Customer portrait key data mining method and system based on spatiotemporal big data

By combining Transformer feature extraction, dynamic spectral clustering and graph neural networks, multiple challenges in multi-source heterogeneous spatiotemporal data processing are solved, high-quality spatiotemporal data processing and personalized preference modeling are achieved, and computing is improved Efficiency and model adaptability.

CN118797542BActive Publication Date: 2024-11-29NANJING EFL E-COMMERCE CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411290436.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-14
Publication Date
2024-11-29
Estimated Expiration
2044-09-14

Smart Images

  • Figure CN118797542B_ABST
    Figure CN118797542B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for mining key data of customer portraits based on spatiotemporal big data, including the following steps: obtaining multi-source heterogeneous spatiotemporal big data, extracting features, dynamically classifying and standardizing, obtaining high-quality spatiotemporal data sets; extracting spatiotemporal feature matrices based on the spatiotemporal data sets, building customer behavior models and personalized spatiotemporal preference models; integrating customer features, building customer portrait sets and dynamic customer groups; generating recommendation models and strategy sets based on dynamic customer portraits and group analysis results. The method can effectively process large-scale heterogeneous spatiotemporal data, can more accurately capture user behavior patterns in sparse and noisy data, and improve the accuracy and timeliness of customer portraits.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to big data processing technology, and in particular to a method and system for mining key data of customer portraits based on spatiotemporal big data. Background Art

[0002] With the advent of the big data era, customer profiling technology based on spatiotemporal big data has shown great potential in areas such as business intelligence, personalized recommendations and risk management. It can extract valuable information from massive spatiotemporal data, build accurate customer portraits, and provide strong support for corporate decision-making.

[0003] At present, the research on customer portraits of spatiotemporal big data has made significant progress. Traditional methods mainly rely on statistical analysis and machine learning algorithms, such as cluster analysis and association rule mining. With the development of deep learning technology, researchers have begun to apply models such as convolutional neural networks (CNN) and recurrent neural networks (RNN) to process spatiotemporal sequence data. In recent years, graph neural networks (GNNs) have shown advantages in processing complex relationship networks and have been widely used in customer social network analysis.

[0004] However, there are still some technical difficulties in current research. For example, the quality of multi-source heterogeneous spatiotemporal data varies greatly, and how to effectively clean and fuse data remains a challenge. Thirdly, the existing dynamic spectral clustering algorithm has low computational efficiency when processing large-scale time-varying graph structures, and it is difficult to meet the needs of real-time processing. In addition, although the anomaly detection method based on Wasserstein distance has theoretical advantages, it has high computational complexity in high-dimensional space. How to improve efficiency while ensuring accuracy is an urgent problem to be solved. Finally, the existing semantic trajectory extraction method is not robust enough for sparse and noisy data, and it is difficult to accurately capture the user's true behavior patterns. In terms of building a personalized spatiotemporal preference model, how to effectively balance the dynamic changes of long-term preferences and short-term interests remains a difficult problem. Therefore, research innovation is needed. Summary of the invention

[0005] The purpose of the invention is to provide a method and system for mining key data of customer portraits based on spatiotemporal big data to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution is a key data mining method for customer portraits based on spatiotemporal big data, including the following steps:

[0007] Step S1, obtain multi-source heterogeneous spatiotemporal big data, and obtain high-quality spatiotemporal data sets through Transformer feature extraction, dynamic classification, adaptive normalization, graph theory quality assessment, Wasserstein distance anomaly detection and graph neural network tensor completion;

[0008] Step S2: Based on the spatiotemporal data set, the spatiotemporal feature matrix is ​​extracted through multi-scale singular spectrum analysis and topological analysis; the customer behavior model is constructed using the graph attention diffusion network to obtain the behavior pattern and time-varying community structure; the trajectory segmentation and hierarchical topic model of the attention mechanism are applied to extract the semantic trajectory and spatiotemporal pattern; a personalized spatiotemporal preference model is constructed by combining the multi-dimensional preference tensor and the graph neural network; finally, the spatiotemporal graph convolutional neural network is designed to generate the anomaly detection model and the abnormal behavior set, and the output data set is constructed;

[0009] The output data set includes spatiotemporal feature matrix F, customer behavior pattern P, time-varying community structure C(t), semantic trajectory ST, spatiotemporal pattern M, personalized spatiotemporal preference model PM, anomaly detection model AD and abnormal behavior set AB;

[0010] Step S3: Based on the output data set, the customer features are integrated to obtain a vector set through the meta-path-guided graph attention mechanism and hierarchical contrastive learning; based on the vector set, a customer portrait set is constructed using a recursive neural topic model and a dynamic factor graph; spectral graph theory and dynamic manifold learning are used to obtain dynamic customer groups and evolution patterns, and dynamic customer portraits and group analysis results are output.

