A large-scale communication network data vectorization processing method and system

Abstract

This invention relates to the field of large-scale graph computing and communication user profiling technology, specifically to a method and system for vectorizing large-scale communication network data. The method includes acquiring user attribute data, call record data, product order data, and application usage behavior data; constructing a heterogeneous graph containing user nodes, product nodes, and application nodes; performing coarse-grained community partitioning of the heterogeneous graph according to business attributes, and then performing fine-grained edge partitioning within each community using a multi-level graph partitioning algorithm to obtain several subgraphs; first, performing probability-weighted sampling on the nodes in each subgraph based on node degree distribution, and then finely filtering the sampling results based on business rules to generate a set of subgraph nodes; performing sparse computation on each subgraph in a distributed computing environment based on GPU sparse matrix computation kernel functions, and employing a pipelined communication mode to overlap the computation and communication processes. This invention achieves real-time vectorization processing of industrial-grade large-scale structured communication data.

Description

Technical Field

[0001] This invention relates to the field of large-scale graph computing and communication user profiling technology, specifically to a method and system for vectorizing data in large-scale communication networks. Background Technology

[0002] With the widespread adoption of mobile communication networks and smart terminals, operators have accumulated massive amounts of user data over the long term, including basic attributes, call records, product subscription information, and application usage behavior. User relationship networks built upon this data typically have the following characteristics: the number of nodes can reach hundreds of millions, the number of edges can reach tens of billions, and there is a complex heterogeneous structure composed of multiple types of nodes and edges.

[0003] In such scenarios, graph neural networks can uncover high-order relationships between users and between users and products through multi-layered message passing and neighborhood aggregation mechanisms, providing powerful representation capabilities for user profiling and strategy recommendation. However, traditional graph neural networks face many challenges when dealing with ultra-large-scale graph structures:

[0004] High memory and storage pressure: Full storage of adjacency matrices and feature matrices with hundreds of millions of nodes and tens of billions of edges consumes a large amount of memory and video memory;

[0005] High computational latency: During training and inference, the computational complexity of adjacency matrix aggregation and multi-layer message passing is high, and the time taken for a single iteration is long, making it difficult to meet the business requirements for real-time performance and throughput.

[0006] Insufficient scalability: Traditional single-machine or simple distributed solutions are difficult to effectively handle horizontal scaling and are prone to performance bottlenecks due to uneven partitioning and frequent cross-partition communication. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, this invention provides a method and system for vectorizing large-scale communication network data. By employing a hybrid partitioning strategy driven by business attributes, a two-stage sampler, and GPU sparse computing and pipelined communication mechanisms, it effectively overcomes the computational efficiency and scalability bottlenecks in the training and inference processes of large-scale graph neural networks. While ensuring the semantics and representation accuracy of the graph structure, it significantly reduces the time and space overhead of adjacency matrix operations, enabling real-time vectorization processing of industrial-grade large-scale structured communication data.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0009] In a first aspect, the present invention proposes a method for vectorization processing of data in large-scale communication networks, comprising the following steps:

[0010] Acquire user attribute data, call record data, product order data, and application usage behavior data to construct a heterogeneous graph containing user nodes, product nodes, and application nodes;

[0011] The heterogeneous graph is divided into coarse-grained communities based on business attributes, and then fine-grained edge partitioning is performed within each community using a multi-level graph partitioning algorithm to obtain several subgraphs.

[0012] First, the nodes in each subgraph are sampled using probability weighting based on the node degree distribution. Then, the sampling results are refined by combining business rules to generate a set of subgraph nodes.

[0013] In a distributed computing environment, sparse computation is performed on each subgraph based on the GPU sparse matrix computation kernel function, and the computation and communication process is overlapped using a pipelined communication mode.

[0014] Preferably, constructing a heterogeneous graph including user nodes, product nodes, and application nodes includes:

[0015] By mapping user, product, and application entities to different types of nodes in the heterogeneous graph, and mapping call, order, and click behaviors to corresponding edge types, a heterogeneous graph containing user nodes, product nodes, and application nodes is obtained.

[0016] Preferably, the optimization objective function of the multi-level graph partitioning algorithm includes two terms: cross-partition edge weights and deviation from partition size. The cross-partition edge weights are used to measure the communication overhead between partitions, and the partition size deviation is used to measure the difference between the number of nodes in each subgraph and the average number of nodes.

