A big data-based user management method and system

Abstract

The application discloses a kind of based on big data's user management method and system, comprising the following steps: S1, user data is collected and preprocessed;S2, construct heterogeneous graph structure, edge indicates the interaction of user and behavior;S3, input graph attention neural network to heterogeneous graph structure, construct weight matrix according to node type and weighted aggregation, generate user feature graph;S4, introduce structure and semantic attention factor, adjust user adjacent edge weight, update user feature graph;S5, input cascade convolution unit to user feature graph, execute local, global and cross-domain sampling, obtain multi-scale behavior representation;S6, decouple interest and social feature to behavior representation, generate independent subspace and construct feature mapping set;S7, aggregate feature mapping, calculate user state vector, output clustering label and management instruction set.The application realizes user behavior fine modeling and intelligent management, improves clustering accuracy and strategy response efficiency.

Description

Technical Field

[0001] This invention relates to the field of user management technology, and in particular to a user management method and system based on big data. Background Technology

[0002] Against the backdrop of rapid digital transformation and the development of intelligent services, user management technology has gradually become a core component in many fields such as big data platforms, social networks, e-commerce operations, and intelligent recommendation systems. Existing user management methods generally rely on static user information, simple behavior records, or manually set rule bases, which have limited capabilities for mining and processing user data, making it difficult to achieve dynamic and accurate segmentation management and personalized operational strategy output.

[0003] Currently, some systems employ traditional data mining algorithms, such as cluster analysis, association rules, or decision trees, to label user data. However, these methods are mostly based on flat feature representations and lack the ability to deeply model multidimensional user behavior, social relationships, and dynamic changes in preferences. In particular, they struggle to effectively utilize the complex structural information and graph-like relationships between users and their behaviors.

[0004] With the development of graph neural networks and graph attention mechanisms, some studies have attempted to shift user behavior modeling towards graph structure representation and introduce neural networks for feature learning. However, these methods often overlook the heterogeneity of heterogeneous node types in graph structures and fail to set differentiated weighting mechanisms for different types of nodes and edge relationships. This leads to generalization problems in the node feature aggregation process, which can easily result in fuzzy user profile representation and insufficient clustering accuracy.

[0005] Furthermore, existing graph models typically employ single-scale sampling strategies, such as fixed-adjacent-node convolution or path-length-limited sampling, which struggle to simultaneously capture local user behavior features, global topological structure features, and cross-domain behavioral jump features, resulting in the loss of user behavior feature information. Simultaneously, in the user behavior modeling process, interest features and social features often coexist and are coupled; without decoupling, representational interference can easily occur, affecting the accuracy and practicality of the final management strategy.

[0006] Therefore, how to provide a user management method and system based on big data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a user management method and system based on big data. This invention fully utilizes key technologies such as graph attention neural networks, multi-scale graph structure convolution, and feature decoupling and fusion. It describes in detail the entire process of constructing user profiles from heterogeneous user behavior data, extracting multi-dimensional behavioral features, and generating accurate grouping and management instructions. It has the advantages of strong structural expression capabilities, detailed user state characterization, and high efficiency of intelligent decision-making.

[0008] A user management method based on big data according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect user data and preprocess it;

[0010] S2. Based on the preprocessed user data, define a node set and an edge set to form a heterogeneous graph structure. The node set includes user nodes and behavior nodes, and the edge set represents the interaction relationship between users and behaviors.

[0011] S3. Input the heterogeneous graph structure into the graph attention neural network, construct attention weight matrices according to node types, perform weighted aggregation on the adjacency relationships between different types of nodes, and output the user feature map;

[0012] S4. Introduce structural attention factors and semantic attention factors to perform dual attention adjustment on the adjacent edges of user nodes, construct an attention propagation mechanism that integrates structural and semantic features, and update the user feature graph.

[0013] S5. Input the updated user feature map into the graph structure cascaded convolutional unit and perform local adjacency sampling, global graph topology sampling and cross-domain skip sampling respectively to obtain multi-scale user behavior representation.

[0014] S6. Decouple the multi-scale user behavior representation from interest features and social features to form an independent representation subspace, and construct a user feature mapping set;

[0015] S7. Perform feature aggregation based on the user feature mapping set, calculate the user state vector through the feature selection weight vector, and output the user group label and management instruction set based on the state vector.

[0016] Optionally, the user data includes identity information, behavior logs, transaction records, preference tags, device fingerprints, and access trajectories.

[0017] Optionally, the preprocessing includes missing value imputation, filtering and denoising, format conversion, outlier removal, and timestamp alignment.

[0018] Optionally, S2 specifically includes:

[0019] S21. Divide the preprocessed user data according to attribute categories, and extract identity information data, behavior log data, transaction record data, preference tag data, device fingerprint data and access trajectory data respectively. Perform feature vectorization processing on each type of data to construct a standardized feature data set.

