User data association mining processing method and system

Abstract

The present invention provides a method and system for user data association mining and processing, which relates to the field of data processing technology. The method preliminarily mines the association relationship between each data through an association model, and mines the implicit association relationship hidden in the data from each node and its neighboring nodes in combination with a graph neural network model; by introducing a time decay factor to dynamically adjust the edge weights of the graph neural network model, the model pays more attention to recent data and reduces the impact of outdated data; at the same time, a local retraining mechanism is preset. When the proportion of new data exceeds a preset threshold, the local retraining mechanism is triggered, thereby performing separate aggregation analysis on the new data, thereby being able to quickly respond to data changes, reduce the computational cost of global updates, and improve the real-time performance of the model. This method can mine implicit association relationships hidden in the data, and pay real-time attention to the latest behavioral data in user data to respond to changes in user data.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a user data association mining processing method and system. Background Art

[0002] User data is growing at an explosive rate across all sectors, fueled by the widespread adoption of the internet, the Internet of Things, mobile devices, and smart devices, as well as the rapid development of cloud computing and big data technologies. Whether it's user interaction records on social media platforms, shopping behavior data on e-commerce platforms, medical and health diagnosis and treatment information, financial transaction flows, or even sensor network data in smart cities, all are experiencing exponential growth. This data contains a wealth of information resources, and its potential value far exceeds the surface value of the data itself.

[0003] Deep mining and analysis of user data can lead to significant breakthroughs in many areas. First, user data can accurately reveal needs, preferences, and behavioral patterns. For example, e-commerce companies can optimize recommendation algorithms, while finance can assess credit risk and design personalized services. Second, data provides a scientific basis for decision-making, such as optimizing inventory management in retail and planning efficient delivery routes in logistics, thereby improving efficiency and reducing costs. Furthermore, data can drive product optimization and service upgrades, for example, by using logs to improve features or leveraging user feedback to identify weaknesses, thereby continuously improving quality and experience.

[0004] However, it's worth noting that the rapid growth of user data also presents numerous challenges. First, how to extract valuable information from this vast amount of data. The surge in data size has rendered traditional analytical methods inefficient, while noise and redundancy further complicate pattern recognition. Technologies like deep learning and graph computing are urgently needed to overcome algorithmic and resource bottlenecks. Second, how to cope with the dynamic evolution of data over time. Traditional graph neural network models, with fixed parameters after training, struggle to adapt to changes in graph structure and temporal characteristics. Dynamic modeling, continuous learning, or incremental update mechanisms are necessary to ensure accurate predictions.

[0005] Chinese patent application number "202210422953.1" discloses "A Social Relationship Mining Method Based on Interaction Graph Propagation," proposing a social relationship mining method based on interaction graph propagation. First, an interaction behavior network is constructed based on raw user social data. Interaction feature vectors between users are extracted and enhanced feature vectors are generated through data augmentation. Second, an adjacency matrix is generated by embedding the interaction graph structure. Finally, the enhanced feature vectors and adjacency matrix are input into a graph neural network model, and after training using cross-entropy loss optimization, classification results are output. This method integrates data augmentation and graph structure learning to fully mine complex user interaction information, significantly improving the accuracy of social relationship prediction and effectively reducing errors. This method relies on a single graph neural network model unit for feature propagation and aggregation, and only uses the adjacency matrix and enhanced feature vector input for relationship prediction. However, a single model structure has difficulty distinguishing the importance of features at different levels, resulting in limited representation of multi-level implicit connections in complex social networks, and reduced timeliness and accuracy of prediction results.

