaluating the importance of each node in the graph structure according to the node activity, the feature correlation and the node graph structure position, and generating a node importance evaluation result.

[0045] In this embodiment, the node importance is a key indicator for evaluating the value of the node in the graph structure, which is used to guide the weight distribution of subsequent information dissemination, thereby improving the accuracy and efficiency of the recommendation system. The evaluation of the node importance integrates the node activity, the feature correlation and the node graph structure position three core indicators.

[0046] Node activity is one of the basic indicators to measure the importance of nodes, which needs to be quantified based on multi-source information according to preset rules for user nodes and item nodes. For user nodes, node activity is based on the number of user behaviors in a preset time window, reflecting the user's participation in the platform. Multi-source information includes all interactive behaviors such as user clicks, views, purchases, collections, comments, and shares. The preset measurement rules include time window selection, usually 30 days (balance recent activity and historical stability), and special scenarios (such as real-time recommendation) can be shortened to 7 days. Behavior weight division, different behaviors contribute differently to activity, assign weights according to importance: purchase (weight 1.0) > collection / comment (0.7) > view / click (0.3) (weights can be adjusted according to business scenarios). For example, a user has purchased 2 times in the past 30 days (1.0x2=2), collected 3 times (0.7x3=2.1), and viewed 10 times (0.3x10=3), so the activity is 2+2.1+3=7.1.

[0047] For item nodes, node activity is based on the number of user interactions with the item and the heat trend, reflecting the popularity of the item. Multi-source information includes the total number of views, purchases, and collections of the item, as well as the change in interaction frequency in the past 7 days / 30 days (heat trend). The preset measurement rules include basic activity, total interaction frequency needs to be standardized to avoid the influence of different item online time, such as "total interaction frequency in the past 30 days / item online days". Heat trend correction, if the interaction frequency in the past 7 days increases by more than 20% compared to the past 30 days, multiply the basic activity by 1.2 (highlight rising trend items); if it decreases by more than 30%, multiply by 0.8 (weaken declining trend items). For example, a certain product has 100 interactions in the past 30 days, and has been online for 50 days, so the basic activity is 100 / 50=2; the interaction in the past 7 days is 30, which is an increase of 30% compared to the average of 3.3 times per day in the past 30 days, and the trend correction coefficient is 1.2, so the final activity is 2x1.2=2.4.

[0048] Feature correlation is used to evaluate the matching degree of node features and target recommendation tasks, which needs to be calculated based on user node feature vectors and item node feature vectors. User node feature vectors include user basic information (age, gender, etc.) and behavior features (browsing preferences, purchase tendencies, etc.) fusion vectors; item node feature vectors include item basic attributes (price, brand, etc.) and key features (style, theme, etc.) fusion vectors. For user nodes, calculate the correlation between their feature vectors and "target recommendation task features"; for item nodes, calculate the correlation between their feature vectors and "user preference feature vectors". The formula for calculating feature correlation is: where: x is the feature vector of the user / item, y is the target feature vector (task feature or user preference feature), , The mean of x and y, respectively; the result range is [-1, 1], and the closer the value is to 1, the higher the feature correlation. For example, the target recommendation task is "recommend fitness equipment", the user node feature vector is "sports commodity browsing proportion = 0.6" and "fitness related collection times = 5", and the target task feature vector is "fitness attribute = 1" and "other attribute = 0". The calculated corr = 0.85 (high correlation), so the importance of the user node in the recommendation is improved.

[0049] The node graph structure position is measured by the degree centrality, which reflects the core degree of the node in the graph network. The degree of the node (deg(v)) is the total number of edges connected to the node v (including user-user edges, user-item edges, and item-item edges). For example, user node A is connected to 3 friends (user-user edges) and 5 interaction items (user-item edges), so deg(A) = 3 + 5 = 8. The total number of nodes (N) is the total number of all user nodes and item nodes in the graph structure. The calculation formula is: The result range is [0, 1], and the closer the value is to 1, the more other nodes the node is connected to, and the more central the node is in the network. For example, the total number of nodes in the graph N = 1000, and item node B is connected to 20 similar items (item-item edges) and 100 interaction users (user-item edges), so deg(B) = 20 + 100 = 120, and the degree centrality C D (B) = 120 / (1000-1)≈0.12, indicating that it is in a relatively central position in the network.

[0050] Node importance is the comprehensive result of the above three indicators, which is used to guide the weight distribution of subsequent information propagation. Due to the different magnitude and range of the three indicators (activity is an absolute number, correlation and degree centrality are [-1, 1] or [0, 1]), it needs to be standardized to the [0, 1] interval first. The activity uses Min-Max standardization, the feature correlation takes the absolute value (since negative correlation has no practical significance) and is standardized, and the degree centrality is already in the [0, 1] interval and does not need additional processing. According to the emphasis of the recommendation task, the three indicators are assigned weights (the weight sum is 1). Through the above steps, the importance of each node in the graph structure can be quantitatively evaluated, providing a core basis for the adjustment of information propagation weight and path in the subsequent "adaptive propagation mechanism" - the information propagation weight of high importance nodes is higher, and the high value path they participate in will be preferentially retained, ultimately improving the capture efficiency of the recommendation system for key information.

