Big data-based online platform service data mining method

Abstract

The application discloses an online platform service data mining method based on big data, and relates to the technical field of big data processing.The method maps structured, semi-structured and unstructured service data to a hyperbolic space to construct a dynamic hyperbolic graph index, realizes bottom-layer cross-modal semantic alignment, detects semantic distribution changes in a streaming data writing channel and triggers index local subgraph reconstruction, so that the index structure has dynamic perception and incremental evolution capabilities, virtual implicit associated edges are generated for sparse service nodes by using limited random walk and graph attention mechanism, and dense features are completed, thereby relieving the problem of mining effect decline in the cold start scene.The application can reduce the calculation and storage resource consumption of index updating, improve the real-time performance and accuracy of online platform service data mining, and can be used for big data analysis and mining of an industrial internet platform.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, specifically to a data mining method for online platform services based on big data. Background Technology

[0002] With the rapid development of the digital economy and online platforms, enterprise platforms generate massive amounts of multi-source heterogeneous service data every day, such as user transaction form data, behavior log data, and customer service text data. In the field of big data analysis and mining, traditional processing frameworks usually use a fixed extraction, transformation, and loading process and a static underlying database index structure to store the above data, and then hand it over to the upper-level machine learning algorithms for mining.

[0003] At the database layer, this traditional architecture typically uses static index structures such as B+ tree indexes or inverted indexes to support retrieval needs for structured and semi-structured data, while for unstructured text data, it uses an independent search engine component to build an inverted index, which is separate from the structured data storage.

[0004] The closest existing technical solution to this invention is a technical approach that combines a fixed extraction, transformation, and loading process with a static index structure and periodic batch mining model training. The processing generally includes: the data collection and cleaning module collects raw data from multiple business subsystems such as the transaction system, log system, and customer service system and performs basic cleaning; the extraction, transformation, and loading module writes the cleaned data into a relational database or a non-relational database according to predefined fixed transformation rules; the underlying database establishes a fixed index structure based on field types; this index structure is only partially maintained when data is added or deleted; full or incremental data is extracted from the database at fixed intervals; the model used by the upper-level mining algorithm is retrained or updated; and the trained mining model is deployed online for downstream business systems to call.

[0005] In the aforementioned existing technical solutions, the underlying index structure remains fixed once established, lacking the ability for semantic-level adaptive reorganization. Structured transaction data, semi-structured log data, and unstructured text data are stored in independent storage engines, with no underlying semantic relationship between them. Cross-modal data fusion relies on upper-layer applications to write manual rules for splicing and matching. Therefore, the existing technical solutions are insufficient in dealing with dynamic changes in data distribution, sparse node association completion, and semantic fusion of multi-source heterogeneous data. To address this, we propose a data mining method for online platform services based on big data. Summary of the Invention

[0006] The purpose of this invention is to provide a data mining method for online platform services based on big data.

[0007] To achieve the above objectives, this invention provides the following technical solution: a data mining method for online platform services based on big data. This method starts with the underlying storage and index structure of the database. By setting up a virtual association layer physically separated from the real index edges in the same dynamic hyperbolic graph index, it ensures that the two types of operations—local subgraph reconstruction driven by semantic distribution drift and the generation of virtual association edges for cold-start sparse nodes—do not interfere with each other and do not destroy the authenticity of the original topology. This constructs a unified dynamic hyperbolic graph index that combines dynamic perception, adaptive reconstruction, and native sparse association completion capabilities. This overcomes the technical bias of traditional solutions where multiple index maintenance mechanisms are coupled and prone to structural pollution. Specifically, this method extracts features from structured service data, semi-structured service data, and unstructured service data respectively, and maps the obtained feature vectors to a consistent Euclidean latent space using a linear projection matrix. The hyperbolic index mapping function is used to map the Euclidean embedding vector to the hyperbolic space to obtain the hyperbolic embedding coordinates of each service node. The hyperbolic distance between service nodes is calculated based on the hyperbolic embedding coordinates, and index edges with time decay weights are established to construct a dynamic hyperbolic graph index. The hyperbolic embedding distribution of service nodes is collected in real time in the streaming data writing channel. The degree of difference in semantic density probability distribution between sliding time windows is used to determine whether semantic distribution changes have occurred. If changes have occurred, subgraph splitting, weight decay, and edge reconstruction are performed on the affected local subgraphs, and the reconstruction results are incrementally written back to the index. For sparse service nodes with a small number of index edge connections, virtual implicit association edges are automatically generated using restricted random walk and graph attention mechanisms and written into a dedicated virtual association layer. When responding to mining requests, the real index edge features and virtual association layer completion features are fused to form a dense feature matrix and sent to the upper-layer mining algorithm.