[0006] Chinese patent application number "202411051769.6" discloses a "User Relationship Mining Method and System Based on the Internet of People," proposing a relationship feature extraction network that effectively improves the accuracy of node feature extraction and relationship prediction in user relationship networks by integrating the weight distribution characteristics of the attention mechanism with the graph structure learning capabilities of GCN. First, GCN learns the graph structure and mines node features. At the same time, the attention mechanism is used to evaluate the importance of node attributes and interaction information. Combining the two, the mutual influence relationships between nodes are mined, and based on this, the GCN weights are dynamically adjusted to ultimately achieve accurate user relationship prediction. Although this method dynamically adjusts the GCN weights through the attention mechanism, the parameters are fixed after model training is completed, making it unable to adapt to the dynamic changes of user relationship networks over time. In real-world scenarios, the topological structure and interaction patterns of social networks are highly temporal, and static models cannot capture such dynamic implicit associations through real-time weight updates, resulting in the decay of prediction accuracy over time. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned existing technologies and provide a user data association mining and processing method and system to mine implicit associations hidden in social network data, quickly respond to data changes, and recommend potential friends.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] In one aspect, the present invention provides a method for user data association mining, comprising:

[0010] Using social network user data as nodes and the relationships between data as edges, we build a graph-structured association model.

[0011] Based on the graph neural network model, the neighbor node features of any node feature in the graph structure association model are aggregated to obtain the aggregated feature vector; each node is aggregated one by one, so that each node feature is propagated in the multi-layer network, thereby capturing the complex association structure;

[0012] Introducing a time decay factor to dynamically adjust the weights of edges in the graph neural network model;

[0013] Set a growth threshold. If the proportion of new user data exceeds the growth threshold, a local retraining mechanism is triggered to retrain the graph neural network model to support real-time data stream processing.

[0014] Based on the trained graph neural network model, implicit correlation information between users is mined to recommend potential friends.

[0015] Furthermore, the construction of the association model with a graph structure includes:

[0016] The multidimensional features of social network user data are converted into node attributes. Each user corresponds to a node, and the attribute vector of the node contains all the features of the user.

[0017] Define three types of edges: behavioral edges, social edges, and spatiotemporal edges;

[0018] Construct adjacency matrix and feature matrix;

[0019] The set nodes and edges are stored based on the graph database to obtain an association model with a graph structure.

[0020] Furthermore, the multidimensional features of the social network user data include behavioral features, social features, and spatiotemporal features; the spatiotemporal features represent the time and geographic location information of the user's behavior;

[0021] The adjacency matrix is used to represent the connection relationship between nodes. The size of the matrix is N×N, where N is the number of nodes. The elements of the adjacency matrix are neighbor node features, which represent the edge weights between nodes and neighbor nodes. The elements of the feature matrix are node feature vectors, which are used to store the attribute information of the nodes.

[0022] Furthermore, the graph neural network model is based on aggregating neighbor node features of any node feature in the association model of the graph structure, including:

[0023] Add the feature vectors of all neighboring nodes of any node feature to obtain the aggregated neighbor feature vector;

[0024] The aggregated neighbor feature vectors are concatenated with the feature vector of the current node to generate a new node feature vector;

[0025] A nonlinear activation function is introduced to perform nonlinear transformation on the new node feature vector to obtain the potential relationship between the current node and its neighboring nodes;

[0026] The feature vectors of all nodes are gradually updated so that each node is aggregated with its neighboring nodes one by one, capturing the complex relationships between the nodes.

[0027] Furthermore, the time decay factor is introduced to dynamically adjust the weight value of the edge in the graph neural network model, including:

[0028] Obtaining historical update trend data of social network user data, dividing the historical update trend data into multiple time period update trend data; assigning different weights to the multiple time period update trend data based on a time decay factor; the time decay factor is calculated using an exponential decay model;

[0029] After updating trend data in multiple time periods to obtain different weights, the weight values of the edges in the graph neural network model are dynamically adjusted based on the association structure between edges and nodes in each time period.

[0030] Furthermore, the weight values of the edges in the graph neural network model are dynamically adjusted based on the association structure between the edges and nodes in each time period, specifically:

[0031] Determine the edge whose weight needs to be adjusted and calculate the time decay factor of the edge;

[0032] Multiply the initial edge weight by the time decay factor to get the new edge weight;

[0033] Using new edge weights, we weight and aggregate the features of neighbor nodes to capture the latest correlation structure in social network user data.