[0017] Preferably, the multi-level graph partitioning algorithm adopts a layer-by-layer shrinkage and projection optimization approach, including a graph shrinkage process, a partitioning result projection, and a local optimization process.

[0018] Preferably, when performing probability-weighted sampling of nodes in each subgraph based on the node degree distribution, the degree of each candidate node in the subgraph is calculated, and a monotonically increasing and sublinear function of the degree is used as the weight to define the probability of the node being sampled.

[0019] Preferably, when performing refined screening of the sampling results in conjunction with business rules, the sampled node set is screened according to preset business rules, retaining nodes with high call frequency, high value level or closely related to the target task, and eliminating noise and irrelevant nodes.

[0020] Preferably, in a distributed computing environment, sparse computation on each subgraph based on the GPU sparse matrix computation kernel function includes:

[0021] Each subgraph is divided into different computing nodes and GPUs. A CUDA-based GPU sparse matrix operation kernel function is used to perform sparse computation on the adjacency matrix and feature matrix. For the adjacency matrix and feature matrix of each subgraph, a sparse matrix storage format is used to store only non-zero elements. During the neighbor aggregation process, a CUDA-based sparse matrix multiplication kernel function is called to perform multiplication and addition operations on non-zero elements, skipping the calculation of zero-value weights or gradients.

[0022] Preferably, the pipelined communication mode for overlapping computation and communication includes:

[0023] Gradient synchronization, parameter update, and intermediate embedded vector transmission are divided into multiple communication stages, which are interleaved with the computation stages of forward and backward propagation to construct a pipelined execution process, so that the communication stage and the computation stage overlap in time.

[0024] Preferably, it further includes:

[0025] For each subgraph, perform forward and backward propagation of the graph neural network to generate node embedding vectors, and update the embeddings and predict the output for newly added or specified nodes during the online inference phase.

[0026] The updated user nodes are embedded and written into the user profile database, and the embedded vectors and prediction results are provided to the precision marketing, customer retention, and churn warning business systems through the online service interface.

[0027] Secondly, this invention proposes a large-scale communication network data vectorization processing system, which applies the large-scale communication network data vectorization processing method described above, including:

[0028] The graph construction module is used to acquire user attribute data, call record data, product order data, and application usage behavior data, and construct a heterogeneous graph containing user nodes, product nodes, and application nodes.

[0029] The partitioning module is used to perform coarse-grained community partitioning of heterogeneous graphs according to business attributes, and to perform fine-grained edge partitioning within the community using a multi-level graph partitioning algorithm to obtain several subgraphs;

[0030] The sampling module is used to first perform probability-weighted sampling on the nodes in each subgraph based on the node degree distribution, and then combine the sampling results with business rules to refine the filtering and generate a set of subgraph nodes.

[0031] The computation acceleration module is used to perform sparse computation on each subgraph based on the GPU sparse matrix computation kernel function in a distributed computing environment, and to overlap the computation and communication process using a pipelined communication mode.

[0032] The present invention has the following beneficial effects:

[0033] (1) This invention combines business attribute-driven hybrid partitioning with multi-level graph partitioning algorithm to effectively reduce cross-partition edges, reduce communication overhead in distributed training, maintain load balance among partitions, and improve system scalability.

[0034] (2) This invention uses a two-stage sampler based on node degree distribution and business rules to significantly reduce the number of samples and computational burden while retaining the nodes and edges that are most critical to the training effect, thereby improving the representativeness of the sampled data and the training efficiency.

[0035] (3) This invention introduces GPU sparse matrix computation and pipelined communication mechanism to achieve efficient sparse computation of adjacency matrix aggregation operation, and performs the computation and communication processes in an overlapping manner, making full use of GPU and other computing resources, and significantly improving the training and inference speed of billion-level graph neural networks;

[0036] (4) This invention can support real-time vectorization processing of industrial-grade large-scale structured data, providing high-performance infrastructure for user profiling and strategy recommendation based on multimodal data, and is applicable to large-scale graph computing scenarios such as telecommunications operations and Internet platforms. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of a large-scale communication network data vectorization processing method.