[0020] S22. Construct a node set based on the standardized feature data set. The node set includes multiple user nodes and multiple behavior nodes, wherein each user node corresponds to a unique user and each behavior node corresponds to a user behavior type.

[0021] S23. Based on the association records between users and behaviors in the original data, construct an edge set. Each edge in the edge set represents the interaction relationship between a user node and a behavior node, and set the weight of the edge based on the interaction frequency and behavior intensity characteristics.

[0022] S24. Construct an interaction matrix based on the connection relationship between user nodes, behavior nodes and edge sets. Each element in the interaction matrix represents a specific interaction value between a user node and a behavior node.

[0023] S25. Integrate the node set, edge set, and interaction matrix, and establish a heterogeneous graph structure according to the principles of node type classification, edge weight assignment, and topology construction to complete the graph model initialization.

[0024] Optionally, S3 specifically includes:

[0025] S31. Input the heterogeneous graph structure into the graph attention neural network, and divide the user type node set and behavior type node set according to the node type, and establish the node type mapping relationship;

[0026] S32. Select the corresponding attention transformation matrix according to the type of the node, and perform a linear transformation on the initial feature vector of each node to generate a typed node representation vector;

[0027] S33. For any user type node, traverse its adjacent behavior type nodes, construct an attention weight matrix using the similarity relationship between the user type node representation vector and the adjacent behavior type node representation vector, and perform normalization processing on each attention weight value.

[0028] S34. All adjacent behavior type nodes of the user type node are weighted and aggregated according to the normalized attention weight value. The representation vectors of all adjacent nodes are multiplied and summed with their corresponding attention weights to generate the updated feature representation of the user type node.

[0029] S35. Perform weighted aggregation and update operations on all user type nodes and output the user feature map.

[0030] Optionally, S4 specifically includes:

[0031] S41. In the set of user type nodes, construct structural attention factors and semantic attention factors respectively to measure the degree of association between user nodes and their adjacent behavioral nodes in terms of structural connection strength and semantic label similarity.

[0032] S42. Perform dual factor adjustment on the attention weight matrix. On the basis of the original attention weight, introduce structural attention factor and semantic attention factor as multiplicative adjustment factors, and normalize the weights of all adjacent nodes of each user node to form an attention weight matrix that integrates structural and semantic features.

[0033] S43. Use the updated attention weight matrix to perform a weighted aggregation operation on the feature vectors of the adjacent behavioral nodes of the user node, complete the feature update of the user node, generate a user feature representation vector that integrates structural features and semantic features, and update the user feature map.

[0034] Optionally, the structural attention factor is quantified by analyzing the topological path density, connection direction consistency, and feature vector dot product relationship between user nodes and adjacent nodes in the graph, while the semantic attention factor is calculated based on the embedding vector of each node in the semantic label space and combined with cosine similarity to calculate the semantic matching degree between nodes.

[0035] Optionally, S5 specifically includes:

[0036] S51. Input the user feature map into the graph structure cascaded convolutional unit and perform local adjacency sampling operation. In the operation, set the adjacency sampling depth and edge weight threshold, filter the local neighbor node set based on the direct connection relationship of user nodes, and process the selected neighbor node feature vectors through average aggregation and convolution transformation to obtain the local adjacency feature representation.

[0037] S52. Perform a global graph topology sampling operation. In this operation, a corresponding graph structure position encoding vector is generated for each user node. The feature similarity coefficient between nodes is calculated by combining the feature vectors of all user nodes. The feature similarity coefficient is used to weight and summarize the features of all other nodes. The node's own features and position encoding are then fused and input into a fully connected neural network layer to obtain a global context-aware feature representation.

[0038] S53. Perform cross-domain skip sampling operation. In the operation, set the skip interval factor, extract the difference features between the local convolution output and the global convolution output, perform affine transformation on the difference features through the skip mapping matrix, and perform weighted fusion processing on the local features, global features and transformed features, and finally output the behavior representation result of each user node under multi-scale information fusion, that is, multi-scale user behavior representation.

[0039] Optionally, S6 specifically includes:

[0040] S61. Input the multi-scale user behavior representation into the feature decoupling unit, set the interest decoupling transformation matrix and the social decoupling transformation matrix respectively, and perform two sets of linear transformation operations on the multi-scale behavior feature vector of each user node. The first set of transformations is used to extract the user's interest feature representation, and the second set of transformations is used to extract the user's social relationship feature representation. Output the interest feature vector and the social feature vector respectively to form two sets of feature representation vectors with different dimensions.

[0041] S62. For interest feature vectors and social feature vectors, an orthogonal constraint mechanism is constructed. By calculating the sum of the dot products of the two vector subspaces on the feature dimension and the norm normalization value, the correlation between the two is measured. The optimization objective is to minimize the squared similarity value, which explicitly suppresses the coupling term between interest features and social features, realizes information isolation between different subspaces, and ensures the mutual independence of interest feature subspace and social feature subspace in the feature representation process.