[0008] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0009] 1. This invention achieves cross-modal semantic alignment at the underlying storage structure level by uniformly mapping structured, semi-structured and unstructured multi-source heterogeneous service data to the same hyperbolic space and constructing a dynamic hyperbolic graph index. This eliminates the data silo phenomenon in traditional solutions where structured, semi-structured and unstructured data are stored separately and separated from each other. Cross-modal semantic association can be completed without relying on manually predefined fusion rules.

[0010] 2. This invention embeds a semantic distribution change detection mechanism into the streaming data writing channel and triggers local adaptive index reconstruction, enabling the index structure itself to possess semantic distribution change perception and incremental evolution capabilities. This allows upper-layer mining algorithms to acquire data features reflecting the latest business status in a more timely manner. Since the local reconstruction process only targets data affected by semantic distribution changes... Processing is done on individual service nodes and their local neighborhoods, rather than on the entire system. Each service node is recalculated, and the time complexity of this reconstruction process is O(n log n). This is lower than the time complexity of the traditional full reconstruction scheme that recalculates all nodes. This can reduce the consumption of computing and storage resources while ensuring timely updates to the index structure;

[0011] 3. This invention automatically generates virtual implicit association edges for sparse service nodes by utilizing restricted random walks and graph attention mechanisms at the graph index layer, so that the upper-layer mining algorithm obtains a dense feature matrix after completion. This alleviates the problem of decreased mining effect caused by the lack of historical interaction data in new user and new service scenarios from the data structure level, and improves the accuracy of mining results in such scenarios. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall method flow in an embodiment of the present invention;

[0013] Figure 2 This is a schematic diagram of the drift detection and local reconstruction sub-process in an embodiment of the present invention;

[0014] Figure 3 This is a schematic diagram of the cold start sparse completion subprocess in an embodiment of the present invention. Detailed Implementation

[0015] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0016] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] In one specific embodiment, the big data-based online platform service data mining method provided by the present invention is applied to the backend data processing system of an online service platform. This system is logically divided from bottom to top into a data source layer, a semantic alignment and embedding layer, a dynamic hyperbolic graph index storage layer, an upper-layer mining algorithm layer, and an application display layer. The data source layer includes three types of data sources: structured transaction form data, semi-structured behavior log data, and unstructured customer service text data. The semantic alignment and embedding layer extracts features from these three types of heterogeneous data and projects them into a unified latent space. The dynamic hyperbolic graph index storage layer serves as the core of the present invention. The core storage structure comprises a hyperbolic graph index body, a drift detection probe, a local subgraph reconstruction module, and a cold-start sparse completion module. These four components are interconnected via data flow and control flow. Specifically, the drift detection probe triggers the local subgraph reconstruction module to perform actions after detecting a change in semantic distribution. The cold-start sparse completion module has a bidirectional read-write relationship with the hyperbolic graph index body. The upper mining algorithm layer includes specific mining task modules such as classification, clustering, association rule mining, and recommendation. Dense feature matrices are obtained from the dynamic hyperbolic graph index storage layer through a unified graph index query interface. The application presentation layer distributes the mining results to the business system.

[0018] I. Heterogeneous Data Semantic Alignment and Index Building Module:

[0019] This section implements the mapping of structured transaction data, semi-structured log data and unstructured text data to a unified high-dimensional hyperbolic space, and constructs a dynamic hyperbolic graph index with time decay weights, specifically including the following steps;

[0020] Step 1, Data Access and Preprocessing: Access structured transaction form data, semi-structured behavior log data, and unstructured customer service text data respectively. Perform basic preprocessing such as noise reduction and missing value imputation on the above three types of data to obtain preprocessed structured data, semi-structured data, and unstructured data.