[0034] Furthermore, the introduced time decay factor is updated in the following two ways:

[0035] Based on periodic updates of the time decay factor; based on the running time of the graph neural network model, a fixed time interval is set, and at the beginning of each time interval, the time decay factors of all edges are recalculated;

[0036] The time decay factor is updated based on event-driven updates. When new data is added, the edges that need to be updated are identified based on the correlation between the new data, and the time decay factors of the related edges are dynamically updated.

[0037] Furthermore, the growth threshold is set. If the proportion of new user data exceeds the growth threshold, a local retraining mechanism is triggered to retrain the graph neural network model, including:

[0038] Obtain historical growth traffic of social network user data and set growth thresholds based on the historical growth traffic;

[0039] Set a local retraining trigger mechanism based on growth thresholds;

[0040] Real-time social network user data is obtained and the proportion of new data is calculated within a set time window. When the proportion of new data exceeds a preset growth threshold, a local subgraph is constructed with the node affected by the new data as the center. The local subgraph only contains nodes and edges related to the new data.

[0041] Retrain the graph neural network model on the local subgraph, update the weights of nodes and edges, obtain the association relationship of the newly added data, and synchronize the updated results of the local subgraph to the global graph neural network model.

[0042] On the other hand, the present invention also provides a user data association mining processing system, comprising a shallow mining unit, a deep mining unit and a real-time mining unit;

[0043] The shallow mining unit constructs an association model with a graph structure based on the social network user data, and preliminarily mines the association relationship between each data in the social network user data based on the association model;

[0044] The deep mining unit aggregates features of each node and neighboring nodes in the graph structure based on the graph neural network model to capture complex association structures;

[0045] The real-time mining unit dynamically adjusts the edge weights of the new data based on the time decay factor, and determines whether the proportion of the new data exceeds the growth threshold. If so, it triggers a local retraining mechanism and uses the graph neural network model to perform separate aggregation analysis on the new data to obtain the correlation relationship of the new data.

[0046] Furthermore, the user data association mining processing system further includes a collection and processing module, a model integration module, a model update module and a result display module;

[0047] The collection and processing module collects social network user data in real time, cleans and normalizes the social network user data, and then transmits the social network user data to the shallow mining unit, deep mining unit and real-time mining unit in real time based on wireless technology, and performs association mining processing based on the association model and graph neural network model;

[0048] The model integration module integrates the association model and graph neural network model used by the shallow mining unit, deep mining unit and real-time mining unit, and connects the association model and graph neural network model with the front-end user data interface based on wireless technology to realize data transmission.

[0049] The model update module regularly updates the time decay factors of all edges and related edges in two ways: periodicity and event-driven;

[0050] The result display module visually displays the association structure diagram of social network user data and the association relationship between each data on the mobile terminal.

[0051] The beneficial effects of adopting the above technical solution are as follows: the user data association mining processing method and system provided by the present invention preliminarily mines the association relationship between each data in the social network user data through an association model with a graph structure, providing a basis for subsequent deep mining; combining with the graph neural network model to perform feature aggregation on each node and neighboring nodes in the graph structure, thereby being able to mine the implicit association relationship hidden in the data; thereby achieving the effect of gradual mining and processing from shallow to deep layers, improving the depth and accuracy of mining. By introducing a time decay factor to dynamically adjust the weight value of the edge in the graph neural network model, a higher weight is given to the user's recent behavior, so that the graph neural network model can pay more attention to recent data, reduce the impact of outdated data, enable it to capture the latest association relationship, and ensure the model's adaptability to dynamic changes; at the same time, a local retraining mechanism is preset. When the proportion of new data exceeds a preset threshold, the local retraining mechanism is triggered, thereby separately performing aggregation analysis on the new data, thereby being able to quickly respond to data changes, reduce the computational cost of global updates, improve the real-time performance of the model, and then recommend potential friends. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of a user data association mining method provided by an embodiment of the present invention;

[0053] Figure 2 This is a structural block diagram of a user data association mining and processing system provided by the first embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0055] In order to solve the problems in the existing technology, such as the data association mining is not comprehensive and the potential relationship between user data cannot be fully mined; the database needs to be scanned multiple times, which makes the candidate processing amount large, resulting in low data processing efficiency and difficulty in meeting the real-time processing requirements of large-scale data; at the same time, the user data update rate is fast, and the traditional method is difficult to adapt to the technical problems of fast processing of real-time updated user data, combined with Figure 1-Figure 2 , this embodiment provides the following technical solutions:

[0056] In this embodiment, a method for user data association mining is provided. First, a correlation model is used to preliminarily associate the relevant relationships in the social network user data to form a graph-structured association relationship of the social network user data.