[0051] For example, in the medical health field, early warning of potential transmission chains of infectious diseases can be achieved by constructing a graph structure and evaluating node importance. Specifically, the graph structure includes three types of key nodes: patient nodes, virus strain nodes, and medical institution nodes. For infectious diseases such as influenza, the importance evaluation criteria for these three types of nodes are different. The importance of patient nodes is evaluated based on their activity, such as the number of visits in the past 7 days and the records of fever clinics. The higher the activity, the more important the node is in the transmission chain. The importance of virus strain nodes is evaluated based on their feature correlation, i.e., the similarity of their genetic sequences to known epidemic strains. The higher the similarity (e.g., reaching 95%), the more important the node is in the transmission early warning. The importance of medical institution nodes is evaluated based on their position in the network structure, such as their degree centrality, i.e., the proportion of infectious disease patients they treat in the region. The higher the proportion (e.g., 30%), the more important the node is as a key node in the transmission chain. By focusing on these high-importance nodes, precise epidemiological investigations can be provided with clear directions, similar to the logic of "high-importance node information transmission weight higher".

[0052] In the financial technology business field, video platforms provide a variety of content, including movies, TV series, variety shows, etc., and users' preferences for video content are influenced by factors such as directors, actors, types, ratings, etc. This scheme constructs a graph structure, which integrates users' viewing history, collection list, ratings, and information such as directors, actors, types, and associations with other video works (such as series works) of video works. Through an adaptive transmission mechanism, the system can identify valuable video association paths for users. For example, if a user likes a director's work, the system can recommend new movies directed by the director, or recommend works of the same type as a TV series the user has watched and with similar actor casts. This mechanism can help users discover more video content that meets their personal preferences and improve user experience.

[0053] In one embodiment, the step S400 includes: S401, calculating node transmission weight according to the node importance evaluation result, feature correlation, and weight coefficient; S402, identifying high-value paths and low-value paths for the target recommendation task through a path analysis algorithm, in combination with the node importance evaluation result, feature correlation, and node graph structure position; S403, performing enhanced transmission processing on the high-value paths and invalid transmission processing on the low-value paths.

[0054] In this embodiment, in the graph neural network, node transmission weight is a key indicator of the "intensity" of information transmission from node u to node v. Its calculation formula is where S uis the importance score of node u, reflecting the influence of the node in the graph; r(u, v) is the feature correlation of node u and node v, measuring the matching degree of the features of the two nodes; is the weight coefficient, used to balance the influence of node importance and feature correlation. In different recommendation scenarios, the value of is different. For example, in social platform recommendation, if more attention is paid to the “overall influence of users / items”, then a larger value (such as 0.6-0.8) is taken; while in e-commerce product recommendation, if more attention is paid to “feature matching accuracy”, then a smaller value (such as 0.3-0.5) is taken. For example, the importance score S u of node u (user A) is 0.7, the feature correlation r(u, v) of node v (item B) is 0.8, and when = 0.5, the propagation weight = 0.5x0.7+0.5x0.8=0.75, indicating that the propagation strength from user A to item B is high.

[0055] The path is the information propagation channel between nodes in the graph structure, for example, “user A→ item B→ item C→ user D”. Through the path analysis algorithm combined with node importance, feature correlation and graph structure location (degree centrality), the paths valuable for the target recommendation task can be distinguished, and the key information can be propagated preferentially. The path analysis algorithm quantifies the value of the path through “path score”, the higher the score, the higher the value of the path. The path score is calculated based on three dimensions: first, the sum of the importance of the nodes on the path, that is, the accumulation of the importance scores S(•) of all nodes on the path, reflecting the “node quality” of the path as a whole; second, the average of the feature correlation of adjacent nodes on the path, that is, the average of the feature correlation r(•,•) of all adjacent node pairs on the path, reflecting the “feature matching stability” of information propagation in the path; third, the degree centrality of the core node of the path, screening the “node with the highest degree centrality” (core node) in the path, and its degree centrality reflects the “core position influence” of the path in the graph structure.

[0056] By the path analysis algorithm, combined with the node importance evaluation results, feature correlation and node graph structure position, high-value paths and low-value paths of the target recommendation task can be identified. The path score of the high-value path is higher than the preset threshold (such as the top 20% of the path), and usually satisfies the following conditions: containing multiple high-importance nodes (such as high-activity users, popular items); adjacent node feature correlation is strong (such as high matching between user preferences and item attributes); passing through a high-degree central node (such as a popular item connecting a large number of users). The path score of the low-value path is lower than the preset threshold (such as the bottom 30% of the path), and usually shows the following characteristics: low node importance (such as low-activity users, unpopular items); weak feature correlation (such as low matching between user historical preferences and item attributes in the path); low-degree central node (the path is located in the edge area of the graph structure).