[0021] Step 2, Modal Feature Extraction: Perform one-hot encoding or numerical normalization on the preprocessed structured fields to obtain structured feature vectors. Extract behavioral sequence statistical features from the preprocessed semi-structured behavior logs according to a preset time window to obtain semi-structured feature vectors. Use a pre-trained semantic coding model to extract semantic vectors from the preprocessed unstructured customer service text to obtain unstructured feature vectors.

[0022] Step 3, cross-modal feature projection: The structured feature vectors, semi-structured feature vectors and unstructured feature vectors mentioned above are mapped to a unified Euclidean latent space with consistent dimensions through a set of learnable linear projection matrices, to obtain the Euclidean embedding vectors corresponding to each service node. The service nodes are used to represent business entities such as users, services or transactions in the online platform.

[0023] Step 4, Hyperbolic Space Mapping: Using the hyperbolic exponential mapping function, the Euclidean embedding vector in the Euclidean latent space is mapped to the Poincaré sphere hyperbolic space to obtain the hyperbolic embedding coordinates of each service node. Since the volume of hyperbolic space grows exponentially with the radius, it is naturally suitable for representing data structures with hierarchical, tree-like or power-law distribution characteristics with lower-dimensional embedding. Therefore, compared with Euclidean space, using hyperbolic space for embedding can reduce the distortion of the hierarchical distance relationship of service nodes under the same embedding dimension.

[0024] Step 5, Node Similarity Calculation and Edge Connection Construction: Calculate the hyperbolic distance between any two service nodes based on the hyperbolic distance. Establish indexed edges between service node pairs whose hyperbolic distance is less than a preset threshold to form the initial graph topology. The hyperbolic distance between nodes is calculated using the distance formula under the Poincaré sphere model.

[0025] ;

[0026] in, Indicates service node With service nodes Distance in hyperbolic space , Let represent the hyperbolic embedding coordinate vectors of the two service nodes in the Poincaré sphere model. express and The magnitude of the vector difference in the Euclidean sense , Representing vectors respectively , The square of the modulus, Represents the inverse hyperbolic cosine function;

[0027] Step Six, Edge Weight Initialization: Assign an initial value to the time decay weight of each index edge established in Step Five. This initial value is determined based on the hyperbolic distance between service nodes and the data generation time. The time decay weight of the index edge is calculated using the following formula:

[0028] ;

[0029] in, Indicates time service node With service nodes The dynamic weights of the indexed edges between them. This indicates the base weight at the time the index edge was created, determined based on the hyperbolic distance. This represents a pre-set time-related decay coefficient, used to control the rate at which historical correlations weaken over time. Indicates the current query or update time. This indicates the timestamp of the last time the indexed edge was accessed or updated;

[0030] Step 7, Index Persistence: The completed hyperbolic graph index, including service node coordinates, index edge connection relationships, and index edge weights, is persistently stored in the underlying graph index storage engine to form an initial dynamic hyperbolic graph index for subsequent semantic distribution change detection, local reconstruction, and sparse completion processes for reading and writing.

[0031] II. Drift Detection and Local Reconstruction Module:

[0032] This section embeds a drift detection probe in the streaming data writing channel to calculate the degree of difference in the semantic density distribution of service nodes within the sliding time window in real time. When a significant change in semantic distribution is detected, a local reconstruction of the dynamic hyperbolic graph index is triggered. The specific steps include the following:

[0033] Step 1, Sliding window acquisition: Embed a drift detection probe in the streaming data writing channel and acquire the hyperbolic embedding distribution of newly written service nodes in real time within the current sliding time window according to the preset window length;

[0034] Step 2, semantic density distribution estimation: The hyperbolic embedding coordinates of the service nodes within the sliding time window are transformed into a discrete semantic density probability distribution through kernel density estimation or histogram statistics.