[0011] Step S4: Based on the dynamic customer portrait and group analysis results, the constructed dynamic heterogeneous graph neural network and multi-task reinforcement learning framework are used to generate a recommendation model and strategy set.

[0012] According to another aspect of the present application, a customer portrait key data mining system based on spatiotemporal big data is also provided, comprising:

[0013] at least one processor; and,

[0014] a memory communicatively connected to at least one of the processors; wherein,

[0015] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the customer portrait key data mining method based on spatiotemporal big data described in any of the above technical solutions.

[0016] Beneficial effects: By introducing Transformer feature extraction and graph theory quality assessment, the processing quality of multi-source heterogeneous spatiotemporal data is effectively improved, and the problem of data cleaning and fusion is solved. The dynamic spectral clustering algorithm combined with the incremental update strategy is adopted to improve the processing efficiency of large-scale time-varying graph structures and meet the needs of real-time analysis. The Wasserstein distance is innovatively combined with the local outlier factor (LOF) algorithm to significantly reduce the computational complexity while ensuring the accuracy of anomaly detection in high-dimensional space. By introducing the trajectory segmentation algorithm and hierarchical topic model with attention mechanism, the robustness of semantic trajectory extraction is improved, and user behavior patterns in sparse and noisy data can be captured more accurately. A personalized spatiotemporal preference model based on multi-task learning is designed, which effectively balances the dynamic changes of long-term preferences and short-term interests and improves the adaptability of the model. By constructing a dynamic heterogeneous graph neural network and introducing a semantic aggregation layer guided by meta-paths, the scalability and computational efficiency of the model in processing large-scale dynamic graph structures are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the present invention.

[0018] Figure 2 It is a flow chart of step S1 of the present invention.

[0019] Figure 3 It is a flow chart of step S2 of the present invention.

[0020] Figure 4 It is a flow chart of step S3 of the present invention.

[0021] Figure 5 It is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0022] like Figure 1 As shown in the figure, the key data mining method of customer portrait based on spatiotemporal big data includes the following steps:

[0023] Step S1, obtain multi-source heterogeneous spatiotemporal big data, and obtain high-quality spatiotemporal data sets through Transformer feature extraction, dynamic classification, adaptive normalization, graph theory quality assessment, Wasserstein distance anomaly detection and graph neural network tensor completion;

[0024] Step S2: Based on the spatiotemporal data set, the spatiotemporal feature matrix is ​​extracted through multi-scale singular spectrum analysis and topological analysis; the customer behavior model is constructed using the graph attention diffusion network to obtain the behavior pattern and time-varying community structure; the trajectory segmentation and hierarchical topic model of the attention mechanism are applied to extract the semantic trajectory and spatiotemporal pattern; a personalized spatiotemporal preference model is constructed by combining the multi-dimensional preference tensor and the graph neural network; finally, the spatiotemporal graph convolutional neural network is designed to generate the anomaly detection model and the abnormal behavior set, and the output data set is constructed;

[0025] The output data set includes spatiotemporal feature matrix F, customer behavior pattern P, time-varying community structure C(t), semantic trajectory ST, spatiotemporal pattern M, personalized spatiotemporal preference model PM, anomaly detection model AD and abnormal behavior set AB;

[0026] Step S3: Based on the output data set, the customer features are integrated to obtain a vector set through the meta-path-guided graph attention mechanism and hierarchical contrastive learning; based on the vector set, a customer portrait set is constructed using a recursive neural topic model and a dynamic factor graph; spectral graph theory and dynamic manifold learning are used to obtain dynamic customer groups and evolution patterns, and dynamic customer portraits and group analysis results are output;

[0027] Step S4: Based on the dynamic customer portrait and group analysis results, the constructed dynamic heterogeneous graph neural network and multi-task reinforcement learning framework are used to generate a recommendation model and strategy set.

[0028] It should be noted that, in some embodiments or scenarios, step S4 may not be included, and the dynamic customer portrait and group analysis results can be pushed to the customer after they are obtained.