[0038] Figure 2 This is a schematic diagram of a large-scale communication network data vectorization processing system. Detailed Implementation

[0039] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0040] When faced with heterogeneous user relationship networks with hundreds of millions of nodes and tens of billions of edges in China Unicom's business scenarios, traditional graph neural networks are prone to problems such as high memory consumption, large computational latency, and difficulty in scaling during training and inference, making it difficult to support the real-time vectorization processing requirements of industrial-grade scenarios. This invention proposes a three-level optimization framework of "partitioning-sampling-computation": In the partitioning stage, a hybrid partitioning strategy based on business attributes is combined with a multi-level graph partitioning algorithm. First, communities are divided according to business attributes such as region. Then, within each community, a multi-level graph partitioning method with minimum cut and load balancing objectives is used to reduce cross-partition edges and balance the computational load. In the sampling stage, a two-stage sampler is designed. First, probability-weighted sampling is performed based on the node degree distribution to suppress the excessive amplification of high-value nodes. Then, the sampling results are finely adjusted by combining business rules such as call frequency and package type to retain high-value nodes and filter irrelevant nodes. In the computation stage, a CUDA-based GPU sparse matrix operation kernel function is used to perform sparse computation on the naturally sparse adjacency matrix and feature matrix. A pipelined communication mechanism is constructed to split the data transmission and computation process in distributed training into multiple stages and execute them overlappingly, reducing the idle waiting time of computing resources such as GPUs.

[0041] like Figure 1 As shown, an embodiment of the present invention provides a method for vectorizing data in a large-scale communication network, comprising the following steps S1 to S4:

[0042] S1. Obtain user attribute data, call record data, product order data, and application usage behavior data, and construct a heterogeneous graph containing user nodes, product nodes, and application nodes;

[0043] In an optional embodiment of the present invention, step S1, constructing a heterogeneous graph including user nodes, product nodes, and application nodes, includes:

[0044] By mapping user, product, and application entities to different types of nodes in the heterogeneous graph, and mapping call, order, and click behaviors to corresponding edge types, a heterogeneous graph containing user nodes, product nodes, and application nodes is obtained.

[0045] This embodiment obtains structured data such as basic user attributes, call records, and product ordering information from telecommunications operators, and application usage behavior data such as user clicks, browsing, searching, and dwell time from the APP log system. It performs missing value imputation, outlier processing, and format unification on the original data, maps entities such as users, products, and applications to different types of nodes in a heterogeneous graph, and maps behaviors such as calls, orders, and clicks to corresponding edge types, thus obtaining a heterogeneous graph structure containing hundreds of millions of nodes and tens of billions of edges.

[0046] S2. Divide the heterogeneous graph into coarse-grained communities according to business attributes, and then perform fine-grained edge partitioning within the communities using a multi-level graph partitioning algorithm to obtain several subgraphs.

[0047] In an optional embodiment of the present invention, the optimization objective function of the multi-level graph partitioning algorithm used in step S2 includes two terms: cross-partition edge weight sum and partition size deviation. The cross-partition edge weight sum is used to measure the communication overhead between partitions, and the partition size deviation is used to measure the difference between the number of nodes in each subgraph and the average number of nodes.

[0048] The multi-level graph partitioning algorithm employs a layer-by-layer shrinkage and projection optimization approach, including graph shrinkage, partitioning result projection, and local optimization.

[0049] This embodiment first performs coarse-grained community partitioning of the heterogeneous graph based on business attributes such as user region or business area. The entire graph is divided into several communities according to business attributes, so that users and their relationships within the same region or business area are preferentially assigned to the same area, preserving the tightness of the community structure. Within each community, a multi-level graph partitioning algorithm is used for fine-grained edge partitioning. A coarse graph is generated by shrinking the graph layer by layer. The partitioning problem is solved at the coarsest layer. The partitioning results are then projected back to the original graph layer by layer. Local optimization is used to reduce the cross-partition edge weights and deviations from the partition size, decomposing the graph into several subgraphs of moderate size, with fewer cross-partition edges and a relatively balanced number of nodes, in order to reduce cross-partition communication overhead and balance the load of distributed training.

[0050] On the global heterogeneous graph, coarse-grained communities are first divided according to attributes such as user region or business area, preserving the tight structure of the communities. Then, within each community, a multi-level graph partitioning algorithm is used, with "sum of cross-partition edge weights + partition size deviation" as the optimization objective. Graph shrinkage, coarse graph partitioning, partitioning result projection and local optimization are performed respectively to finally obtain a high-quality partition with fewer cross-partition edges and balanced load across partitions.