[0035] Step 3, Calculate the distribution difference metric: Calculate the distribution difference metric between the semantic density probability distribution of the current sliding time window and the semantic density probability distribution of the previous sliding time window. In a preferred embodiment, the distribution difference metric is... The divergence value is calculated using the following formula:

[0036] ;

[0037] in, Indicates the current sliding time window time. The semantic density probability distribution relative to the previous sliding time window time The degree of difference in semantic density probability distributions This represents the total number of discrete intervals divided by the semantic density probability distribution. Indicates the number of times within the current sliding time window. The probability values ​​corresponding to each interval Indicates the first time within the previous sliding time window The probability values ​​corresponding to each interval;

[0038] Step 4, Dynamic Threshold Maintenance: Based on the moving mean and moving standard deviation of the historical distribution difference measurement value sequence, the adaptive threshold used to determine whether a semantic distribution change has occurred is dynamically updated, specifically calculated using the following formula:

[0039] ;

[0040] in, express An adaptive threshold is used to determine whether a local reconstruction is triggered. This represents the moving average of a sequence of historical distribution variance measures. The moving standard deviation of a series of historical distribution variance measures. This represents a pre-set sensitivity coefficient, used to adjust the sensitivity of the judgment according to the business scenario;

[0041] Step 5, Drift Determination and Localization: Compare the distribution difference metric obtained in Step 3 with the dynamic determination threshold obtained in Step 4. If the distribution difference metric does not exceed the dynamic determination threshold, return to Step 1 to continue monitoring the next sliding time window. If the distribution difference metric exceeds the dynamic determination threshold, determine that the service nodes in the current sliding time window have undergone semantic distribution changes, and identify the service nodes whose semantic density changes exceed the preset amplitude threshold and their k-hop neighbor service node sets as the affected local subgraph.

[0042] Step 6, Subgraph Splitting: Perform a subgraph splitting operation on the local subgraph determined in Step 5, and remove service nodes whose semantic distribution differences exceed the preset tolerance range from the original local subgraph structure to obtain a set of isolated service nodes to be reconstructed.

[0043] Step 7, Time-based weight decay: For the index edges associated with the stripped service nodes within the local subgraph, recalculate and update their weights according to the aforementioned index edge time-based weight decay calculation method.

[0044] Step 8, Dynamic Edge Reconstruction: Based on the updated hyperbolic embedding coordinates of the stripped service node, recalculate its hyperbolic distance with surrounding service nodes, add index edges that meet the preset similarity conditions, delete invalid index edges whose hyperbolic distance exceeds the preset threshold, and complete the reconstruction of the local graph topology.

[0045] Step 9, Incremental Write-back: Write the reconstructed local subgraph structure, including the new service node coordinates, index edge connection relationships, and index edge weights, back to the underlying dynamic hyperbolic graph index storage engine to replace the original local structure, complete a closed-loop incremental update, and return to Step 1 to continue monitoring the subsequent data flow.

[0046] Since steps five through nine only apply to those affected by changes in semantic distribution. Processing is done on individual service nodes and their local neighborhoods, rather than on the entire system. Each service node is recalculated, therefore the time complexity of this partial reconstruction process is O(n log n). This is lower than the time complexity of the traditional full reconstruction scheme that recalculates all nodes. .

[0047] III. Cold Start Sparse Completion Module:

[0048] This section addresses sparse service nodes corresponding to newly launched services or new users. Before the data retrieval is sent to the mining algorithm, virtual implicit association edges are generated for them at the graph index layer using restricted random walk and graph attention mechanisms. The specific steps include the following:

[0049] Step 1, Sparse Node Identification: Traverse the dynamic hyperbolic graph index, count the number of index edge connections for each service node, and mark service nodes with an index edge connection number lower than the preset sparse threshold as sparse service nodes to be completed.