[0042] S63. Perform subspace combination on the interest feature vector and social feature vector after orthogonal constraint optimization, apply weight factors to control the influence ratio of the two in the fusion process, and perform product operation according to the corresponding position of the elements to perform nonlinear combination of interest features and social features to obtain the fused user feature mapping vector, and summarize the feature mapping vectors corresponding to all user nodes to form a user feature mapping set.

[0043] According to an embodiment of the present invention, a user management system based on big data includes:

[0044] The data processing module is used to collect user data and perform preprocessing operations on the user data;

[0045] The heterogeneous graph construction module is used to define the node set and edge set, and to build a heterogeneous graph structure containing user nodes and behavior nodes. The edge set is used to represent the interaction relationship between users and behaviors.

[0046] The feature aggregation module is used to input heterogeneous graph structures into the graph attention neural network, construct attention weight matrices according to node types, perform weighted aggregation on the adjacency relationships between different types of nodes, and output user feature maps.

[0047] The attention fusion module is used to introduce structural attention factors and semantic attention factors, perform dual attention adjustment on the adjacent edges of user nodes, construct an attention propagation mechanism that integrates structural and semantic features, and update the user feature graph.

[0048] The convolutional sampling module is used to input the updated user feature map into the graph structure cascaded convolutional unit, and perform local adjacency sampling, global graph topology sampling and cross-domain skip sampling respectively to obtain multi-scale user behavior representations;

[0049] The feature decoupling module is used to decouple the multi-scale user behavior representation from interest features and social features, forming an independent representation subspace and constructing a user feature mapping set.

[0050] The status determination module is used to perform feature aggregation based on the user feature mapping set, calculate the user status vector through the feature selection weight vector, and output the user group label and management instruction set based on the status vector.

[0051] The beneficial effects of this invention are:

[0052] First, this invention utilizes a dual regulation mechanism of structural attention factor and semantic attention factor to enhance the sensitivity of graph neural network in the process of multi-type node aggregation. This allows user features to retain graph topology information while incorporating contextual features of behavioral semantics during aggregation, effectively solving the problem of ambiguous expression in existing user profiles.

[0053] Secondly, this invention constructs a graph-structured cascaded convolutional unit and designs three sampling mechanisms: local adjacency sampling, global topological sampling, and cross-domain skip sampling. This allows user behavior features to be fully captured at different scales and connection levels, thereby forming a hierarchical and structurally complete expression of behavior patterns during the graph structure representation process, which significantly improves the accuracy of user grouping and the system's generalization ability.

[0054] Finally, this invention establishes an orthogonal constraint mechanism by decoupling interest features and social features from multi-scale user behavior representations, and introduces a fusion strategy to generate user feature mapping sets, ensuring the independence and integrity of user information expression during state modeling. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a flowchart of a user management method based on big data proposed in this invention;

[0057] Figure 2 This is a schematic diagram of graph neural network modeling and multi-scale sampling for a user management method based on big data proposed in this invention;

[0058] Figure 3 This is a module structure diagram of a user management system based on big data proposed in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figure 1-2 A user management method based on big data includes the following steps:

[0061] S1. Collect user data and preprocess it;

[0062] S2. Based on the preprocessed user data, define a node set and an edge set to form a heterogeneous graph structure. The node set includes user nodes and behavior nodes, and the edge set represents the interaction relationship between users and behaviors.

[0063] S3. Input the heterogeneous graph structure into the graph attention neural network, construct attention weight matrices according to node types, perform weighted aggregation on the adjacency relationships between different types of nodes, and output the user feature map;

[0064] S4. Introduce structural attention factors and semantic attention factors to perform dual attention adjustment on the adjacent edges of user nodes, construct an attention propagation mechanism that integrates structural and semantic features, and update the user feature graph.

[0065] S5. Input the updated user feature map into the graph structure cascaded convolutional unit and perform local adjacency sampling, global graph topology sampling and cross-domain skip sampling respectively to obtain multi-scale user behavior representation.

[0066] S6. Decouple the multi-scale user behavior representation from interest features and social features to form an independent representation subspace, and construct a user feature mapping set;

[0067] S7. Perform feature aggregation based on the user feature mapping set, calculate the user state vector through the feature selection weight vector, and output the user group label and management instruction set based on the state vector.

[0068] In this embodiment, the user data originates from multi-channel terminal access, specifically including identity information extracted from user registration information, behavior logs collected based on page click sequences and dwell time, transaction records synchronized through the payment system, preference tags formed based on user content selection and interaction behavior, device fingerprints generated by the client device's unique identifier, and access trajectory data obtained through IP address, GPS signal, and network request path parsing.