[0057] According to the evaluation results of the path value, different propagation strategies are adopted for high-value paths and low-value paths to optimize the information propagation efficiency and avoid invalid information interference with the recommendation results. For high-value paths, the propagation strength is improved and the propagation range is extended to allow the information of high-value paths to be more fully transmitted. Specific measures include: increasing the propagation weight, multiplying the propagation weight w_{uv} of all adjacent nodes on the high-value path by an enhancement coefficient (such as 1.2 1.5), amplifying the information transmission strength; extending the propagation steps, allowing information to propagate more than 1 2 steps on the high-value path (such as 3 steps for ordinary paths and 4 5 steps for high-value paths), to mine deeper associated information (such as “user A→ item B → item C → item D” can be extended to “item E”); preferentially propagating markers, preferentially aggregating information transmitted by high-value paths during the information aggregation stage of the graph neural network, and increasing their weight in node feature updating. For low-value paths, the propagation strength is reduced or directly truncated to reduce the interference of low-value paths on the recommendation results. Specific measures include: reducing the propagation weight, multiplying the propagation weight w_{uv} on the low-value path by a decay coefficient (such as 0.3 -0.5), weakening the information transmission strength; truncating the propagation path, directly truncating the propagation at the intermediate node for paths with extremely low scores (such as the bottom 10%), such as “user X → item Y → item Z” stopping propagation at item Y and not propagating to item Z; filtering invalid information, ignoring information transmitted by low-value paths during the information aggregation stage, or only retaining a very small proportion (such as 5%) for diversity supplementation to avoid dominating the recommendation results.

[0058] Through the above steps, the information propagation mechanism of the graph neural network realizes "precision" and "differentiation". The information of the high-value path is strengthened and propagated to ensure that the user's potential interest is fully captured; the invalid information of the low-value path is suppressed to reduce noise interference. This mechanism directly improves the utilization efficiency of the recommendation system for complex correlation relationships, provides better input for the learning of the subsequent hybrid model architecture, and ultimately achieves more accurate and personalized recommendations.

[0059] For example, in the medical health field, dynamically adjusting the information propagation path and the node propagation weight can achieve personalized antibacterial drug recommendations. Specifically, the propagation weight is adjusted according to the node importance, and the high-value path is strengthened and the low-value path is truncated. For example, when recommending antibacterial drugs for a patient with a bacterial infection, the initial propagation path is "patient (symptom: pneumonia) → pathogen (Klebsiella pneumoniae) → antibacterial drug (ceftazidime, meropenem)". If the patient has a new test result of "renal insufficiency", the "patient-drug feature correlation" will decrease in the node importance evaluation, so the propagation weight of "meropenem (metabolized by the kidney) → patient" is reduced from 0.7 to 0.3. If it is detected that the pathogen is resistant to ceftazidime, the "pathogen-ceftazidime" feature correlation will drop sharply, and the system will directly truncate this path and strengthen the propagation strength of the "pathogen → meropenem (sensitive)" path, with the weight increasing from 0.5 to 0.9. By avoiding invalid or high-risk recommendations in real time, this mechanism is similar to "dynamic evaluation and adjustment of path value".

[0060] In a music platform, there are various song styles, and user listening preferences are easily influenced by factors such as singers, song styles, and lyric themes, and there are a large number of long-tail songs. This solution constructs a graph structure by fusing user listening records, collected singers, created playlists, and information such as singers, albums, song styles, and similar songs (such as songs with the same style or often played together). Using a hybrid model architecture, the system can recommend songs that meet the user's taste, including some niche but high-quality songs. This recommendation mechanism not only meets the user's demand for mainstream music, but also helps users discover more personalized music content, improving the user's listening experience.

[0061] In one embodiment, the step S500 includes: S501, inputting the user node feature vector and the item node feature vector into an attention mechanism for preprocessing, and outputting user attention weights and item attention weights; S502, inputting the user attention weights and the item attention weights into a graph neural network for processing, and outputting new feature key information; S503, feeding back the new feature key information to the attention mechanism for dynamic adjustment of the focus point; S504, forming a collaborative processing model with the attention mechanism and the graph neural network; S505, inputting the target recommendation task into the collaborative processing model, collaboratively processing the target recommendation task, and outputting a recommendation result.

[0062] In this embodiment, the core role of the attention mechanism is to filter out key features highly related to the recommendation task from the user node feature vector and the item node feature vector. By calculating the "attention weight", important features are highlighted and secondary features are weakened, thereby providing more efficient input for subsequent graph neural network processing. The user node feature vector integrates user basic information (such as age, gender) and behavior features (such as browsing time proportion, purchase conversion rate), while the item node feature vector integrates item basic attributes (such as price, brand) and key features (such as potential style, theme). The calculation of user attention weight focuses on behavior features that are strongly related to the user's "recent interest" or "core preference". For example, if a user frequently browses "wireless earphones" in the past 7 days, features such as "wireless earphone browsing proportion" and "audio device purchase times" will be given high weights, while historical features unrelated to recent interest (such as purchase records half a year ago) will be weakened. The calculation of item attention weight focuses on item attribute features that have a significant impact on user decision-making. For example, in the e-commerce scenario, features such as "price range", "brand popularity", and "user rating" have a greater impact on purchase decisions, so they will be given high weights; while in book recommendation, features such as "theme matching degree" and "author popularity" have higher weights. After preprocessing, the user attention weight vector (each dimension corresponds to the weight of the user feature) and the item attention weight vector (each dimension corresponds to the weight of the item feature) are output. In the subsequent step, the feature vector will be multiplied by the corresponding attention weight to obtain a "weighted feature vector", thereby highlighting high-weight features.