[0057] Furthermore, the graph neural network model is used to perform feature aggregation analysis on each data node in the graph structure and its associated neighboring data nodes, further mining the deep relationships in the user data;

[0058] A time decay factor is introduced into the graph neural network model to dynamically adjust the weight values of the edges in the graph neural network model, so that the graph neural network model can pay attention to the latest behavioral data in user data in real time to respond to changes in user data.

[0059] The specific steps include:

[0060] Step S1: Using social network user data as nodes and the relationships between data as edges, a graph-structured association model is constructed. Specifically:

[0061] The multidimensional features of social network user data are converted into node attributes. These multidimensional features include: behavioral features, representing users' purchasing behavior, browsing history, and click behavior; social features, representing social relationships and interactions between users; and spatiotemporal features, representing the time and location of user behaviors. Each user corresponds to a node, and the node's attribute vector contains all the user's features. Three types of edges are defined: behavioral edges, social edges, and spatiotemporal edges. Behavioral edges are based on user activities and reflect the connections between various actions performed by users in specific scenarios. Spatiotemporal edges are based on both time and space and represent the connections between users or entities at specific time and spatial locations. Social edges represent direct social relationships between users. For example, if two users are friends or follow each other on a social network, a social edge is added between them. The weight of the social edge can be calculated based on the frequency of interaction or the strength of the relationship.

[0062] Construct an adjacency matrix and a feature matrix. The adjacency matrix is used to represent the connection relationship between nodes. The size of the matrix is N×N, where N is the number of nodes. The elements of the adjacency matrix are neighbor node features, which represent the edge weights between nodes and neighbor nodes. If there is an edge between a node and a neighbor node, the elements of the adjacency matrix are greater than 0; if there is no edge, the elements of the adjacency matrix are equal to 0. The elements of the feature matrix are node feature vectors, which are used to store the attribute information of the node. The size of the matrix is N×D, where D is the dimension of the node attribute vector. Each row of the feature matrix represents the attribute vector of the node.

[0063] The set nodes and edges are stored based on the graph database to obtain an association model with a graph structure; the association model is updated in real time based on changes in social network user data; specifically, by converting the multidimensional features of social network user data into node attributes and defining three types of edges, the association relationships in social network user data can be fully captured, providing important application value for subsequent deep mining.

[0064] Step S2: Based on the graph neural network model, the neighbor node features of any node feature in the association model of the graph structure are aggregated to obtain the aggregated feature vector; each node is aggregated one by one, so that each node feature is propagated in the multi-layer network, thereby capturing the complex association structure; specifically:

[0065] Step S21: Based on the summation aggregation function in the graph neural network, the feature vectors of all neighboring nodes of any node feature are added to obtain an aggregated neighbor feature vector; the aggregated feature vector can be used to update the feature vector of the node, so that the node can learn the information of its neighbor nodes.

[0066] Step S22: concatenate the aggregated neighbor feature vectors with the feature vector of the current node to generate a new node feature vector;

[0067] Step S23: Introduce a nonlinear activation function to perform nonlinear transformation on the new node feature vector to obtain the potential relationship between the current node and its neighboring nodes;

[0068] Step S24: Repeat the above steps S21 to S23, and gradually update the feature vectors of all nodes, so that each node is aggregated with its neighbor nodes one by one, and the complex relationship between the nodes is captured; wherein, in each layer of the graph neural network model, the feature vector of the node will be updated according to the feature information of its neighbor nodes; through multi-layer propagation, the graph neural network model can learn more distant neighbor information, thereby capturing the complex association structure in the graph structure.