[0069] In this embodiment, during the data preprocessing process, missing fields are first filled by combining K-nearest neighbor interpolation and logistic regression prediction. Then, the abnormal spike signals in the behavior log are filtered and denoised using the wavelet thresholding method. Subsequently, the structured and unstructured data are uniformly converted from JSON to vector format, and the fields are mapped according to the preset data dictionary rules. Finally, the time sequence alignment of cross-terminal data is achieved through the timestamp normalization function.

[0070] In this embodiment, S2 specifically includes:

[0071] S21. Divide the preprocessed user data into a set of identity information, a set of behavior logs, a set of transaction records, a set of preference tags, a set of device fingerprints, and a set of access trajectories. Perform feature vectorization processing on each set to construct a unified feature space.

[0072] S22. Based on a unified feature space, construct a node set V = V u ∪V b The user node set V u ={u1, u2, ..., u m}, the set of behavior nodes V b ={b1, b2, ..., b n Define the edge set E = {e} ij |u i ∈V u b j ∈V b} represents the relationship between users and behaviors, where m represents the number of user nodes and n represents the number of behavior types;

[0073] S23. Construct the node feature matrix X and the interaction matrix A, where x k ∈X represents the initial feature vector of the corresponding node, a ij ∈ A Represents user node u i With behavior node b j Frequency of interaction between them;

[0074] S24. Based on the node feature matrix X and the interaction matrix A, calculate the association weight matrix W between nodes:

[0075]

[0076] Among them, w ij Represents node v i With node v j The weight values ​​between them, x ik Represents node v i The eigenvalues ​​in the k-th dimension, x jk Represents node v j In the k-th eigenvalue, d represents the feature dimension, λ represents the interaction modulator, ∈ represents a non-zero constant, and log2(·) represents the logarithmic function;

[0077] S25. Based on the node type and edge connection relationship, the node set V, the edge set E, and the interaction matrix A are jointly represented to construct a heterogeneous graph structure G = (V, E, A).

[0078] In this embodiment, S3 specifically includes:

[0079] S31. Input the heterogeneous graph structure G into the graph attention neural network, and construct type grouping mapping sets based on user type nodes and behavior type nodes respectively. and Where t i , t j These represent different node type labels;

[0080] S32. Initialize the attention transformation matrix W according to the node type. (u) With W (b) For any node v k eigenvectors x ∈V k Perform a linear mapping on ∈X to obtain the node projection vector, satisfying:

[0081]

[0082] Among them, z k Represents node v k The projection vector, W (u) W (c) b represents the attention transformation matrix for different types. (u) b (b) These represent different bias vectors;

[0083] S33. Construct the attention weight matrix A (att) For each pair of nodes v that are connected by an edge i With v j Calculate the weighted adjacency attention coefficient α ij The definition is as follows:

[0084]

[0085] Where, α ij Represents node v i For adjacent node v j The attention weights, where σ(·) represents the activation function. This represents the weight vector in the attention mechanism, where λ represents the interaction modulator. Represents node v i The set of adjacent nodes, [·||·] denotes the vector concatenation operation, a ij ∈A represents user node u i With behavior node b j The frequency of interaction between them, a ik ∈A represents user node u i With behavior node b k Interaction frequency between them, z i Represents node v i The projection vector, z j Represents node v j The projection vector of exp(·) represents the natural exponential function with base e;

[0086] S34, Based on attention weight matrix A (att) For each user type node Perform weighted aggregation operations on features from different types of nodes to obtain the updated user feature vector h. i The calculation formula is:

[0087]

[0088] The user feature vectors of all nodes are concatenated and fused to form a user feature map.

[0089] In this embodiment, S4 specifically includes:

[0090] S41, in the type grouping mapping set Above, we introduce the structural attention factor matrix S = {s} ij} and semantic attention factor matrix M={m ij}, where the structural attention factor s ij Based on the consistency of topological path density and connection direction, the semantic attention factor m is constructed. ij Based on the semantic similarity between node label embedding vectors, the following definitions are used:

[0091]

[0092] Where, δ ij Represents node vi With v j In the graph structure, whether a connection is represented by a direct link is indicated by a value of 1 or 0. μ represents the structure adjustment factor, β represents the directional weighting coefficient, and θ represents the directional weighting coefficient. ij The x represents the difference in steering angles along the path connecting the nodes. ik Represents node v i The eigenvalues ​​in the k-th dimension, e il Represents node v i The embedding vector value in the l-th semantic label dimension, where d represents the feature dimension, r represents the semantic embedding dimension, and η represents the smoothing term constant;

[0093] S42. Introduce a structural semantic joint factor adjustment mechanism, and combine it with the weighted adjacency attention coefficient α. ij Generate double-weighted fusion attention coefficients satisfy:

[0094]

[0095] Where, α ij Represents node v i For adjacent node v j Attention weights, α ik Represents node v i For adjacent node v k The attention weights are κ and λ, which are power-law moderating coefficients that control the degree of influence of structural and semantic factors on attention propagation, respectively. Represents node v i The set of adjacent nodes;

[0096] S43. Using the Fusion Attention Coefficient Perform propagation aggregation on the feature vectors of adjacent nodes to update the nodes in the user feature graph. The feature vectors in the dataset are used to generate the updated user feature map.