[0029] In this embodiment, a solid foundation is laid for subsequent analysis through the fusion and quality improvement of multi-source heterogeneous data. Secondly, technologies such as multi-scale feature extraction and dynamic community detection are used to deeply explore customer behavior patterns and preferences. Thirdly, advanced algorithms such as recursive neural topic models and dynamic factor graphs are used to construct dynamically evolving customer portraits. Finally, based on dynamic heterogeneous graph neural networks and reinforcement learning frameworks, the generation of personalized recommendation strategies is realized. This series of steps not only solves the efficiency and accuracy problems of traditional methods in processing large-scale spatiotemporal data, but also overcomes the limitation that static portraits cannot reflect the dynamic changes of customers. Through the innovative application of deep learning and graph structure algorithms, the solution can capture subtle changes and long-term trends in customer behavior, providing strong support for precision marketing and personalized services. In addition, the anomaly detection and adaptive optimization mechanisms in the solution greatly improve the robustness and real-time performance of the model.

[0030] According to one aspect of the present application, step S1 specifically comprises:

[0031] Step S11, obtaining different types of original spatiotemporal big data from multiple heterogeneous data sources, and for each data source, using a pre-trained Transformer model to perform feature extraction, including: serializing the spatiotemporal data into a format suitable for Transformer processing to obtain an input sequence; inputting the input sequence into the Transformer model, and processing it through a multi-layer self-attention mechanism and a feedforward neural network to obtain an enhanced feature sequence; aggregating the enhanced feature sequence to obtain an overall feature representation of the data source, and obtaining and outputting an enhanced feature data set, including original data and its corresponding enhanced features, and the original data includes original trajectory data;

[0032] Step S12, obtaining an enhanced feature data set, first extracting feature vectors for each enhanced feature data set, constructing a dynamic similarity matrix, calculating its Laplace matrix, solving the generalized eigenvalue problem to obtain time-varying feature vectors, clustering based on the time-varying feature vectors, completing the process of classifying the feature vectors using a dynamic spectral clustering algorithm, and obtaining a classified data set; the time-varying feature vectors include time granularity, spatial precision, and data type;

[0033] Step S13, obtaining the classified data set, performing time standardization, detecting the timestamp format and uniformly converting it into a Unix timestamp, applying wavelet transform to decompose the time series, extracting multi-scale time features, resampling the time series using an adaptive time window method, and then performing spatial standardization, converting all spatial coordinates into a unified geographic coordinate system, for non-point data, applying a curvature-based shape descriptor to extract feature points, and using the Riemann manifold learning method to perform nonlinear normalization on the spatial data, and then completing the spatiotemporal standardization of each classified data set to obtain a standardized spatiotemporal data set;

[0034] Step S14: for the standardized spatiotemporal data set, call to build a spatiotemporal hypergraph, apply the hypergraph theory, calculate the hypergraph Laplace eigenvalue of each vertex, call the threshold of the quality score, and mark the data points below the threshold as the data set to be repaired; obtain the data set with quality score and the data set to be repaired, and perform data quality assessment on each data set;

[0035] Among them, the vertices of the spatiotemporal hypergraph represent data points, and the hyperedges represent spatiotemporal relationships; the quality score is the result of the processing of the eigenvalue by the nonlinear mapping function based on deep learning;

[0036] Step S15, obtaining the quality-scored data set, constructing a spatiotemporal feature vector, applying an anomaly detection algorithm based on Wasserstein distance, and estimating the empirical distribution of the data; for each data point, calculating its Wasserstein distance with the empirical distribution, and dynamically setting the anomaly threshold using extreme value theory; marking the data point with a distance greater than the threshold as an outlier and adding it to the outlier set; obtaining the cleaned data set and the outlier set;

[0037] Among them, the spatiotemporal feature vector includes time feature, space feature and attribute feature;

[0038] Step S16, obtain the cleaned data set, the data set to be repaired and the set of outliers, construct a spatiotemporal-attribute fourth-order tensor, introduce an additional modal dimension to represent data reliability, adopt a tensor completion algorithm based on graph neural network, represent the tensor as a heterogeneous graph, construct a graph attention network layer to learn node representation, and use the graph autoencoder architecture to reconstruct missing values; for areas with insufficient spatiotemporal density, apply a conditional generation model based on Gaussian Copula to generate new data points; complete data completion and enhancement to obtain a completed high-quality spatiotemporal data set.