[0069] The present invention uses an association model to preliminarily mine the association relationships between various data in social network user data, providing a basis for subsequent in-depth mining; and uses a graph neural network model to mine the implicit association relationships hidden in the data from the feature vectors of each node and its neighboring nodes, thereby achieving the effect of gradual mining and processing from shallow to deep layers, and improving the depth and accuracy of mining.

[0070] Step S3: Introduce the time decay factor to dynamically adjust the weight value of the edge in the graph neural network model; specifically:

[0071] Obtain historical update trend data of social network user data, and divide the historical update trend data into multi-period update trend data; assign different weights to the multi-period update trend data based on the time decay factor; for example: a smaller decay coefficient, such as 0.1, has a faster time decay, and the graph neural network model pays more attention to recent data; a larger decay coefficient, such as 0.9, has a slower time decay, and the graph neural network model has a milder weight decay on historical data; specifically, as time passes, the importance of early data gradually decreases, and the time decay factor can reduce the impact of early data on the graph neural network model, preventing the graph neural network model from being misled by outdated information; through the time decay factor, the graph neural network model can pay more attention to recent data, so that the graph neural network model can reflect the time characteristics of the data in real time; at the same time, it provides more accurate weight information for the local retraining mechanism, ensuring that local updates can better reflect the latest changes in the data, thereby quickly capturing the latest changes in social network user data.

[0072] In this embodiment, in a social network, nodes represent users, and edges represent relationships between users. Through the aggregation mechanism of graph neural networks, the graph neural network model can learn complex relationships between users. For example, regarding community detection, through multi-layer propagation, the model can identify groups of users with similar characteristics. Regarding link prediction, by aggregating neighbor features, the model can predict whether there are potential connections between users.

[0073] After updating trend data in multiple time periods and obtaining different weights, the weight values of edges in the graph neural network model are dynamically adjusted based on the association structure between edges and nodes in each time period; specifically:

[0074] Determine the edge whose weight needs to be adjusted, and calculate the time decay factor of the edge according to the calculation formula of the time decay factor;

[0075] In this embodiment, an exponential decay model is used as the calculation formula for the time decay factor to assign a higher weight to the most recent user behavior. The calculation formula is as follows:

[0076] ;

[0077] in, Expressed as a time decay factor; It is expressed as the time difference between the current time and the time when the user behavior occurs; the value range of the time decay factor is: 0 < time decay factor < 1, the smaller the time decay factor, the faster the time decay; the initial edge weight is multiplied by the time decay factor to obtain a new edge weight; using the new edge weight, the characteristics of the weighted aggregated neighbor nodes are weighted to capture the latest association structure in the social network user data; specifically, in the multi-period updated trend data, by determining the edges that need to adjust the weight, calculating their time decay factors, and then updating the edge weights based on the calculation results, as well as the characteristics of the weighted aggregated neighbor nodes, the graph neural network model can dynamically adjust the edge weights, thereby capturing the latest association structure in the social network user data; this not only improves the dynamic adaptability of the graph neural network model, but also enhances the accuracy and timeliness of the graph neural network model, and has broad application prospects.

[0078] After the time decay factor is introduced, the time decay factor is updated in the following two ways:

[0079] Update the time decay factor based on periodicity; set fixed time intervals based on the running time of the graph neural network model, such as daily or weekly, and recalculate the time decay factors of all edges at the beginning of each time interval.

[0080] The time decay factor is updated based on event-driven updates. When new data is added, the edges that need to be updated are identified based on the correlation between the new data, and the time decay factors of the related edges are dynamically updated. For example, if a new user behavior is added, the edges related to the behavior need to be updated, so as to quickly respond to new user behaviors and dynamically adjust the recommendation weights.