[0063] The graph neural network receives the user / item feature vectors weighted by the attention mechanism, and through information propagation and aggregation between nodes, it mines the deep relationships between users and items, users and users, and items and items, and outputs newly discovered feature key information. The input includes the weighted user feature vector (user feature vector x user attention weight), the weighted item feature vector (item feature vector x item attention weight), and the edge weight of the graph structure (the weight of the user-user, user-item, and item-item edge). Multi-layer graph convolution or graph attention network (GAT) is used for information propagation: each node (user / item) aggregates the features of its neighbor nodes (e.g., a user node aggregates the features of its interaction items and social friends), and the edge weight is referred to during aggregation (the neighbor information of high weight edges has a greater impact). The node features are updated through a nonlinear activation function (such as ReLU) to capture high-order correlations (e.g., the correlation between "items liked by a user's friends" and "the user's potential interests"). The "new feature key information" output by the graph neural network includes the "latent preference features" of the user nodes (e.g., interests that are not explicitly expressed by the user but inferred through correlation, such as a user who likes "wireless earphones" may prefer "noise reduction function"), the "hidden association features" of the item nodes (e.g., the association between items that is not directly labeled, such as the compatibility between "a certain notebook" and "a specific model of mouse"), and the "dynamic matching features" of users and items (e.g., the change in user preference for an item in different time windows).

[0064] The bidirectional interaction of the graph neural network and the attention mechanism is realized through the "information sharing module", the key information of new features discovered by the graph neural network is fed back to the attention mechanism, the focus of the attention mechanism on the features is dynamically adjusted, and a closed loop of "preprocessing-propagation-feedback" is formed. The information sharing module receives the key information of new features output by the graph neural network in real time, analyzes the correlation related to the user / item features (for example, "the matching degree of certain item attribute X and user preference is significantly improved"). For the user attention mechanism, if the graph neural network finds that "the item category that the user has not frequently browsed recently but is highly recommended by friends" is strongly related to the user preference, the weight of the behavior feature corresponding to the category (for example, the feature weight of "the number of times of browsing the item mentioned in the friend interaction") is increased. For the item attention mechanism, if the graph neural network finds that "the potential feature Y (such as "environmentally friendly material") of a certain item has increased influence on user decision-making", the attention weight of the potential feature in the item feature vector is increased. Initially, the item attention mechanism may pay more attention to explicit features such as "price" and "brand", but the graph neural network discovers through propagation that "among the users who purchase the item, 80% pay attention to 'after-sales service score' at the same time", and after the key information of new features is fed back to the attention mechanism, the attention mechanism dynamically increases the weight of the "after-sales service score" feature, and will focus more on this feature in subsequent preprocessing. Through the cooperative mode of "preprocessing-propagation-feedback", the two form a unified model, and the core is to realize the complementation of "key feature focusing" and "deep association mining" through bidirectional information interaction.

[0065] The specific target recommendation task (such as "e-commerce platform recommends goods to users" and "video platform recommends video content to users") is input into the collaborative processing model. The model focuses on key features combined with attention mechanisms and captures deep associations through graph neural networks, outputting accurate, diverse, and novel recommendation results. The model first extracts scene features of the target task (such as "during the promotion period" and "first-time user login"), and determines the core target of the recommendation (such as "improve purchase conversion rate" and "increase content browsing time"). The attention mechanism is based on the task target and the current focus point, and the user / item feature vector is preprocessed again to generate weighted features that adapt to the task. The graph neural network is based on weighted features and graph structure to mine association paths that match the task target (for example, "in the promotion scenario, the user's joint preference for items with high 'discount strength' and friend-recommended items"). Combined with a hybrid loss function (including accuracy, diversity, and novelty loss), the final recommendation result (such as Top-N goods / content list) is output. For example, the target task is to recommend goods to a user who has browsed laptops for the past 30 days but has not purchased. The attention mechanism focuses on key features such as "browsing time proportion", "price interval preference", and "friend purchase record". The graph neural network mines new features such as "the user browsed laptop models and the adaptation association of a certain mouse" and "small niche brand with rising recent evaluation rate in the same price interval". The collaborative processing output recommendation result contains mainstream laptops of the same type, adaptive mice, and high-performance small niche brand laptops, taking into account accuracy, diversity, and novelty. The collaborative processing model of attention mechanism and graph neural network realizes accurate capture of user preferences and item associations through the process of "focusing on key features-mining deep associations-dynamic feedback adjustment". This model not only solves the problem of inefficient feature utilization and insufficient association mining in traditional recommendation systems, but also adapts to data changes through dynamic adjustment mechanisms to provide high-quality recommendation results for target recommendation tasks.

[0066] For example, in the medical health field, the attention mechanism cooperates with the graph neural network to process personalized recommendations for prenatal examination projects. Specifically, the attention mechanism can focus on key features and interact with the graph neural network in real time for information exchange and collaborative optimization. When recommending prenatal examination projects for pregnant women, the attention mechanism focuses on recent key features, such as "abnormal glucose tolerance test results" and "reduced fetal movement frequency" at 24 weeks of pregnancy, and assigns high attention weights to these features (e.g., 0.8 for abnormal glucose tolerance and 0.2 for regular blood pressure). The graph neural network then propagates information based on these focused features in the path "pregnant woman → pregnancy complication (gestational diabetes) → examination project (glycated hemoglobin, fetal ultrasound)." When the graph neural network discovers a strong correlation between "glycated hemoglobin detection and fetal movement abnormalities," it feeds this information back to the attention mechanism, allowing it to further increase the weight of the "fetal movement monitoring" feature in subsequent processing. This real-time information exchange and focus adjustment logic between models can more accurately recommend necessary prenatal examination projects for pregnant women.