[0039] In this embodiment, first, the pre-trained Transformer model is used for feature extraction, which not only improves the representation ability of complex spatiotemporal data, but also captures the long-distance dependency between data through the self-attention mechanism. The application of the dynamic spectral clustering algorithm solves the problem that traditional clustering methods are difficult to handle high-dimensional spatiotemporal data, and realizes the adaptive classification of data. The wavelet transform and Riemann manifold learning method introduced in the spatiotemporal standardization process effectively handle the multi-scale characteristics of time series and the nonlinear characteristics of spatial data, and improve the comparability and consistency of data. The quality assessment link innovatively uses spatiotemporal hypergraphs and spectral theory, which can not only comprehensively evaluate data quality, but also identify and locate problem data. The introduction of the Wasserstein distance anomaly detection algorithm improves the sensitivity to minor anomalies, and at the same time dynamically sets the threshold through extreme value theory, which enhances the adaptability of detection. Finally, the tensor completion algorithm based on the graph neural network solves the problem that traditional methods are difficult to handle high-dimensional sparse data, and realizes the accurate repair of missing data. The comprehensive application of this series of technologies not only greatly improves the data quality and integrity, but also lays a solid foundation for subsequent in-depth analysis. The innovation of this step lies in the organic combination of deep learning, graph theory and statistical methods, which realizes the comprehensive processing of complex spatiotemporal big data and overcomes the limitations of traditional methods in processing heterogeneous, high-dimensional and noisy data.

[0040] According to one aspect of the present application, step S2 is specifically:

[0041] Step S21, obtaining a high-quality spatiotemporal data set, applying an adaptive multi-scale singular spectrum analysis method and a dynamic manifold learning method to sequentially perform feature extraction on the spatiotemporal sequence of each customer, obtaining and combining a time feature vector and a space feature vector to form a spatiotemporal feature matrix;

[0042] Step S22, obtain the spatiotemporal feature matrix and construct a spatiotemporal behavior graph; apply the graph attention diffusion network to capture different types of spatiotemporal dependencies and perform multi-scale aggregation to learn customer representations; perform dynamic community detection on the learned customer representations, construct a time-varying modularity matrix and optimize it; use the spectral clustering method to cluster the optimized customer representations; finally output a set of customer behavior patterns and a time-varying community structure;

[0043] Step S23, obtaining the original trajectory data, applying the trajectory segmentation algorithm based on the attention mechanism to identify the semantic stay points and movement segments; applying the hierarchical topic model to the stay points to extract the multi-level semantic labels; using the temporal association rule mining algorithm to extract the frequent movement patterns for the movement segments; integrating the processed stay point and movement segment information to generate a semantic trajectory set; summarizing the analysis results of all trajectories to form a frequent spatiotemporal pattern set;

[0044] Step S24, obtain semantic trajectories and frequent spatiotemporal patterns, and construct a multi-dimensional preference tensor; convert the preference tensor into a heterogeneous graph, and apply a relationship-aware graph convolutional network to learn the embedding representation of entities and relationships; based on the learned embedding, calculate the user-location preference score, and estimate the time and context preference distribution; integrate this information to construct a multimodal preference vector, and generate a personalized spatiotemporal preference model for each user; finally output a set of personalized spatiotemporal preference models for customers;

[0045] Step S25, obtain customer behavior patterns and time-varying community structures, as well as personalized preference models; construct a dynamic spatiotemporal graph and apply the spatiotemporal attention mechanism and graph differential pooling to learn the hierarchical representation of user behavior; based on the hierarchical representation, train the spatiotemporal graph convolutional neural network as the basic model for anomaly detection; enhance the model's discrimination ability through comparative learning methods; perform anomaly judgment on newly observed user behaviors in combination with the learned behavior patterns and personalized preference models; and finally output the abnormal behavior detection model and the detected abnormal behavior set.

[0046] In this embodiment, first, the application of the adaptive multi-scale singular spectrum analysis method solves the problem that traditional time series analysis is difficult to capture multiple time scale features at the same time, and improves the recognition ability of complex time patterns such as periodicity, trend and mutation. The introduction of the dynamic manifold learning method overcomes the limitations of traditional spatial analysis methods in processing high-dimensional nonlinear data, and realizes the effective representation of complex spatial structures. The application of the graph attention diffusion network innovatively combines the graph structure with the attention mechanism, which can not only capture the complex dependencies of customer behavior, but also adaptively focus on the most relevant spatiotemporal features, greatly improving the accuracy of behavioral pattern recognition. In the process of semantic trajectory extraction, the combination of hierarchical topic model and temporal association rule mining realizes multi-level and multi-angle analysis of customer activity patterns, providing rich semantic information for in-depth understanding of customer behavior. The construction of the personalized spatiotemporal preference model innovatively integrates tensor decomposition, graph neural network and sequence modeling technology, which can not only capture the multi-dimensional characteristics of user preferences, but also effectively model the temporal dynamics and spatial correlation of preferences. Finally, the application of spatiotemporal graph convolutional neural networks solves the problem that traditional anomaly detection methods are difficult to consider time, space, and behavior characteristics at the same time, and achieves efficient recognition of complex abnormal patterns through end-to-end learning. The comprehensive application of this series of technologies has greatly improved the depth and breadth of customer behavior analysis.