[0081] Step S4: Set a growth threshold. If the proportion of new user data exceeds the growth threshold, a local retraining mechanism is triggered to retrain the graph neural network model to support real-time data stream processing. Specifically:

[0082] Obtain the historical growth traffic of social network user data and set a growth threshold based on the historical growth traffic; set a local retraining trigger mechanism based on the growth threshold; obtain real-time social network user data and calculate the proportion of new data within a set time window, such as every hour, every day, etc.; when the proportion of new data exceeds the preset growth threshold, construct a local subgraph centered on the node affected by the new data, and the local subgraph only contains nodes and edges related to the new data; retrain the graph neural network model on the local subgraph, update the weights of the nodes and edges, obtain the association relationship of the new data, and synchronize the updated results of the local subgraph to the global graph neural network model to ensure that the graph neural network model can promptly reflect the changes in the new data; specifically, the local retraining mechanism only updates the parts that are most affected by the new data, rather than fully retraining the entire model, thereby reducing the computational burden caused by global updates and significantly reducing the computational cost.

[0083] This invention introduces a time decay factor to give higher weight to users' recent behaviors, so that the model can pay more attention to recent data and reduce the impact of outdated data. At the same time, a local retraining mechanism is preset so that the graph neural network model can perform separate aggregation analysis on new data when the new data reaches a certain proportion, thereby quickly responding to data changes, reducing the computational cost of global updates, and improving the real-time performance of the model.

[0084] Step S5: Based on the trained graph neural network model, the implicit association information between users is mined and potential friends are recommended.

[0085] This embodiment uses a trained graph neural network model to generate embedding vectors for user nodes, and then recommends potential friends by calculating the similarity between vectors (such as cosine similarity).

[0086] In this embodiment, a user data association mining processing system includes a shallow mining unit, a deep mining unit and a real-time mining unit;

[0087] The shallow mining unit constructs an association model with a graph structure based on social network user data, and preliminarily mines the association relationships between various data in the social network user data based on the association model; specifically, by defining users or user behaviors in the social network as nodes in the graph, for example: each user is a node; defining edges based on the relationship between users or the association between behaviors, for example: if two users are friends, then a social edge is added between them; if two users purchase the same product, then a behavioral edge is added between them; based on the definition of nodes and edges, a graph-structured association model is constructed; based on predefined rules, the association relationships in the social network user data are mined; for example: if two users visit the same web page in the same time period, it is considered that there is some kind of association between them. The preliminary results obtained provide important support for subsequent deep mining and real-time mining.

[0088] The deep mining unit aggregates features of each node and neighboring nodes in the graph structure based on the graph neural network model to capture complex association structures. Specifically, by constructing a graph neural network model and using the feature aggregation mechanism to capture complex association relationships in the graph structure, the system is provided with deeper feature representation and association mining capabilities. Combined with multi-layer propagation and nonlinear transformation, the model can more effectively process complex data structures and improve the overall performance and adaptability of the system.

[0089] The real-time mining unit dynamically adjusts the edge weights of the new data based on the time decay factor algorithm, and determines whether the proportion of the new data exceeds the growth threshold. If so, the local retraining mechanism is triggered, and the graph neural network model is used to perform aggregation analysis on the new data separately to obtain the correlation relationship of the new data; specifically, assuming that the proportion of new data is calculated within a certain time window, the calculation formula is: proportion of new data = amount of new data / total amount of data; if the proportion of new data exceeds the preset growth threshold, the local retraining mechanism is triggered; when the proportion of new data exceeds the growth threshold, the local retraining mechanism is triggered, and the new data can be processed separately. The specific steps are as follows: construct a local subgraph with the node affected by the new data as the center; retrain the graph neural network model on the local subgraph, update the weights of the nodes and edges, so as to capture the correlation relationship of the new data; synchronize the updated results of the local subgraph to the global model to ensure that the global model can reflect the changes in the new data in a timely manner; thereby reducing the computational cost of the global update and improving the real-time performance and processing efficiency of the model.