[0067] In the financial technology business field, the attention mechanism focuses on dynamic features related to repayment ability for loan applicants. For example, the stability of the past 6 months' salary stream (weight 0.6) and credit card overdue times (weight 0.3) are given priority, while secondary features such as "industry label" in the occupation type (weight 0.1) are weakened. A graph structure is constructed for "user-credit product-credit node-social relationship," where user nodes are associated with their income, debt, and other features, and credit product nodes are associated with interest rates, limits, and overdue rates. The graph neural network mines high-order relationships during propagation, such as "products jointly applied for by the user and high-credit users in their social circle" and "the matching degree of repayment performance of users in the same income level to a certain product." When the graph neural network discovers that "the pass rate of a certain low-interest product in the user's company is significantly higher than the industry average," it feeds this correlation information back to the attention mechanism, temporarily increasing the weight of the "enterprise attribute" feature (from 0.1 to 0.3). Finally, the system outputs the recommendation result, prioritizing the two products with the highest matching degree (e.g., "white-collar exclusive loan" and "public accumulation fund credit loan"), and providing the recommendation reasons (e.g., "matching degree 92% with your income and credit, pass rate 85% with users in the same industry").

[0068] In one embodiment, the step S504 further comprises: S5041, obtaining a historical cross-domain data set and preprocessing the historical cross-domain data set; S5042, inputting the preprocessed historical cross-domain data set into a collaborative processing model for training, and outputting a recommendation result; S5043, evaluate the recommendation result by a mixed loss function; S5044, optimize the collaborative processing model by an adaptive gradient descent algorithm; S5045, adjust the parameters of the collaborative processing model by grid search and Bayesian optimization.

[0069] In this embodiment, the historical cross-domain data set is the basis for model training, and needs to cover multiple industry scenarios to improve the generalization ability of the model. Preprocessing ensures data quality and consistency. The acquisition range of the historical cross-domain data set is "large-scale and diversified", covering core data in different fields (cross-domain), including industry fields such as e-commerce (user-goods interaction, goods attributes), video (user-movie interaction, movie features), music (user-song interaction, song style), social networking (user social relationship, content sharing record), etc.; data types such as user behavior data (browsing, purchasing, watching, collecting, etc.), item information data (basic attributes, latent features), social relationship data (friend list, interaction frequency).

[0070] Preprocessing aims to remove noise, fill in missing data, and unify formats to ensure that data can be directly input into model training. Specifically, data cleaning removes noise data, filters misoperation records (such as invalid clicks with single browsing duration <1 second), and removes outliers (such as abnormal behavior of a user purchasing the same product 100 times in a day). Duplicate data is deleted, and duplicate interaction records of the same user on the same item are merged (such as a user browsing the same product multiple times, keeping the latest timestamp and cumulative duration). Missing data is filled using an "interpolation method based on similar node features", where if a user's age is missing, the average age of the 10 users with the most similar behavior characteristics (such as purchase preferences, browsing categories) is calculated as the fill-in value; if an attribute of an item (such as the "theme" of a book) is missing, the missing value is inferred based on the latent features of similar items (such as the theme distribution of similar books). Data standardization and normalization, numerical features (such as browsing times, price) are mapped to the [0,1] interval using Min-Max standardization, and categorical features (such as gender, item category) are converted into low-dimensional vectors using one-hot encoding or embedding to avoid misleading the model due to the "numerical size" between categories. Cross-domain feature alignment unifies the feature dimensions of different domains (such as standardizing "purchase times" in e-commerce and "view times" in video to the same order of magnitude), ensuring that cross-domain data can be trained together.

[0071] The synergistic processing model (attention mechanism + graph neural network) learns general recommendation rules by training across-domain data and outputs preliminary recommendation results. The pre-processed cross-domain data needs to be converted into a format recognizable by the model, including node feature matrix, user node feature vector (including cross-domain behavior features such as e-commerce purchase + video viewing preferences), item node feature vector (including cross-domain attributes such as product brand + film type); graph structure data, cross-domain association edges (such as the fusion of user friends on e-commerce and video platform attention relationships, cross-domain co-occurrence edges such as "users who purchase projectors often watch high-definition videos"); label data, actual interaction results of users on items (such as "purchase" marked as 1, "non-purchase" marked as 0) for loss function calculation. The attention mechanism preprocesses cross-domain node features, focusing on key features in each domain (such as "price sensitivity" in the e-commerce domain and "type preference" in the video domain). The graph neural network propagates information based on cross-domain graph structure to capture cross-domain associations (such as "users who like 'classical music' on the music platform may prefer'retro style goods' on the e-commerce platform"). The model learns through multiple iterations to gradually optimize node feature representation and association edge weights. For each user, the model outputs a cross-domain or single-domain recommendation list (such as recommending "classical music albums + retro earphones" to the user), sorted by predicted scores (user preference probabilities).

[0072] The hybrid loss function evaluates the recommendation quality from accuracy, diversity, and novelty, guiding the model optimization direction. The hybrid loss function is designed to include recommendation accuracy loss (such as cross-entropy loss, used to measure the matching degree of the recommendation result and the user's actual behavior), diversity loss (encouraging the recommendation result to cover different types of items to avoid excessive concentration), and novelty loss (promoting the recommendation of a certain proportion of items that the user has not contacted but may be interested in). The Adam algorithm (adaptive moment estimation) is used to dynamically adjust the model parameters, improving training efficiency and convergence speed. The adaptive learning rate dynamically adjusts the learning rate for different parameter gradients (large gradient parameters have small learning rates to avoid oscillation; small gradient parameters have large learning rates to accelerate convergence); the momentum mechanism combines first-order moments (momentum) and second-order moments (exponential moving average of gradient squares) to reduce training fluctuations and stabilize the convergence path. By minimizing the hybrid loss function L, the model parameters (such as the weight matrix of the graph neural network and the feature score network parameters of the attention mechanism) are iteratively updated until the loss function converges to a stable value.