[0047] According to one aspect of the present application, step S3 is specifically:

[0048] Step S31, obtain the customer-related data output by step S2, and construct a heterogeneous information network; construct and apply a set of key meta-paths, and use a graph attention mechanism to calculate the semantic-specific embedding of nodes; design an adaptive attention mechanism, integrate multi-path information to obtain a comprehensive embedding of nodes; further optimize node representation through hierarchical contrast learning tasks; and finally output a fused customer feature vector set; customer-related data include a spatiotemporal feature matrix, a set of customer behavior patterns and a time-varying community structure, a set of semantic trajectories and a set of frequent spatiotemporal patterns, a set of customer personalized spatiotemporal preference models, an abnormal behavior detection model, and a set of detected abnormal behaviors.

[0049] Step S32, obtain a fused feature vector set, apply a recursive neural topic model to capture the evolution of the topic over time; construct a dynamic factor graph model to capture the time-varying dependencies between features; integrate the outputs of the two models to generate a multi-granular customer portrait, including the topic evolution trajectory at the macro level, the time-varying distribution of key features at the meso level, and the instantaneous feature vector at the micro level; output a dynamic customer portrait set;

[0050] Step S33, obtain a dynamic customer portrait set and construct a time-varying customer similarity graph; apply spectral graph theory and incremental singular value decomposition to perform dynamic community detection; construct a community evolution manifold and use a dynamic manifold learning algorithm to capture the continuous changes in group structure; analyze the manifold trajectory and extract key events in the customer group evolution process; finally output the dynamic customer group structure and evolution pattern set.

[0051] In this embodiment, first, the application of the meta-path guided graph attention mechanism solves the problem that traditional feature fusion methods are difficult to capture complex semantic relationships in heterogeneous information networks. By constructing a set of key meta-paths and applying an adaptive attention mechanism, the method can effectively integrate multivariate path information and generate a more comprehensive and accurate customer feature representation. The introduction of hierarchical contrastive learning further enhances the model's ability to identify subtle feature differences and improves the accuracy and discrimination of customer portraits. The application of the recursive neural topic model innovatively combines the topic model with the recurrent neural network, which can not only capture the temporal evolution of the topic, but also effectively model long-term dependencies, solving the problem that traditional static topic models cannot reflect the dynamic changes of topics. The introduction of the dynamic factor graph model overcomes the limitation that traditional probabilistic graph models are difficult to handle time-varying dependencies, and through dynamic community detection and message passing algorithms, accurate modeling of complex time-varying dependencies between features is achieved. The application of the multi-task learning framework enables the model to simultaneously optimize topic evolution and feature dependency modeling, improving the generalization and robustness of the model. In terms of group analysis, the combination of time-varying customer similarity graphs and dynamic manifold learning algorithms solves the problem that traditional group analysis methods are difficult to capture continuous changes in group structure. By constructing community evolution manifolds and analyzing manifold trajectories, this method can effectively identify key events and turning points in the evolution of customer groups, providing an important basis for companies to formulate dynamic marketing strategies. The comprehensive application of this series of technologies not only realizes the dynamic and multi-dimensional characterization of customer portraits, but also can deeply understand the evolution laws of customer groups, providing powerful decision-making support for precision marketing and customer relationship management. The innovation of this step lies in the organic combination of deep learning, probabilistic graphical models and dynamic system analysis methods, realizing the all-round and dynamic modeling of individual and group characteristics of customers, overcoming the limitations of traditional methods in dealing with large-scale, high-dimensional, and time-varying data.

[0052] According to one aspect of the present application, step S4 is specifically:

[0053] Step S41, obtaining dynamic customer portraits and group analysis results; constructing a multi-relationship graph and a dynamic heterogeneous graph neural network to capture the user's long-term and short-term interest changes and the specific semantics of heterogeneous relationships;

[0054] Dynamic customer portraits include the theme evolution trajectory at the macro level, the time-varying distribution of key features at the meso level, and the instantaneous feature vectors at the micro level.