[0090] In this embodiment, a user data association mining processing system further includes a collection and processing module, a model integration module, a model update module and a result display module;

[0091] The collection and processing module collects social network user data in real time, cleans and normalizes the user data, and then transmits the social network user data to the shallow mining unit, deep mining unit and real-time mining unit in real time based on wireless technology, and performs association mining processing based on the association model and graph neural network model. Specifically, by collecting social network user data, including user basic information, behavior records, social relationships, etc., and cleaning, denoising and formatting the user data, invalid data, abnormal data and duplicate data are removed, and then the data is converted into a unified format to ensure data quality and facilitate subsequent processing.

[0092] The model integration module integrates the association model and graph neural network model used by the shallow mining unit, deep mining unit and real-time mining unit, and connects the association model and graph neural network model with the front-end user data interface based on wireless technology to realize data transmission.

[0093] The model update module uses both periodic and event-driven methods to regularly update the time decay factors of all edges and related edges; specifically, by regularly recalculating the time decay factors of all edges according to preset time intervals, it can ensure that the model adjusts weights regularly to reduce the impact of outdated data; based on event-driven dynamic updates of the time decay factors of related edges, it can quickly respond to data changes and ensure the timeliness and accuracy of the graph neural network model.

[0094] The result display module visually displays the association structure diagram of social network user data and the association relationship between each data on the mobile terminal.

[0095] Working principle: Through the association model with a graph structure, the association relationship between each data in the social network user data is preliminarily mined, and the features of each node and neighboring nodes in the graph structure are aggregated in combination with the graph neural network model to mine the implicit association relationship hidden in the data; by introducing the time decay factor to dynamically adjust the weight value of the edge in the graph neural network model, the graph neural network model can pay more attention to recent data, reduce the impact of outdated data, enable it to capture the latest association relationship, and ensure the adaptability of the graph neural network model to dynamic changes; at the same time, a local retraining mechanism is preset. When the proportion of new data exceeds the preset threshold, the local retraining mechanism is triggered, thereby performing separate aggregation analysis on the new data to quickly respond to data changes.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A user data association mining method, characterized in that: include: Using social network user data as nodes and the relationships between data as edges, we build a graph-structured association model. The multidimensional features of social network user data are converted into node attributes. Each user corresponds to a node, and the attribute vector of the node contains all the features of the user. Define three types of edges: behavioral edges, social edges, and spatiotemporal edges; Construct adjacency matrix and feature matrix; The set nodes and edges are stored based on the graph database to obtain an association model with a graph structure; Based on the graph neural network model, the neighbor node features of any node feature in the graph structure association model are aggregated to obtain the aggregated feature vector; each node is aggregated one by one, so that each node feature is propagated in the multi-layer network, thereby capturing the complex association structure; Introducing a time decay factor to dynamically adjust the weights of edges in the graph neural network model; Obtain historical update trend data of social network user data, and divide the historical update trend data into update trend data for multiple time periods; Assign different weights to multi-period updated trend data based on time decay factors; The time decay factor is calculated using an exponential decay model; After updating trend data in multiple time periods to obtain different weights, the weight values of edges in the graph neural network model are dynamically adjusted based on the association structure between edges and nodes in each time period; Set a growth threshold. If the proportion of new user data exceeds the growth threshold, a local retraining mechanism is triggered to retrain the graph neural network model to support real-time data stream processing. Obtain historical growth traffic of social network user data and set growth thresholds based on the historical growth traffic; Based on the growth threshold, a local retraining trigger mechanism is set; real-time social network user data is obtained, and the proportion of new data is calculated within the set time window; when the proportion of new data exceeds the preset growth threshold, a local subgraph is constructed with the node affected by the new data as the center. The local subgraph only contains nodes and edges related to the new data; the graph neural network model is retrained on the local subgraph, the weights of the nodes and edges are updated, the association relationship of the new data is obtained, and the updated results of the local subgraph are synchronized with the global graph neural network model to ensure that the graph neural network model can promptly reflect the changes of the new data; Based on the trained graph neural network model, implicit correlation information between users is mined to recommend potential friends.

2. A user data association mining method according to claim 1, characterized in that: The multidimensional features of the social network user data include behavioral features, social features, and spatiotemporal features; the spatiotemporal features represent the time and geographic location information of the user's behavior; The adjacency matrix is used to represent the connection relationship between nodes. The size of the matrix is N×N, where N is the number of nodes. The elements of the adjacency matrix are neighbor node features, which represent the edge weights between nodes and neighbor nodes. The elements of the feature matrix are node feature vectors, which are used to store the attribute information of the nodes.