[0073] Parameter tuning is crucial to improve model performance, and requires efficient combination of two methods to find the optimal parameter combination. For each parameter, preset discrete candidate values (such as the number of layers of graph neural networks: [2, 3, 4]; the number of attention heads: [2, 4, 6, 8]), traverse all combinations (grid), calculate the validation set loss L corresponding to each parameter, and filter out the parameter range with lower loss (such as the optimal range of the number of layers of graph neural networks is 3-4 layers). Based on the range determined by grid search, the parameter optimization is regarded as a "black box function" (input parameter combination, output loss L), the loss of untried parameters is predicted by Gaussian process model, and the parameter combination that "may reduce the loss" is preferentially selected for verification, and the iteration is optimized until the parameter combination corresponding to the minimum loss is found. Using historical results for prediction can reduce the amount of calculation and is suitable for precise tuning; for example, within the range of 3-4 layers of graph neural networks and 4-6 heads of attention heads, the optimal combination (such as 3 layers of graph neural networks + 6 heads of attention) can be quickly located.

[0074] Through preprocessing of cross-domain datasets, high-quality and diverse training data is provided for the model; the collaborative processing model learns the cross-domain correlation rules during training and outputs the recommended results; the hybrid loss function evaluates the quality of the results from multiple dimensions; the Adam algorithm efficiently optimizes the model parameters; and the combination of grid search and Bayesian optimization realizes precise parameter tuning. This process ensures that the model maintains excellent recommendation performance in different domains, balancing precision, diversity, and adaptability.

[0075] For example, in the medical health field, a rare disease diagnosis and treatment plan recommendation system is trained by fusing multiple data sets, uses a hybrid loss function and an adaptive optimization algorithm to improve the accuracy and adaptability of recommendations. Specifically, the system fuses multi-center rare disease cases (such as Fabry disease), gene databases, international diagnosis and treatment guidelines, and doctor consultation records, etc. In the data preprocessing stage, noise data (such as misdiagnosed cases) is cleaned up, and missing gene test results are filled in (based on similar case interpolation). The loss function includes accuracy loss (the matching degree of diagnosis and treatment plan and patient condition, such as the effectiveness of enzyme replacement therapy for patients with specific gene mutations), safety loss (avoiding high-risk plans, such as contraindicated drugs for patients with renal dysfunction), and novelty loss (including the latest clinical trial plans, such as gene editing therapy). The optimization algorithm uses the Adam algorithm, which optimizes the number of layers of graph neural networks (such as 3 layers) and the number of attention heads (such as 4 heads) through grid search and Bayesian optimization to improve the adaptability of the model to rare diseases with "sparse data but precise recommendations".

[0076] In the field of financial technology, by selecting cross-platform transaction data (including online payment, transfer, credit application), device fingerprint data (such as login IP, mobile phone model) and user behavior log (such as operation frequency, transaction period), a fraud identification model is constructed. In the data preprocessing stage, noise data (such as simulated transactions of test accounts) is cleaned, the missing "device location information" is filled with the average location of similar users, and the transaction amount, frequency and other features are normalized. The loss function includes accuracy loss (using cross-entropy loss to measure the accuracy of the model in identifying "fraudulent transactions"), false positive loss (penalizing the misjudgment of normal transactions as fraudulent to avoid damage to user experience) and coverage loss (encouraging the model to identify new fraud patterns such as "AI-generated fake identity information"). The optimization algorithm also uses the Adam algorithm to determine the range of graph neural network layers (2-4 layers) and attention heads (4-8 heads) through grid search, and then use Bayesian optimization to find the optimal parameters (such as 3-layer network + 6-head attention), so that the fraud identification rate of the model is improved to 98% and the false positive rate is reduced to 0.5%.

[0077] In an embodiment, a recommendation device based on an enhanced graph neural network is provided, which corresponds to the recommendation method based on the enhanced graph neural network in the above-mentioned embodiments. Referring to Figure 3 , Figure 3 The functional module diagram of a preferred embodiment of the recommendation device based on the enhanced graph neural network of the present application is shown in FIG. 1. The data acquisition and processing module 10, the graph structure construction module 20, the node importance evaluation module 30, the propagation dynamic adjustment module 40 and the hybrid model prediction module 50. The detailed description of each functional module is as follows: The data acquisition and processing module 10 is used to acquire user multi-source information and pre-process the user multi-source information; The graph structure construction module 20 is used to calculate node attributes and associated edge weights according to the pre-processed user multi-source information, and construct a graph structure through the node attributes and associated edge weights; The node importance evaluation module 30 is used to generate a node importance evaluation result according to the graph structure and the pre-processed user multi-source information; The propagation dynamic adjustment module 40 is used to dynamically adjust the information propagation path and node propagation weight according to the node importance evaluation result and feature correlation relevance; The hybrid model prediction module 50 is used to cooperatively process the target recommendation task through the attention mechanism and the graph neural network according to the information propagation path and the node propagation weight, and output a recommendation result.