[0050] Step 2, Candidate Related Node Sampling: Starting with each sparse service node to be completed, based on the restricted random walk transition probability, perform multi-step walk sampling with a limited number of steps and a limited range on the dynamic hyperbolic graph index to obtain the set of candidate related service nodes corresponding to the sparse service node. The restricted random walk transition probability is calculated using the following formula:

[0051] ;

[0052] in, This indicates that during a restricted random walk, the process starts from the current service node. Transfer to neighbor service node The probability, Indicates service node With service nodes The index edge between them at time Dynamic weights, Indicates service node and The hyperbolic distance between them This represents a pre-defined distance attenuation coefficient, used to control the degree to which hyperbolic distance affects the transition probability. Indicates service node The set of neighbor service nodes in a dynamic hyperbolic graph index This represents any candidate neighbor service node in the set of neighbor service nodes;

[0053] Step 3, Feature Extraction: Extract the original feature embedding vectors of each service node in the dynamic hyperbolic graph index from the sparse service node to be completed and its candidate associated service node set.

[0054] Step 4, Attention Weight Calculation: Using the graph attention mechanism, calculate the attention weight coefficient between the sparse service node to be completed and each service node in the candidate associated service node set. This attention weight coefficient represents the contribution of the candidate associated service node to the semantic completion of the sparse service node, and is calculated using the following formula:

[0055] ;

[0056] in, This represents sparse service nodes that need to be completed. Related service nodes with candidates Attention weight coefficients for generating virtual implicit association edges between them. , Representing service nodes With service nodes The original feature embedding vector, This represents the learnable linear transformation weight matrix. Represents a learnable attention weight vector, with the symbol... This represents a vector concatenation operation. This represents the linear rectified activation function with leakage correction. Represented as a service node The set of candidate associated service nodes obtained by sampling Represents any service node in the set of candidate associated service nodes;

[0057] Step 5, Virtual Edge Generation and Feature Completion: Based on the attention weight coefficients obtained in Step 4, the feature vectors of each service node in the candidate associated service node set are weighted and aggregated to generate virtual implicit associated edges corresponding to the sparse service nodes to be completed and the completed dense feature vectors, specifically calculated using the following formula:

[0058] ;

[0059] in, Indicates service node The dense feature vector obtained after virtual implicit correlation edge completion. Indicates service node The original sparse feature vector, , , The meaning is the same as the previous formula;

[0060] Step 6, Virtual Layer Writing: The virtual implicit association edges generated in Step 5 and the completed dense feature vectors are written into the virtual association layer specially set up in the dynamic hyperbolic graph index, and are distinguished from the real index edges by a specific label, so that the virtual association layer is stored separately from the graph topology structure composed of real index edges, avoiding the virtual implicit association edges from affecting the authenticity of the original topology structure.

[0061] Step 7, Feature Distribution: When the upper-layer mining algorithm initiates a retrieval request, the features corresponding to the real index edges and the complete features in the virtual association layer are read simultaneously through the graph index query interface. The two are then merged to form a dense feature matrix, which is then distributed to the upper-layer mining algorithm for data mining processing.

[0062] As a specific example, suppose that the number of index edge connections of a sparse service node corresponding to a newly launched service on an online service platform is lower than a preset sparsity threshold in the dynamic hyperbolic graph index. This service node is marked as a sparse service node to be completed. Starting from this sparse service node, a multi-step walk sampling with a preset number of steps is performed on the dynamic hyperbolic graph index according to the restricted random walk transition probability. This results in a set of candidate associated service nodes consisting of several candidate associated service nodes. For each service node in this set of candidate associated service nodes, the corresponding attention weight coefficient is calculated using a graph attention mechanism. The feature vector of the candidate associated service node is then weighted and aggregated based on the obtained attention weight coefficient to obtain the dense feature vector of the sparse service node after completion. When the upper-layer recommendation mining algorithm initiates a retrieval request for the newly launched service, the graph index query interface simultaneously reads and merges the real index edge features corresponding to the service node with the completed features in the virtual association layer. This allows the recommendation mining algorithm to obtain a relatively complete feature representation even when the service lacks historical interaction data, thereby alleviating the problem of decreased mining effect caused by the lack of historical interaction data.

[0063] In another embodiment of the present invention, hyperbolic space embedding and hyperbolic distance calculation can also be implemented using a hyperbolic surface model, while the other semantic distribution change detection, local reconstruction and sparse completion mechanisms remain unchanged, thus achieving the purpose of the present invention. In yet another embodiment of the present invention, the distribution difference metric used to determine the degree of semantic distribution difference can also be replaced by other distribution difference metrics instead of KL divergence, and combined with the corresponding dynamic threshold mechanism to trigger local subgraph reconstruction, thus achieving the purpose of drift-driven local adaptive reconstruction.