[0097] In this embodiment, S5 specifically includes:

[0098] S51, User Feature Map The input is fed into a graph-structured cascaded convolutional unit, where a local adjacency sampling operation is performed, and the sampling depth d is set. l Neighbor selection threshold τ l Construct a set of local adjacency sampling results Perform local feature convolution based on the sampled adjacent subgraph:

[0099]

[0100] in, Represents node vi The convolution output under local adjacency sampling, where σ(·) is the activation function, W (l) Let b be the local convolution weight matrix. (l) For bias vectors, To sample the number of neighboring nodes;

[0101] S52. Perform global graph topology sampling, and set the topology location encoding matrix P = {p i}, where p i For node v i The graph structure depth encoding vector is used to generate a global context-aware vector using topologically weighted convolution. The formula is as follows:

[0102]

[0103]

[0104] Where tanh(·) represents the hyperbolic tangent function, W (g) Let b represent the global convolution weight matrix. (g) This represents the bias vector. Representing user feature maps Different user feature vectors, ρ ij The normalized topological correlation coefficient is represented by γ, and the temperature regulation coefficient is represented by γ. Representing the eigenvector and The cosine similarity between them, where m represents the number of user nodes;

[0105] S53. Perform cross-domain skip sampling operation, and set the skip interval coefficient Δ and the skip mapping matrix W. (c) Grouping mapping sets by type Any node v i Feature fusion is performed during convolution of images at different scales, and the calculation formula is as follows:

[0106]

[0107] Where φ(·) represents the fusion activation function, α1, α2, and α3 represent different multi-scale weight coefficients, and f i This represents a multi-scale representation of user behavior.

[0108] In this embodiment, S6 specifically includes:

[0109] S61, representing multi-scale user behavior f i Input is given to the feature decoupling unit, and the interest decoupling transformation matrix W is set. (int) With social decoupling transformation matrix W (soc) For each multi-scale feature fusion vector fi Perform a linear projection transformation to generate the interest feature vector r. i With social feature vectors s i The calculation formula is:

[0110] r i =ReLU(w (int) ·f i +b (int) );s i =ReLU(W (soc) ·f i +b (soc) );

[0111] Among them, b (int) b (soc) For different bias terms, d f For multi-scale representation dimensions, ReLU(·) is the modified linear unit activation function;

[0112] S62. Construct an orthogonal decoupling constraint function to ensure the independence of the interest feature subspace and the social feature subspace, and define an orthogonal decoupling loss function. for:

[0113]

[0114] Where, r ik Represents node v i In the k-th dimension component of the interest feature subspace, n il Represents node v i In the l-th dimension of the social feature subspace, ∈ represents a non-zero constant, m represents the number of user nodes, and d r d s Representing different dimensional components;

[0115] S63, Regarding the set of interest vectors With social vector set Perform subspace combination to form a feature mapping vector, calculated using the following formula:

[0116] m i =φ(α·r i +β·s i +γ·(r i ⊙s i ));

[0117] Where φ(·) is the fusion activation function, α, β, and γ are weighting factors, ⊙ represents the Hadamard product of corresponding elements, and m i Represents node v i The feature mapping vectors form the user feature mapping set.

[0118] refer to Figure 3 A user management system based on big data includes:

[0119] The data processing module is used to collect user data and perform preprocessing operations on the user data;

[0120] The heterogeneous graph construction module is used to define the node set and edge set, and to build a heterogeneous graph structure containing user nodes and behavior nodes. The edge set is used to represent the interaction relationship between users and behaviors.

[0121] The feature aggregation module is used to input heterogeneous graph structures into the graph attention neural network, construct attention weight matrices according to node types, perform weighted aggregation on the adjacency relationships between different types of nodes, and output user feature maps.

[0122] The attention fusion module is used to introduce structural attention factors and semantic attention factors, perform dual attention adjustment on the adjacent edges of user nodes, construct an attention propagation mechanism that integrates structural and semantic features, and update the user feature graph.

[0123] The convolutional sampling module is used to input the updated user feature map into the graph structure cascaded convolutional unit, and perform local adjacency sampling, global graph topology sampling and cross-domain skip sampling respectively to obtain multi-scale user behavior representations;

[0124] The feature decoupling module is used to decouple the multi-scale user behavior representation from interest features and social features, forming an independent representation subspace and constructing a user feature mapping set.

[0125] The status determination module is used to perform feature aggregation based on the user feature mapping set, calculate the user status vector through the feature selection weight vector, and output the user group label and management instruction set based on the status vector.