[0051] Let the set of nodes of the entire graph be . The edge set is Divide the graph into There are several subgraphs, and the node sets of each subgraph are denoted as follows: Each edge The weight is The objective function for partitioning is then defined as:

[0052] ;

[0053] in, This represents the set of edges that are cut off and connected to different subgraphs. The first term is the sum of edge weights across partitions, used to measure the communication overhead between partitions; the second term is the partition size deviation, used to measure the difference between the number of nodes in each subgraph and the average number of nodes. The gap; This is a balancing factor used to weigh the trade-off between reducing communication overhead and maintaining load balance. The multi-level graph partitioning algorithm employs a layer-by-layer shrinkage and projection optimization approach.

[0054] (1) The process of graph shrinkage:

[0055] Record No. Layer diagram The matching set is Where each matching pair consists of adjacent nodes, the shrinking operation is represented as:

[0056] ;

[0057] The graph size is reduced layer by layer by merging nodes in matching pairs into supernodes.

[0058] (2) Projection of partitioning results and local optimization process:

[0059] After the initial partitioning is completed at the coarsest layer, projection and optimization are performed layer by layer upwards, denoted as the first... The result of the layer partitioning is Then in the first The result of layer partitioning can be represented as:

[0060] ;

[0061] in, This represents the projection operation that maps the coarse map partitioning results back to the current layer node. As a local optimization process, the partitioning objective function is further reduced through operations such as node migration. This process continues until the original image is restored and a high-quality partitioning result is obtained.

[0062] S3. First, perform probability-weighted sampling on the nodes in each subgraph based on the node degree distribution, and then refine the sampling results by combining business rules to generate a set of subgraph nodes.

[0063] In an optional embodiment of the present invention, when performing probability-weighted sampling of nodes in each subgraph based on the node degree distribution in step S3, the degree of each candidate node in the subgraph is calculated, and a monotonically increasing and sublinear function with respect to the degree is used as the weight to define the probability of the node being sampled.

[0064] When performing refined filtering of sampling results in conjunction with business rules, the sampled node set is filtered according to preset business rules, retaining nodes with high call frequency, high value level or closely related to the target task, and removing noise and irrelevant nodes.

[0065] This embodiment designs a two-stage sampler for the training nodes in each subgraph:

[0066] The first stage performs probability-weighted sampling based on the node degree distribution. By appropriately attenuating the sampling probability of height nodes, it avoids oversampling them during training, while retaining nodes that are representative of the overall structure.

[0067] The second stage involves refining and supplementing the node set based on business rules (including but not limited to call frequency, package type, and user value level) on the coarse sampling results. Important business nodes are retained first, while noisy nodes are removed, thereby reducing the amount of computational data while ensuring the representativeness and effectiveness of the sampled data.

[0068] In this embodiment, a two-stage sampler is used within each subgraph to select the central node and its neighbors for training. The first stage performs probability-weighted sampling based on node degree distribution, appropriately attenuating the sampling probability of high-degree nodes to prevent training from being dominated by a few high-degree nodes. The second stage refines the sampled nodes according to business rules such as call frequency, package type, and user value level, retaining business-important nodes and removing noisy nodes, thereby reducing the size of the training data while maintaining representativeness.

[0069] In the first stage, when performing probability-weighted sampling based on node degree distribution, for each candidate node in the subgraph... Calculate degrees It employs a monotonically increasing and sublinear function with respect to degree. As a weight, the probability of a node being sampled is defined as:

[0070] ;

[0071] in, Given a set of candidate nodes, the function The weights of height nodes are appropriately attenuated so that the sampling results can cover key nodes while avoiding excessive amplification of height nodes during training.

[0072] In the second stage, the sampled node set is filtered according to the preset business rules, retaining nodes with high call frequency, high value level or closely related to the target task, and eliminating noise and irrelevant nodes.

[0073] S4. In a distributed computing environment, sparse computation is performed on each subgraph based on the GPU sparse matrix computation kernel function, and the computation and communication process is overlapped using a pipelined communication mode.