[0064] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a computer program controlling related hardware devices. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement some or all of the steps of the methods described in the foregoing embodiments. Correspondingly, the functional modules that implement the above methods can also be integrated into a big data-based online platform service data mining system. This system includes a heterogeneous data semantic alignment and index construction module, a drift detection and local reconstruction module, and a cold start sparse completion module. The three modules share the same dynamic hyperbolic graph index data structure and work together to provide real-time, dense, and semantically aligned data features for upper-level mining algorithms.

[0065] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A data mining method for online platform services based on big data, characterized in that, include: Feature extraction is performed on structured service data, semi-structured service data, and unstructured service data respectively to obtain structured feature vectors, semi-structured feature vectors, and unstructured feature vectors. The structured feature vectors, semi-structured feature vectors, and unstructured feature vectors are then mapped to a Euclidean latent space of the same dimension using a linear projection matrix to obtain the Euclidean embedding vector of each service node. The Euclidean embedding vector is mapped to hyperbolic space using the hyperbolic exponential mapping function to obtain the hyperbolic embedding coordinates of each service node; The hyperbolic distance between any two service nodes is calculated based on the hyperbolic embedded coordinates. Index edges are established between service nodes whose hyperbolic distance is less than a preset threshold, and each index edge is assigned an initial value for the time decay weight to construct a dynamic hyperbolic graph index. In the streaming data writing channel, the hyperbolic embedding distribution of newly written service nodes within the sliding time window is collected in real time. The semantic density probability distribution is determined based on the hyperbolic embedding distribution. The distribution difference metric between the semantic density probability distribution of the current sliding time window and the semantic density probability distribution of the previous sliding time window is calculated. Based on the historical distribution difference metric sequence, a dynamic judgment threshold for determining whether the semantic distribution has changed is determined. When the distribution difference metric exceeds the dynamic determination threshold, the service node whose semantic density change exceeds the preset magnitude threshold and its k-hop neighbor service nodes are identified as the affected local subgraph. The local subgraph is split into subgraphs, and the service nodes whose semantic density distribution difference exceeds the preset tolerance range are separated from the local subgraph to obtain the service nodes to be reconstructed. The hyperbolic distance between the service nodes to be reconstructed and the surrounding service nodes is recalculated based on the hyperbolic embedding coordinates of the service nodes to be reconstructed. Index edges that meet the preset similarity conditions are added and index edges whose hyperbolic distance exceeds the preset threshold are deleted. The reconstructed local subgraph structure is written back to the dynamic hyperbolic graph index. The number of index edge connections for each service node in the dynamic hyperbolic graph index is counted. Service nodes with an index edge connection number lower than a preset sparsity threshold are marked as sparse service nodes to be completed. Starting from the sparse service nodes to be completed, multi-step walk sampling with a limited number of steps and a limited range is performed on the dynamic hyperbolic graph index according to the restricted random walk transition probability to obtain a set of candidate associated service nodes. The attention weight coefficient between the sparse service node to be completed and each service node in the set of candidate associated service nodes is calculated using a graph attention mechanism. The feature vectors of each service node in the set of candidate associated service nodes are weighted and aggregated according to the attention weight coefficient to generate the virtual implicit association edge and the completed dense feature vector corresponding to the sparse service node to be completed. The virtual implicit association edge and the completed dense feature vector are written into the virtual association layer in the dynamic hyperbolic graph index. In response to the mining request, the features corresponding to the real index edges are read from the dynamic hyperbolic graph index and the completion features in the virtual association layer are fused to obtain a dense feature matrix, and the dense feature matrix is ​​sent to the mining algorithm for data mining processing.

2. The data mining method for online platform services based on big data according to claim 1, characterized in that: The hyperbolic embedding coordinates are located in the Poincaré sphere hyperbolic space, between any two service nodes. and hyperbolic distance between Calculate using the following formula: ； in, Indicates service node With service nodes The magnitude of the vector difference in the Euclidean sense. , Representing vectors respectively , The square of the modulus, This represents the inverse hyperbolic cosine function.