[0126] Example 1:

[0127] To verify the feasibility of this invention in practice, it was applied to the refined user management system of a comprehensive online service platform. This platform has over 8 million daily active users, covering multiple business modules such as education and training, e-commerce, and content distribution. Due to the complexity and diversity of user behavior, traditional management methods based on static tags and rule engines have significant shortcomings in actual operation, failing to effectively identify users' true intentions and behavioral paths, leading to problems such as insufficient recommendation accuracy, delayed strategy response, and decreased user retention. The platform urgently needs an intelligent user management solution with dynamic behavioral analysis capabilities, high segmentation accuracy, and automatic strategy output capabilities.

[0128] In this embodiment, multidimensional behavioral data of platform users is first collected and preprocessed. Specific data includes identity registration information (such as gender, age group, and education level), page access history over the past six months, operation click logs, product browsing and transaction records, content liking and commenting behavior, device usage patterns, network connection types, IP activity areas and their trends, etc. On average, each user generates 38 behavioral records per day, with a total collection scale exceeding 1.8 billion records.

[0129] During the preprocessing of the raw data, a multi-model prediction method (a hybrid strategy of KNN and random forest) is used to restore the missing parts of the fields; at the same time, high-frequency abnormal operations (such as clicking the same page more than 5 times within 10 seconds) are removed using the local outlier algorithm; in the multi-terminal synchronous recording, behavior is uniformly aligned through timestamp mapping and session merging operations to ensure data order consistency and integrity.

[0130] The preprocessed data is input into the heterogeneous graph construction module. Users, products, content, and actions (clicks, purchases, comments, favorites, etc.) are used as node types in the heterogeneous graph, and different interaction behaviors are used as weighted edges. This ultimately constructs a heterogeneous graph structure containing approximately 42 million nodes and 310 million edges. After inputting the graph structure, the graph attention neural network model dynamically constructs a typified attention matrix based on node types and introduces a dual-factor adjustment mechanism of structural attention and semantic attention to ensure that user features maintain the hierarchical nature of the behavioral structure and the integrity of the semantic context during the aggregation process.

[0131] In the graph-structured cascaded convolution process, local adjacency sampling, with a range of 2-hop adjacency, focuses on capturing short-term preference features of users and their directly interacting content; global topological sampling, based on a graph Laplacian position encoder, processes cross-topic behavioral paths to extract long-term user behavior migration trends; while cross-domain jump sampling focuses on fusing jump paths of user behavior between modules (such as jumps from educational content to e-commerce products). After these three scales of collaborative sampling, the constructed user features outperform the original aggregation method in terms of representation completeness, semantic clarity, and discriminative ability.

[0132] Subsequently, through a decoupling mechanism between interest features and social features, this embodiment obtains the user's personal behavior preference space and group relationship influence space, respectively. By constructing orthogonal constraints and fusing weight vectors, a final feature mapping vector is generated for state determination and clustering strategy generation.

[0133] During the trial period, 50,000 newly registered users were selected as the test group. The method of this invention was used to construct dynamic profiles and allocate management strategies. Another 50,000 users served as the control group, using the platform's original rule engine model. During the three-week observation period, the test group performed excellently in several indicators: the recommendation click-through rate increased from 7.1% to 11.3%, an increase of 59.1%; the average daily user dwell time increased from 5.6 minutes to 8.9 minutes; the accuracy of user segmentation (based on manual annotation and behavioral backtracking) improved from 72.4% to 91.8%; the product conversion rate increased from 1.9% to 3.2%, an increase of 68.4%; and the user retention rate increased by 9.7 percentage points. The table below shows a comparison of some key data indicators:

[0134] Table 1 Comparison of User Management Effectiveness Indicators

[0135]

[0136] As can be seen from the above embodiments, the present invention has significant advantages in complex user behavior modeling, feature learning and management strategy output, and can achieve high-precision and high-real-time intelligent user management, which is suitable for personalized operation scenarios in various big data platforms.