[0074] In an optional embodiment of the present invention, step S4, which performs sparse computation on each subgraph based on the GPU sparse matrix computation kernel function in a distributed computing environment, includes:

[0075] Each subgraph is divided into different computing nodes and GPUs. A CUDA-based GPU sparse matrix operation kernel function is used to perform sparse computation on the adjacency matrix and feature matrix. For the adjacency matrix and feature matrix of each subgraph, a sparse matrix storage format is used to store only non-zero elements. During the neighbor aggregation process, a CUDA-based sparse matrix multiplication kernel function is called to perform multiplication and addition operations on non-zero elements, skipping the calculation of zero-value weights or gradients.

[0076] The pipelined communication mode for overlapping computation and communication includes:

[0077] Gradient synchronization, parameter update, and intermediate embedded vector transmission are divided into multiple communication stages, which are interleaved with the computation stages of forward and backward propagation to construct a pipelined execution process, so that the communication stage and the computation stage overlap in time.

[0078] In this embodiment, in a distributed training environment, each subgraph is divided across different computing nodes and GPUs. A CUDA-based sparse matrix operation kernel function is used to perform sparse computation on the adjacency matrix and feature matrix of the subgraph, performing multiplication and addition operations only on non-zero elements. This significantly reduces the time and space overhead of adjacency matrix aggregation computation, thereby reducing memory usage and computational load. At the same time, a pipelined communication mode is adopted, which breaks down the parameter synchronization and intermediate embedding transmission process into multiple stages and executes them in parallel with the forward and backward propagation computation stages. This achieves overlap between communication and computation, reducing the idle time of computing resources such as GPUs.

[0079] In this embodiment, under a distributed training architecture, each subgraph is divided across different computing nodes and GPUs. A CUDA-based GPU sparse matrix operation kernel function is used to perform sparse computation on the adjacency matrix and feature matrix. For the adjacency matrix and feature matrix of each subgraph, a sparse matrix storage format is used, storing only non-zero elements. During the neighbor aggregation process, a CUDA-based sparse matrix multiplication kernel function is called to perform multiplication and addition operations on non-zero elements, skipping the calculation of zero-value weights or gradients, significantly reducing computational load and GPU memory usage. At the same time, a pipelined communication mechanism is constructed. In distributed training, gradient synchronization, parameter updates, and intermediate embedding vector transmission are divided into multiple communication stages, which are interleaved with the forward and backward propagation computation stages to construct a pipelined execution flow. This allows the communication stage and the computation stage to overlap in time, reducing the idle time of GPUs and other computing resources due to waiting for communication, and improving overall throughput and real-time inference performance.

[0080] Graph data is typically very sparse (a node connects to only a few neighbors, and the adjacency matrix is ​​mostly 0).

[0081] Sparse storage format: In GPU memory, the complete adjacency matrix and feature matrix are not stored. Instead, a compressed format is used to store only non-zero elements.

[0082] CUDA sparse matrix multiplication kernel function design: "Neighbor aggregation" is essentially the multiplication of a sparse matrix with a dense eigenvalue matrix. This is the kernel function.

[0083] In the CUDA kernel function, each node in the graph is assigned to a computation thread. The computation thread can read neighboring nodes and node features in parallel. Because the final computation is based on a sparse storage format, it will only contain non-zero elements, so the computation of zero-value weights or gradients can be skipped.

[0084] Pipeline communication mechanism: Based on GPUs, which have independent computing and replication engines.

[0085] Pipeline partitioning: The computation steps are divided into pipelines according to each layer of the neural network, and the communication steps are divided into pipelines according to the order of gradient synchronization, parameter update and intermediate embedding vector transmission.

[0086] For computational flow: the core matrix multiplication kernel function responsible for performing forward / backward propagation.

[0087] For communication streams: responsible for performing operations such as gradient synchronization, parameter updates, and intermediate embedding vector transfer.

[0088] By using CPU scheduling, the results of the computation stream can be directly handed over to the communication stream for transmission and synchronization, while the computation stream can perform subsequent calculations simultaneously, achieving overlap between computation and communication.

[0089] In an optional embodiment of the present invention, the method further includes steps S5 and S6:

[0090] S5. Perform forward and backward propagation of the graph neural network for each subgraph to generate node embedding vectors, and update the embeddings and predict the output of newly added or specified nodes during the online inference phase.

[0091] S6. Embed the updated user nodes into the user profile database, and provide the embedding vectors and prediction results to the precision marketing, customer retention, and churn warning business systems through the online service interface.