[0078] In an embodiment, the data acquisition and processing module 10 comprises: obtain user basic information and user behavior sequence, and quantitatively process the user behavior sequence to generate user behavior features; obtain item basic attributes and potential features, analyze the item basic attributes and potential features, and extract item key features; obtain user social information, and count user interaction frequency in the user social information; obtain item association information, distribute the item association information, and generate same-type item attribute information and non-same-type item co-occurrence records.

[0079] In one embodiment, the graph structure construction module 20 comprises: a user node unit configured to fuse the user basic information and the user behavior features to generate a user node feature vector; an item node unit configured to fuse the item basic attributes and the item key features to generate an item node feature vector; a node attribute unit configured to construct node attributes based on the user node feature vector and the item node feature vector; a social edge weight unit configured to calculate a social edge weight according to the user interaction frequency; an item association edge weight unit configured to calculate an item association edge weight according to the same-type item attribute information and the non-same-type item co-occurrence records; an association edge weight unit configured to construct an association edge weight from the social edge weight and the item association edge weight; a graph structure unit configured to construct a graph structure from the node attributes and the association edge weight; an updated graph structure unit configured to continuously obtain user new behavior information, and update the graph structure according to the user new behavior information.

[0080] In one embodiment, the node importance evaluation module 30 comprises: a node activity unit configured to determine the activity of each node in the graph structure according to a preset measurement rule through the user multi-source information, and generate a node activity; a feature correlation unit configured to calculate feature correlation through the user node feature vector and the item node feature vector; a node graph structure position unit configured to evaluate the position of each node in the graph structure through the degree centrality, and generate a node graph structure position; a node evaluation unit configured to evaluate the importance of each node in the graph structure according to the node activity, the feature correlation, and the node graph structure position, and generate a node importance evaluation result.

[0081] In one embodiment, the propagation dynamic adjustment module 40 comprises: A propagation weight unit is configured to calculate a node propagation weight according to the node importance evaluation result, the feature correlation, and a weight coefficient; A path analysis unit is configured to identify a high-value path and a low-value path of the target recommendation task by a path analysis algorithm in combination with the node importance evaluation result, the feature correlation, and a node graph structure position; A path propagation unit is configured to perform an enhanced propagation process on the high-value path and an invalid propagation process on the low-value path.

[0082] In an embodiment, the hybrid model prediction module 50 comprises: An attention weight unit is configured to input the user node feature vector and the item node feature vector into an attention mechanism for preprocessing, and output a user attention weight and an item attention weight; A key information unit is configured to input the user attention weight and the item attention weight into a graph neural network for processing, and output new feature key information; An information feedback unit is configured to feed back the new feature key information to the attention mechanism for dynamic adjustment of a focus point; A collaborative processing model unit is configured to form a collaborative processing model from the attention mechanism and the graph neural network; A model analysis unit is configured to input the target recommendation task into the collaborative processing model, perform collaborative processing on the target recommendation task, and output a recommendation result.

[0083] In an embodiment, the collaborative processing model unit further comprises: A historical cross-domain data set is obtained and preprocessed; The preprocessed historical cross-domain data set is input into the collaborative processing model for training, and a recommendation result is output; The recommendation result is evaluated by a hybrid loss function; The collaborative processing model is optimized by an adaptive gradient descent algorithm; The parameters of the collaborative processing model are adjusted by grid search and Bayesian optimization.

[0084] In an embodiment, a computer device is provided, which can be a server, and an internal structure diagram of the computer device can be as shown in Figure 4As shown in the figure. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes non-volatile and / or volatile storage media, internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with the external user terminal through the network connection. The computer program is executed by the processor to realize the functions or steps of the recommendation method based on the enhanced graph neural network on the server side.

[0085] In one embodiment, a computer device is provided, which can be a user terminal, and its internal structure diagram can be as shown in the figure. Figure 5 As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes non-volatile storage media, internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with the external server through the network connection. The computer program is executed by the processor to realize the functions or steps of the recommendation method based on the enhanced graph neural network on the user terminal side In one embodiment, a computer device is provided, including a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to realize the following steps: Obtain user multi-source information, and preprocess the user multi-source information; According to the preprocessed user multi-source information, calculate node attributes and associated edge weights, and construct a graph structure through the node attributes and the associated edge weights; According to the graph structure and the preprocessed user multi-source information, generate a node importance evaluation result; According to the node importance evaluation result and the feature correlation dynamic adjustment information, adjust the information propagation path and the node propagation weight; Through the attention mechanism and the graph neural network, the target recommendation task is cooperatively processed according to the information propagation path and the node propagation weight, and a recommendation result is output.

[0086] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to realize the following steps: Obtain user multi-source information, and preprocess the user multi-source information; Calculate node attributes and associated edge weights according to the preprocessed user multi-source information, and construct a graph structure through the node attributes and associated edge weights; Generate node importance evaluation results according to the graph structure and the preprocessed user multi-source information; Dynamically adjust the information propagation path and node propagation weight according to the node importance evaluation results and feature correlation; Cooperate with the graph neural network to process the target recommendation task according to the information propagation path and node propagation weight through the attention mechanism, and output the recommendation result.

[0087] It should be noted that the functions or steps that the computer readable storage medium or computer device can achieve correspond to the related descriptions of the server side and the user side in the foregoing method embodiments, and to avoid repetition, they will not be described one by one here.