[0137] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A big data-based user management method, characterized by, The method comprises the following steps: S1, collecting user data and performing preprocessing; S2, defining a node set and an edge set based on the preprocessed user data to form a heterogeneous graph structure, the node set comprising user nodes and behavior nodes, and the edge set representing the interaction relationship between users and behaviors; S3, inputting the heterogeneous graph structure into a graph attention neural network, constructing an attention weight matrix according to the node type, weighting and aggregating the adjacency relationship between different types of nodes, and outputting a user feature graph; S4, introducing a structure attention factor and a semantic attention factor to perform double attention adjustment on the adjacent edges of the user nodes, constructing an attention propagation mechanism that fuses structure and semantic features, and updating the user feature graph; S5, inputting the updated user feature graph into a graph structure cascaded convolution unit to perform local adjacency sampling, global graph topology sampling and cross-domain jump layer sampling, and obtaining multi-scale user behavior representation; The S5 specifically comprises: S51, inputting the user feature graph into the graph structure cascaded convolution unit to perform local adjacency sampling operation, setting the adjacency sampling depth and edge weight threshold in the operation, filtering the local neighbor node set based on the direct connection relationship of the user nodes, and obtaining the local adjacency feature representation by averaging and aggregating the filtered neighbor node feature vectors and performing convolution transformation processing; S52, performing global graph topology sampling operation, generating a corresponding graph structure position encoding vector for each user node in the operation, combining the feature vectors of all user nodes, calculating the feature similarity coefficient between nodes, weighting and summarizing all other node features through the feature similarity coefficient, and inputting the fused node features and position encoding into a fully connected neural network layer to obtain the global context-aware feature representation; S53, performing cross-domain jump layer sampling operation, setting a jump layer interval factor in the operation, extracting the difference features between the local convolution output and the global convolution output, performing affine transformation on the difference features through a jump layer mapping matrix, and weighting and fusing the local features, global features and transformed features to finally output the behavior representation result of each user node under multi-scale information fusion, i.e., multi-scale user behavior representation; S6, decoupling the interest feature and the social feature of the multi-scale user behavior representation to form independent representation subspaces, and constructing a user feature mapping set; The S6 specifically comprises: S61, inputting the multi-scale user behavior representation into a feature decoupling unit, setting an interest decoupling transformation matrix and a social decoupling transformation matrix respectively, and performing two sets of linear transformation operations on the multi-scale behavior feature vector of each user node, wherein the first set of transformation is used to extract the interest feature representation of the user, and the second set of transformation is used to extract the social relationship feature representation of the user, and the interest feature vector and the social feature vector are output respectively to form two sets of feature representation vectors with different dimensions; S62, construct an orthogonal constraint mechanism for the interest feature vector and the social feature vector, measure the correlation between the two by calculating the dot product sum of the two vector subspaces in the feature dimension and the norm normalization value, and minimize the square value of the similarity as the optimization target, explicitly suppress the coupling term between the interest feature and the social feature, realize the information isolation between different subspaces, and ensure the mutual independence of the interest feature subspace and the social feature subspace in the feature representation process; S63, perform subspace combination on the interest feature vector and the social feature vector optimized by the orthogonal constraint, respectively apply weight factor to control the influence proportion of the two in the fusion process, and perform product operation on the corresponding positions of the elements, nonlinearly combine the interest feature and the social feature, obtain the fused user feature mapping vector, and collect the feature mapping vectors corresponding to all user nodes to form a user feature mapping set; S7, based on the user feature mapping set, perform feature aggregation, calculate the user state vector through the feature selection weight vector, and output the user clustering label and the management instruction set according to the state vector. 2.The user management method based on big data according to claim 1, characterized in that, The user data includes identity information, behavior log, transaction record, preference label, device fingerprint and access trajectory. 3.The user management method based on big data of claim 1, wherein, The preprocessing includes missing value filling, filtering denoising, format conversion, outlier removal and timestamp alignment. 4.The user management method based on big data according to claim 1, characterized in that, The S2 specifically includes: S21, divide the preprocessed user data according to attribute categories, respectively extract identity information data, behavior log data, transaction record data, preference label data, device fingerprint data and access trajectory data, and perform feature vectorization processing on each type of data to construct a standardized feature data set; S22, according to the standardized feature data set, construct a node set, the node set includes a plurality of user nodes and a plurality of behavior nodes, wherein each user node corresponds to a unique user, and each behavior node corresponds to a user behavior type; S23, according to the association record between users and behaviors in the original data, construct an edge set, each edge in the edge set represents an interaction relationship between a user node and a behavior node, and the weight of the edge is set based on the interaction frequency and behavior intensity features; S24, according to the connection relationship between the user nodes, the behavior nodes and the edge set, construct an interaction matrix, each element in the interaction matrix represents a specific interaction value between a user node and a behavior node; S25, integrate the node set, the edge set and the interaction matrix, classify the nodes according to the node type, assign the edge weight and construct the topology structure according to the construction principle, establish a heterogeneous graph structure, and complete the graph model initialization. 5.The user management method based on big data according to claim 1, characterized in that, The S3 specifically includes: S31, input the heterogeneous graph structure into the graph attention neural network, and divide the user type node set and the behavior type node set according to the node type respectively, and establish the node type mapping relationship; S32, select the corresponding attention transformation matrix according to the type of the node, perform linear transformation on the initial feature vector of each node to generate a typed node representation vector; S33, for any user type node, traverse its adjacent behavior type node, use the similarity relationship between the user type node representation vector and the adjacent behavior type node representation vector to construct the attention weight matrix, and perform normalization processing on each attention weight value; S34, weight aggregation is performed on all adjacent behavior type nodes of the user type node according to the normalized attention weight value, the representation vector of all adjacent nodes is multiplied by the corresponding attention weight to obtain the updated feature representation of the user type node; S35, perform weighted aggregation and update operation on all user type nodes, and output the user feature graph. 6.The user management method based on big data according to claim 1, wherein, The S4 specifically comprises: S41, respectively construct a structural attention factor and a semantic attention factor in the user type node set, which are used to measure the correlation degree between the user node and its adjacent behavior node in terms of structural connection strength and semantic label similarity; S42, double factor adjustment is performed on the attention weight matrix, the structural attention factor and the semantic attention factor are introduced as multiplicative adjustment factors on the basis of the original attention weight, and the weight of all adjacent nodes of each user node is normalized to form an attention weight matrix that integrates structural and semantic features; S43, use the updated attention weight matrix to perform weighted aggregation operation on the adjacent behavior node feature vector of the user node, complete the feature update of the user node, generate the user feature representation vector integrating structural and semantic features, and update the user feature graph.