[0092] In this embodiment, under the three-level optimization framework of "partitioning-sampling-computation", forward and backward propagation of graph neural networks are performed for each subgraph. Node embedding vectors are generated using sparse aggregation of adjacency matrices and multi-layer message passing mechanism. After convergence, the node embeddings of each subgraph are mapped back to the global graph. In the online inference stage, the same subgraph partitioning and sampling strategy is used to efficiently update the embeddings and predict the output of newly added or specified nodes.

[0093] The updated user nodes are embedded into the user profile database, and the embedded vectors and prediction results are provided to business systems such as precision marketing, customer retention, and churn warning through online service interfaces, enabling real-time vectorization processing and online services for large-scale structured communication data.

[0094] In this embodiment, under the above three-level optimization framework, graph neural network training is performed on each subgraph, neighbor information is aggregated layer by layer and node embeddings are generated; during the inference phase, the same partitioning and sampling strategy is used to perform efficient inference on newly added or specified nodes, avoiding full graph recalculation.

[0095] The trained nodes are embedded back into the user profile database, and high-quality vector features and prediction results are provided to business systems such as marketing recommendation, customer retention and churn warning through standardized API interfaces.

[0096] This "partition-sampling-computation" three-level optimization framework is suitable for graph computation scenarios with hundreds of millions of nodes and tens of billions of edges. By prioritizing a multi-level partitioning strategy that combines edge partitioning and community partitioning, it reduces cross-partition edges. Combined with a sampler jointly driven by node degree distribution and business rules, it ensures that the semantics and distribution characteristics of the graph structure are preserved while reducing the amount of computational data. Combined with GPU sparse computing and pipelined communication, it achieves real-time vectorized processing of large-scale graph computation.

[0097] This invention also provides a large-scale communication network data vectorization processing system, which applies a large-scale communication network data vectorization processing method as described in the above embodiments, such as... Figure 2 As shown, it includes:

[0098] The graph construction module is used to acquire user attribute data, call record data, product order data, and application usage behavior data, and construct a heterogeneous graph containing user nodes, product nodes, and application nodes.

[0099] The partitioning module is used to perform coarse-grained community partitioning of heterogeneous graphs according to business attributes, and to perform fine-grained edge partitioning within the community using a multi-level graph partitioning algorithm to obtain several subgraphs;

[0100] The sampling module is used to first perform probability-weighted sampling on the nodes in each subgraph based on the node degree distribution, and then combine the sampling results with business rules to refine the filtering and generate a set of subgraph nodes.

[0101] The computation acceleration module is used to train and infer graph neural networks for each subgraph in a distributed computing environment based on GPU sparse matrix computation kernel functions, and to overlap the computation and communication processes using a pipelined communication mode.

[0102] The partitioning module includes:

[0103] Attribute pre-partitioning units are used to coarsely divide the entire map according to user region or business area, forming multiple communities;

[0104] Multi-level partitioning units are used to perform multi-level graph partitioning within a single community, including graph shrinking, coarse graph partitioning, partitioning result projection and local optimization, to obtain subgraph partitioning results with fewer cross-partition edges and balanced node size.

[0105] The sampling module includes:

[0106] Degree base sampling unit is used to perform probability-weighted sampling based on the node degree distribution, and the sampling probability of height nodes is appropriately attenuated;

[0107] The rule fine-tuning unit is used to refine and supplement the sampling results based on business rules such as call frequency, package type, and user value level.

[0108] In an optional embodiment of the present invention, the system further includes:

[0109] The graph neural network training and inference module is used to perform forward and backward propagation of the graph neural network for each subgraph, generate node embedding vectors, and update the embeddings and predict the output of newly added or specified nodes during the online inference phase.

[0110] The embedded output and service module is used to embed the updated user nodes into the user profile library and provide the embedded vectors and prediction results to the precision marketing, customer retention, and churn warning business systems through the online service interface.

[0111] This invention significantly reduces the time and space overhead of adjacency matrix aggregation without requiring frequent retraining of the entire graph, improves the throughput and response speed of training and inference of graph neural networks with hundreds of millions of nodes, supports real-time vectorization processing of industrial-grade large-scale structured communication data, and provides a high-performance computing foundation for multimodal user profiling and real-time strategy recommendation.