3. The method for user data association mining according to claim 2, characterized in that: The graph neural network model is based on aggregating neighbor node features of any node feature in the association model of the graph structure, including: Add the feature vectors of all neighboring nodes of any node feature to obtain the aggregated neighbor feature vector; The aggregated neighbor feature vectors are concatenated with the feature vector of the current node to generate a new node feature vector; A nonlinear activation function is introduced to perform nonlinear transformation on the new node feature vector to obtain the potential relationship between the current node and its neighboring nodes; The feature vectors of all nodes are gradually updated so that each node is aggregated with its neighboring nodes one by one, capturing the complex relationships between the nodes.

4. The method for user data association mining according to claim 3, characterized in that: The weight values of the edges in the graph neural network model are dynamically adjusted based on the association structure between the edges and nodes in each time period, specifically: Determine the edge whose weight needs to be adjusted and calculate the time decay factor of the edge; Multiply the initial edge weight by the time decay factor to get the new edge weight; Using new edge weights, we weight and aggregate the features of neighbor nodes to capture the latest correlation structure in social network user data.

5. A user data association mining method according to claim 4, characterized in that: The introduced time decay factor is updated in the following two ways: Based on periodic update time decay factor; Based on the running time of the graph neural network model, a fixed time interval is set, and at the beginning of each time interval, the time decay factors of all edges are recalculated; The time decay factor is updated based on event-driven updates. When new data is added, the edges that need to be updated are identified based on the correlation between the new data, and the time decay factors of the related edges are dynamically updated.

6. The method for user data association mining according to claim 1, characterized in that: The growth threshold is set. If the proportion of new user data exceeds the growth threshold, a local retraining mechanism is triggered to retrain the graph neural network model, including: Obtain historical growth traffic of social network user data and set growth thresholds based on the historical growth traffic; Set a local retraining trigger mechanism based on growth thresholds; Real-time social network user data is obtained and the proportion of new data is calculated within a set time window. When the proportion of new data exceeds a preset growth threshold, a local subgraph is constructed with the node affected by the new data as the center. The local subgraph only contains nodes and edges related to the new data. Retrain the graph neural network model on the local subgraph, update the weights of nodes and edges, obtain the association relationship of the newly added data, and synchronize the updated results of the local subgraph to the global graph neural network model.

7. A user data association mining processing system, implemented based on the user data association mining processing method according to claim 1, characterized in that: Including shallow excavation unit, deep excavation unit and real-time excavation unit; The shallow mining unit constructs an association model with a graph structure based on the social network user data, and preliminarily mines the association relationship between each data in the social network user data based on the association model; The deep mining unit aggregates features of each node and neighboring nodes in the graph structure based on the graph neural network model to capture complex association structures; The real-time mining unit dynamically adjusts the edge weights of the new data based on the time decay factor, and determines whether the proportion of the new data exceeds the growth threshold. If so, it triggers a local retraining mechanism and uses the graph neural network model to perform separate aggregation analysis on the new data to obtain the correlation relationship of the new data.

8. A user data association mining and processing system according to claim 7, characterized in that: It also includes a collection and processing module, a model integration module, a model update module and a result display module; The collection and processing module collects social network user data in real time, cleans and normalizes the social network user data, and then transmits the social network user data to the shallow mining unit, deep mining unit and real-time mining unit in real time based on wireless technology, and performs association mining processing based on the association model and graph neural network model; The model integration module integrates the association model and graph neural network model used by the shallow mining unit, deep mining unit and real-time mining unit, and connects the association model and graph neural network model with the front-end user data interface based on wireless technology to realize data transmission; The model update module regularly updates the time decay factors of all edges and related edges in two ways: periodicity and event-driven; The result display module visually displays the association structure diagram of social network user data and the association relationship between each data on the mobile terminal.

Images