7. The user management method based on big data according to claim 6, characterized in that, The structural attention factor is quantified by analyzing the topological path density, connection direction consistency and feature vector dot product relationship between the user node and the adjacent node in the graph, and the semantic attention factor is based on the embedding vector of each node in the semantic label space, and the semantic matching degree between nodes is calculated by combining the cosine similarity. 8.A big data based user management system, which executes the big data based user management method according to any one of claims 1 to 7. Comprise: A data processing module is configured to collect user data and perform preprocessing operation on the user data; A heterogeneous graph construction module is configured to define a node set and an edge set, construct a heterogeneous graph structure comprising user nodes and behavior nodes, and the edge set is used to represent the interaction relationship between users and behaviors; A feature aggregation module is configured to input the heterogeneous graph structure into a graph attention neural network, construct an attention weight matrix according to the node type, perform weighted aggregation on the adjacent relationship between different types of nodes, and output a user feature graph; An attention fusion module is configured to introduce a structural attention factor and a semantic attention factor, perform double attention adjustment on the adjacent edges of the user node, construct an attention propagation mechanism integrating structural and semantic features, and update the user feature graph; A convolution sampling module is configured to input the updated user feature graph into a graph structure cascading convolution unit to perform local adjacent sampling, global graph topology sampling and cross-domain jump layer sampling respectively, and obtain multi-scale user behavior representation, comprising: The user feature map is input into a graph structure cascaded convolution unit, a local adjacency sampling operation is performed, an adjacency sampling depth and an edge weight threshold are set in the operation, a local neighbor node set is filtered based on a direct connection relationship of a user node, and a local adjacency feature representation is obtained by performing average aggregation and convolution transformation processing on a filtered neighbor node feature vector; A global graph topology sampling operation is performed, a corresponding graph structure position encoding vector is generated for each user node in the operation, a feature similarity coefficient between nodes is calculated by combining feature vectors of all user nodes, all other node features are weighted and aggregated through the feature similarity coefficient, and a global context-aware feature representation is obtained by inputting the node features and the position encoding vector into a fully connected neural network layer; A cross-domain skip layer sampling operation is performed, a skip layer interval factor is set in the operation, difference features between local convolution outputs and global convolution outputs are extracted, the difference features are affine transformed through a skip layer mapping matrix, and local features, global features, and transformed features are weighted and fused, and finally, a behavior representation result of each user node under multi-scale information fusion is output, i.e., a multi-scale user behavior representation; A feature decoupling module is configured to decouple interest features and social features from the multi-scale user behavior representation, form independent representation subspaces, and construct a user feature mapping set, including: The multi-scale user behavior representation is input into a feature decoupling unit, an interest decoupling transformation matrix and a social decoupling transformation matrix are set, and two sets of linear transformation operations are performed on the multi-scale behavior feature vector of each user node, wherein the first set of transformations is used to extract interest feature representations of the user, and the second set of transformations is used to extract social relationship feature representations of the user, and interest feature vectors and social feature vectors are output, forming two sets of feature representation vectors with different dimensions; An orthogonal constraint mechanism is constructed for the interest feature vectors and the social feature vectors, the correlation between the two vector subspaces in the feature dimension is measured by calculating the dot product sum and the norm normalization value, and the optimization objective is to minimize the square value of the similarity, which explicitly suppresses the coupling term between the interest features and the social features, realizes information isolation between different subspaces, and ensures the mutual independence of the interest feature subspace and the social feature subspace in the feature representation process; The interest feature vectors and the social feature vectors optimized by the orthogonal constraint are combined in a subspace, weight factors are applied to control the influence proportion of the two in the fusion process, and product operations are performed on corresponding element positions, non-linear combinations of interest features and social features are performed, user feature mapping vectors are obtained, and all user node corresponding feature mapping vectors are aggregated to form a user feature mapping set; A state determination module is configured to aggregate features based on the user feature mapping set, calculate a user state vector through a feature selection weight vector, and output a user clustering label and a management instruction set according to the state vector.

Images