[0112] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0113] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0114] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0115] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0116] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for data vectorization processing in a large-scale communication network, characterized in that, Includes the following steps: Acquire user attribute data, call record data, product order data, and application usage behavior data to construct a heterogeneous graph containing user nodes, product nodes, and application nodes; The heterogeneous graph is divided into coarse-grained communities based on business attributes, and then fine-grained edge partitioning is performed within each community using a multi-level graph partitioning algorithm to obtain several subgraphs. First, the nodes in each subgraph are sampled using probability weighting based on the node degree distribution. Then, the sampling results are refined by combining business rules to generate a set of subgraph nodes. In a distributed computing environment, sparse computation is performed on each subgraph based on the GPU sparse matrix computation kernel function, and the computation and communication process is overlapped using a pipelined communication mode.

2. The method of claim 1, wherein, Constructing a heterogeneous graph that includes user nodes, product nodes, and application nodes includes: By mapping user, product, and application entities to different types of nodes in the heterogeneous graph, and mapping call, order, and click behaviors to corresponding edge types, a heterogeneous graph containing user nodes, product nodes, and application nodes is obtained.

3. The method of claim 1, wherein, The optimization objective function of the multi-level graph partitioning algorithm includes two terms: the cross-partition edge weight sum and the partition size deviation. The cross-partition edge weight sum is used to measure the communication overhead between partitions, and the partition size deviation is used to measure the difference between the number of nodes in each subgraph and the average number of nodes.

4. The method of claim 1, wherein, The multi-level graph partitioning algorithm employs a layer-by-layer shrinkage and projection optimization approach, including graph shrinkage, partitioning result projection, and local optimization.

5. The method of claim 1, wherein, When sampling nodes in each subgraph using probability weighting based on the node degree distribution, the degree of each candidate node in the subgraph is calculated, and a monotonically increasing and sublinear function of degree is used as the weight to define the probability of a node being sampled.

6. The method of claim 1, wherein, When performing refined filtering of sampling results in conjunction with business rules, the sampled node set is filtered according to preset business rules, retaining nodes with high call frequency, high value level or closely related to the target task, and removing noise and irrelevant nodes.

7. The method of claim 1, wherein, In a distributed computing environment, sparse computation of each subgraph based on the GPU sparse matrix computation kernel function includes: Each subgraph is divided into different computing nodes and GPUs. A CUDA-based GPU sparse matrix operation kernel function is used to perform sparse computation on the adjacency matrix and feature matrix. For the adjacency matrix and feature matrix of each subgraph, a sparse matrix storage format is used to store only non-zero elements. During the neighbor aggregation process, a CUDA-based sparse matrix multiplication kernel function is called to perform multiplication and addition operations on non-zero elements, skipping the calculation of zero-value weights or gradients.

8. The method of claim 1, wherein, The pipelined communication mode for overlapping computation and communication includes: Gradient synchronization, parameter update, and intermediate embedded vector transmission are divided into multiple communication stages, which are interleaved with the computation stages of forward and backward propagation to construct a pipelined execution process, so that the communication stage and the computation stage overlap in time.

9. The method of claim 1, wherein, Also includes: For each subgraph, perform forward and backward propagation of the graph neural network to generate node embedding vectors, and update the embeddings and predict the output for newly added or specified nodes during the online inference phase. The updated user nodes are embedded and written into the user profile database, and the embedded vectors and prediction results are provided to the precision marketing, customer retention, and churn warning business systems through the online service interface.

10. A large-scale communication network data vectorization processing system, applying a large-scale communication network data vectorization processing method as claimed in any one of claims 1 to 9, characterized in that, include: The graph construction module is used to acquire user attribute data, call record data, product order data, and application usage behavior data, and construct a heterogeneous graph containing user nodes, product nodes, and application nodes. The partitioning module is used to perform coarse-grained community partitioning of heterogeneous graphs according to business attributes, and to perform fine-grained edge partitioning within the community using a multi-level graph partitioning algorithm to obtain several subgraphs; The sampling module is used to first perform probability-weighted sampling on the nodes in each subgraph based on the node degree distribution, and then combine the sampling results with business rules to refine the filtering and generate a set of subgraph nodes. The computation acceleration module is used to perform sparse computation on each subgraph based on the GPU sparse matrix computation kernel function in a distributed computing environment, and to overlap the computation and communication process using a pipelined communication mode.

Images