3. The data mining method for online platform services based on big data according to claim 1, characterized in that: service node With service nodes The index edge between them at time Dynamic weights Calculate using the following formula: ； in, The base weight at the time the indexed edge is established is represented by the service node. With service nodes The hyperbolic distance between them and the time of data generation are determined. This represents the pre-set aging decay coefficient. This indicates the timestamp of the last time the indexed edge was accessed or updated.

4. The data mining method for online platform services based on big data according to claim 1, characterized in that: The distribution difference measure is Divergence value, the semantic density probability distribution of the current sliding time window. semantic density probability distribution of the previous sliding time window Between divergence value Calculate using the following formula: ； in, This represents the total number of discrete intervals divided by the semantic density probability distribution. Indicates the number of times within the current sliding time window. The probability values ​​corresponding to each interval Indicates the first time within the previous sliding time window The probability values ​​corresponding to each interval.

5. The data mining method for online platform services based on big data according to claim 4, characterized in that: The dynamic determination threshold Based on history Moving average of divergence value sequence With sliding standard deviation Determine by the following formula: ； in, This represents the preset sensitivity coefficient.

6. The data mining method for online platform services based on big data according to claim 1, characterized in that: The restricted random walk transition probability, i.e., from the current service node Transfer to neighbor service node probability Calculate using the following formula: ； in, Indicates service node With service nodes The index edge between them at time Dynamic weights, Indicates service node and The hyperbolic distance between them This represents the preset distance attenuation coefficient. Indicates service node The set of neighbor service nodes in the dynamic hyperbolic graph index This refers to any candidate neighbor service node in the set of neighbor service nodes.

7. The data mining method for online platform services based on big data according to claim 1, characterized in that: The attention weight coefficient, i.e., the sparse service node to be completed. Related service nodes with candidates Attention weight coefficients between Calculate using the following formula: ； in, , Representing service nodes With service nodes The original feature embedding vector, This represents the learnable linear transformation weight matrix. Represents a learnable attention weight vector, with the symbol... This represents a vector concatenation operation. This represents the linear rectified activation function with leakage correction. Represented as a service node The set of candidate associated service nodes obtained by sampling This refers to any service node in the set of candidate associated service nodes.

8. The data mining method for online platform services based on big data according to claim 1, characterized in that: The sparse service nodes to be completed Completed dense feature vector Calculate using the following formula: ； in, Indicates service node The original sparse feature vector.

9. A data mining system for online platform services based on big data, characterized in that, It includes a heterogeneous data semantic alignment and index building module, a drift detection and local reconstruction module, and a cold start sparse completion module, among which: The heterogeneous data semantic alignment and index construction module is used to extract features from structured service data, semi-structured service data and unstructured service data respectively and project them to a unified Euclidean latent space, and then map them to hyperbolic space to obtain the hyperbolic embedding coordinates of each service node. Based on the hyperbolic embedding coordinates, a dynamic hyperbolic graph index with time decay weight is constructed. The drift detection and local reconstruction module is used to monitor the distribution difference metric of semantic density probability distribution between sliding time windows in real time in the streaming data writing channel. When the distribution difference metric exceeds the dynamic judgment threshold, the module performs subgraph splitting, weight decay and edge reconstruction on the affected local subgraph and writes the reconstruction result back to the dynamic hyperbolic graph index. The cold start sparse completion module is used to identify sparse service nodes in the dynamic hyperbolic graph index whose number of index edge connections is lower than a preset sparse threshold. It uses restricted random walk and graph attention mechanisms to generate virtual implicit association edges and completed dense feature vectors for the sparse service nodes and writes them into a dedicated virtual association layer in the dynamic hyperbolic graph index. The heterogeneous data semantic alignment and index construction module, the drift detection and local reconstruction module, and the cold start sparse completion module share the same dynamic hyperbolic graph index for data reading and writing, and provide the upper-layer mining algorithm with a dense feature matrix obtained by fusing real index edge features and virtual association layer completion features.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the data mining method for online platform services based on big data as described in any one of claims 1 to 8.

Images