[0055] Group analysis results, including:

[0056] Dynamic customer group structure: reflects the structure of customer groups changing over time.

[0057] Evolutionary pattern collection: summarizes typical patterns in the process of group evolution.

[0058] Step S42: Based on the dynamic heterogeneous graph neural network, a hierarchical reinforcement learning framework is constructed to construct a state space including customer portraits, group structure and context information, and an action space including recommended items, interaction methods and recommendation timing; a personalized recommendation strategy is generated by training a reinforcement learning agent;

[0059] Step S43, construct a temporal causal graph and apply dual machine learning methods, and use counterfactual reasoning technology to optimize recommendation decisions; integrate the above components (dynamic heterogeneous graph neural network, hierarchical reinforcement learning framework, personalized recommendation strategy, temporal causal graph, dual machine learning methods, counterfactual reasoning technology) to form a complete personalized recommendation model and interactive strategy set.

[0060] In this embodiment, first, the construction of a dynamic heterogeneous graph neural network solves the problem that traditional recommendation systems are difficult to handle multiple types of entities and relationships at the same time. By designing a heterogeneous graph attention layer and a temporal graph evolution module, the network can effectively capture the long-term and short-term changes in user interests and the complex interactions between different types of entities, greatly improving the depth of understanding of user behavior. The introduction of the meta-path-guided semantic aggregation layer further enhances the model's ability to model complex semantic relationships, making the recommendation results more in line with the user's potential needs and preferences. The application of a hierarchical reinforcement learning framework innovatively transforms the recommendation problem into a sequential decision-making process, which can not only optimize long-term benefits, but also adjust the recommendation strategy according to dynamically changing user states and environmental information. By constructing a rich state space containing customer portraits, group structures, and contextual information, as well as a multidimensional action space including recommended items, interaction methods, and recommendation timing, the framework achieves all-round optimization of the recommendation process. The introduction of temporal causal graphs and dual machine learning methods solves the problem that traditional recommendation systems are difficult to handle causal relationships and counterfactual reasoning. By constructing a temporal causal graph and applying counterfactual reasoning technology, the method can more accurately evaluate the potential impact of recommendation decisions, reduce recommendation bias, and improve the fairness and interpretability of recommendations. The adoption of end-to-end training methods and incremental learning strategies enables the entire recommendation system to efficiently adapt to dynamically changing user preferences and market environments, maintaining the real-time and effectiveness of the model.

[0061] According to one aspect of the present application, in step S21, the process of applying the adaptive multi-scale singular spectrum analysis method includes: constructing a multi-scale trajectory matrix, which includes different time scales; performing singular value decomposition on each trajectory matrix to obtain feature triples; using the information entropy criterion to adaptively select the optimal decomposition scale; extracting time feature vectors, including periodicity, trend, mutation and entropy;

[0062] The process of extracting spatial features using the dynamic manifold learning method is as follows: using local linear embedding or isometric mapping algorithms to reduce the high-dimensional spatiotemporal data to a low-dimensional manifold space, obtaining a low-dimensional embedding of the spatiotemporal data, and capturing the dynamic changes of the data structure in the low-dimensional manifold space by sliding the time window; calculating the geometric and topological features on the manifold; extracting spatial feature vectors; the geometric and topological features include curvature, geodesic distance, and spectral features; the spatial feature vectors include local connectivity, global structure, and complexity.

[0063] According to one aspect of the present application, step S12 is specifically:

[0064] Step S121, receiving an enhanced feature data set, constructing a distance matrix using Euclidean distance calculation; generating a similarity matrix based on the distance matrix by applying a Gaussian kernel function; calculating a normalized Laplace matrix based on the similarity matrix; solving a generalized eigenvalue problem to obtain eigenvalues ​​and eigenvectors; selecting an eigenvector corresponding to the minimum non-zero eigenvalue to form a time-varying feature matrix;

[0065] Enhanced feature dataset D = {d1, d2, ..., dn}, where di represents the feature vector of the i-th data point;

[0066] For each data point di, calculate its Euclidean distance to all other data points and construct the distance matrix E.

[0067] Use the Gaussian kernel function K(di, dj) = exp(-||di - dj|| 2 / (2σ 2 ))Construct the similarity matrix S, where σ is the kernel width parameter, selected by cross-validation.