[0088] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0089] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is exemplified. In actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0090] It should be explained that if the software tools or components of other companies appear in the embodiments of the present application, they are only used for example introduction and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A recommendation method based on an enhanced graph neural network, characterized in that, The method comprises the following steps: obtaining user multi-source information and preprocessing the user multi-source information; calculating node attributes and associated edge weights according to the preprocessed user multi-source information, and constructing a graph structure through the node attributes and associated edge weights; generating a node importance evaluation result according to the graph structure and the preprocessed user multi-source information; dynamically adjusting an information propagation path and a node propagation weight according to the node importance evaluation result and feature correlation; cooperatively processing a target recommendation task according to the information propagation path and the node propagation weight through an attention mechanism and a graph neural network, and outputting a recommendation result. 2.The recommendation method based on the enhanced graph neural network according to claim 1, wherein, The method comprises the following steps: obtaining user multi-source information and preprocessing the user multi-source information; obtaining user basic information and user behavior sequences, quantitatively processing the user behavior sequences, and generating user behavior features; obtaining item basic attributes and potential features, analyzing the item basic attributes and potential features, and extracting item key features; obtaining user social information, and counting user interaction frequencies in the user social information; 3.The recommendation method based on the enhanced graph neural network according to claim 2, wherein, obtaining item association information, distributing the item association information, generating same-type item attribute information, and non-same-type item co-occurrence records. The method comprises the following steps: fusing the user basic information and the user behavior features, and generating a user node feature vector; fusing the item basic attributes and the item key features, and generating an item node feature vector; constructing node attributes based on the user node feature vector and the item node feature vector; calculating social edge weights according to the user interaction frequencies; calculating item association edge weights according to the same-type item attribute information and the non-same-type item co-occurrence records; constructing associated edge weights from the social edge weights and the item association edge weights; constructing a graph structure from the node attributes and the associated edge weights; 4.The recommendation method based on the enhanced graph neural network according to claim 1, wherein, continuously obtaining user new behavior information, and updating the graph structure according to the user new behavior information. The method comprises the following steps: determining the activity of each node in the graph structure according to a preset measurement rule to generate a node activity through the user multi-source information, and generating a node activity; calculating feature correlation through the user node feature vector and the item node feature vector; evaluating the position of each node in the graph structure through the degree centrality to generate a node graph structure position; 5.The recommendation method based on the enhanced graph neural network according to claim 1, wherein, evaluating the importance of each node in the graph structure according to the node activity, the feature correlation, and the node graph structure position to generate a node importance evaluation result. The method comprises the following steps: calculating node propagation weights according to the node importance evaluation result, the feature correlation, and a weight coefficient; identifying high-value paths and low-value paths of the target recommendation task through a path analysis algorithm in combination with the node importance evaluation result, the feature correlation, and the node graph structure position; performing enhanced propagation processing on the high-value paths, and performing invalid propagation processing on the low-value paths. 6.The recommendation method based on the enhanced graph neural network according to claim 3, wherein, The attention mechanism and the graph neural network are used to cooperatively process the target recommendation task according to the information propagation path and the node propagation weight, and output a recommendation result, including: The user node feature vector and the item node feature vector are input into the attention mechanism for preprocessing, and user attention weight and item attention weight are output; The user attention weight and the item attention weight are input into the graph neural network for processing, and new feature key information is output; The new feature key information is fed back to the attention mechanism for dynamic adjustment of the focus point; The attention mechanism and the graph neural network form a cooperative processing model; The target recommendation task is input into the cooperative processing model, and the target recommendation task is cooperatively processed to output a recommendation result.

7. The recommendation method based on the enhanced graph neural network according to claim 6, wherein, The attention mechanism and the graph neural network form a cooperative processing model, further including: Obtain a historical cross-domain data set, and preprocess the historical cross-domain data set; The preprocessed historical cross-domain data set is input into the cooperative processing model for training, and a recommendation result is output; The recommendation result is evaluated by a hybrid loss function; The cooperative processing model is optimized by an adaptive gradient descent algorithm; The parameters of the cooperative processing model are adjusted by grid search and Bayesian optimization.

8. A recommendation device based on an enhanced graph neural network, characterized in that, The recommendation device based on the enhanced graph neural network includes: A data acquisition and processing module for acquiring user multi-source information and preprocessing the user multi-source information; A graph structure construction module for calculating node attributes and associated edge weights based on the preprocessed user multi-source information, and constructing a graph structure based on the node attributes and associated edge weights; A node importance evaluation module for generating a node importance evaluation result based on the graph structure and the preprocessed user multi-source information; A propagation dynamic adjustment module for dynamically adjusting the information propagation path and the node propagation weight based on the node importance evaluation result and the feature correlation; A hybrid model prediction module for cooperatively processing the target recommendation task according to the information propagation path and the node propagation weight by the attention mechanism and the graph neural network, and outputting a recommendation result.

9. A computer device, comprising: The computer device includes a memory, a processor, and a recommendation program based on an enhanced graph neural network stored on the memory and executable on the processor, and the recommendation program based on the enhanced graph neural network is executed by the processor to implement the steps of the recommendation method based on the enhanced graph neural network in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a recommendation program based on an enhanced graph neural network, and the recommendation program based on the enhanced graph neural network is executed by the processor to implement the steps of the recommendation method based on the enhanced graph neural network in any one of claims 1-7.

Images