[0068] Calculate the normalized Laplacian matrix L = I - D (-1 / 2) SD (-1 / 2) , where D is the degree matrix and I is the identity matrix.

[0069] Solve the generalized eigenvalue problem Lv = λDv and obtain the eigenvalue λ and eigenvector v.

[0070] Select the eigenvectors corresponding to the k smallest non-zero eigenvalues ​​to form the time-varying characteristic matrix V.

[0071] Step S122, obtaining the time-varying feature matrix of step S121, setting the maximum number of clusters; executing the k-means algorithm for each possible number of clusters to obtain clustering results; calculating the silhouette coefficient and the Calinski-Harabasz index based on the clustering results; drawing curves of the above two indicators changing with the number of clusters; analyzing the curves using the inflection point method to determine the optimal number of clusters;

[0072] Iterate k from 2 to Kmax: a. Apply k-means algorithm to cluster V and get clustering result Ck. b. Calculate silhouette coefficient SC(k). c. Calculate Calinski-Harabasz index CH(k).

[0073] Use the "inflection point method" to determine k*: a. Draw the curves of SC(k) and CH(k) with respect to k. b. Find the inflection point of the curve and determine the optimal number of clusters k*.

[0074] Step S123, receiving the time-varying feature matrix of step S121 and the optimal number of clusters of step S122; calculating the cosine similarity between data points in the matrix; constructing an adjacency matrix using the k-nearest neighbors algorithm; calculating the graph Laplacian matrix based on the adjacency matrix; solving the eigenvector of the graph Laplacian matrix; performing l2 normalization on the eigenvector; performing clustering on the normalized eigenvector using the k-means++ algorithm to obtain a clustering result;

[0075] Construct the similarity graph G: a. Calculate the cosine similarity between each pair of data points in V. b. Use the k-nearestneighbors (kNN) algorithm to construct the adjacency matrix A, retaining the top-k similar neighbors of each point. Calculate the graph Laplacian matrix L = D - A, where D is the degree matrix. Calculate the eigenvectors corresponding to the first k* smallest eigenvalues ​​of L to form the matrix U. Perform l2 normalization on each row of U. Use the k-means++ algorithm to cluster the normalized U and obtain the clustering result C.

[0076] Step S124, obtain the clustering result of step S123, receive new data points; extract the time-varying feature vector of the new data point; calculate the distance between the vector and the existing cluster center; assign the new data point to the nearest cluster; update the center position of the affected cluster; determine whether the cluster structure change exceeds a preset threshold, if so, trigger the re-execution of steps S122 and S123;

[0077] Extract the time-varying feature vector vnew of dnew. Calculate the distance between vnew and the existing cluster center. Assign dnew to the cluster with the closest distance. Update the center of the affected cluster. If the cluster structure changes by more than a preset threshold, trigger the re-execution of S122 and S123.

[0078] Step S125, receiving the clustering results and original data set of step S123 or S124; calculating the Davies-Bouldin index and Dunn index as internal evaluation indicators; if there is labeled data, calculating the adjusted Rand index and normalized mutual information as external evaluation indicators; combining internal and external indicators to generate an overall quality indicator; comparing the overall quality indicator with a preset threshold, if it is lower than the threshold, adjusting the kernel function parameters in step S121, and re-executing steps S121 to S124; outputting the final optimized clustering results and quality indicators.

[0079] Calculate the internal evaluation index of clustering: a. Calculate the Davies-Bouldin index DB. b. Calculate the Dunn index DI.

[0080] Calculate external evaluation indicators for clustering (if there is labeled data): a. Calculate the adjusted Rand index ARI. b. Calculate the normalized mutual information NMI. Combining internal and external indicators, we can get the overall quality indicator Q.

[0081] If Q is lower than the preset threshold: a. Adjust the kernel function parameter σ in S121. b. Re-execute S121 to S124. c. Update the clustering result to C_opt. Output the final optimized clustering result C_opt and the quality index Q.

[0082] In this embodiment, first, the construction of the time-varying feature matrix solves the problem that traditional clustering methods are difficult to handle dynamically changing data. By calculating the normalized Laplace matrix and solving the generalized eigenvalue problem, this method can effectively capture the intrinsic structure and time evolution characteristics of the data, and provide a robust feature representation for subsequent clustering. The automatic determination process of the optimal number of clusters innovatively combines the silhouette coefficient and the Calinski-Harabasz index, and realizes the adaptive selection of the number of clusters through the inflection point analysis, overcoming the limitation that the traditional method requires artificial setting of the number of clusters. The application of the spectral clustering algorithm solves the problem that traditional K-means is difficult to handle non-convex data sets. By constructing a similarity graph and calculating the graph Laplace matrix, this method can better capture the global structure of the data and improve the accuracy and stability of clustering. The introduction of the incremental clustering update strategy enables the model to efficiently process streaming data, and by dynamically adjusting the clustering structure, the real-time and adaptability of the classification results are guaranteed. The design of the clustering quality evaluation mechanism combines internal evaluation indicators (such as the Davies-Bouldin index and the Dunn index) and external evaluation indicators (such as the adjusted Rand index and normalized mutual information) to achieve a comprehensive evaluation of the clustering results. By comparing with the preset threshold and adjusting the kernel function parameters through feedback, the mechanism can continuously optimize the clustering performance and improve the robustness and generalization ability of the algorithm. The comprehensive application of this series of technologies not only greatly improves the classification accuracy of complex spatiotemporal data, but also realizes the adaptive optimization and quality assurance of the clustering process.

[0083] According to one aspect of the present application, step S13 is specifically:

[0084] Step S131, obtaining the original spatiotemporal data, detecting the timestamp format therein; converting the detected timestamps into the Unix timestamp format; decomposing the converted time series by wavelet transform; extracting multi-scale time features from the decomposition results; resampling the extracted time series by using the adaptive time window method to obtain standardized time features;

[0085] Step S132, receiving the original spatial data, converting all the spatial coordinates therein into the WGS84 geographic coordinate system; for non-point type spatial data, extracting feature points using a curvature-based shape descriptor algorithm; inputting the extracted feature points into a Riemann manifold learning model, performing a nonlinear normalization operation, and outputting standardized spatial features;

[0086] Step S133, obtaining the standardized time features of step S131 and the standardized space features of step S132; constructing a spatiotemporal tensor, integrating the time and space features into the same data structure; applying the Tucker decomposition method to the constructed spatiotemporal tensor to perform feature fusion; introducing an attention mechanism, dynamically adjusting the weights of the time and space features, and outputting the fused spatiotemporal features;

[0087] Step S134, receiving the fused spatiotemporal features of step S133, calculating the local and global statistical features thereof; dynamically adjusting the scale parameters of the spatiotemporal analysis based on the calculated statistical features; applying the adjusted scale parameters to the multi-scale analysis method, capturing the spatiotemporal patterns at different scales, and outputting the spatiotemporal features of the adaptive scale;

[0088] Step S135, obtain the original spatiotemporal data and the adaptive scale spatiotemporal features of step S134; calculate the information entropy of the data before and after standardization to evaluate the degree of information retention; use the Kolmogorov-Smirnov test method to test the effect of standardization; compare the test result with the preset threshold, if it does not meet the requirements, adjust the relevant parameters of steps S131 to S134, and re-execute these steps until the standardization quality requirements are met, and output the final high-quality standardized spatiotemporal data.

[0089] In this embodiment, first, the wavelet transform introduced in the time standardization process solves the problem that traditional methods are difficult to handle multi-scale time features at the same time. By performing multi-resolution decomposition on the time series, the method can effectively capture the periodicity, trend and mutation characteristics of the data, and provide rich time dimension information for subsequent analysis. The application of the adaptive time window method overcomes the limitation that the fixed sampling interval is difficult to adapt to different time scales, realizes the dynamic resampling of the time series, and improves the accuracy and representativeness of the standardization results. In the process of spatial standardization, the introduction of the curvature-based shape descriptor algorithm solves the problem that traditional methods are difficult to handle complex geometric shapes. By extracting key feature points, the method can effectively retain the topological structure and geometric features of spatial data. The application of Riemann manifold learning innovatively maps nonlinear data to manifold space, realizes effective dimensionality reduction and normalization of high-dimensional spatial data, and overcomes the limitations of the Euclidean space standardization method. The construction of spatiotemporal tensors and the application of Tucker decomposition solve the problem that traditional methods are difficult to handle time and space features at the same time. Through feature fusion and dimensionality reduction, the method can effectively capture the intrinsic structure and correlation of spatiotemporal data. The introduction of the attention mechanism further enhances the model's ability to identify key spatiotemporal features and achieves dynamic adjustment of the importance of temporal and spatial features. The application of multi-scale analysis methods overcomes the limitation that a single scale is difficult to fully capture spatiotemporal patterns. By dynamically adjusting the analysis scale, this method can simultaneously identify local details and global trends. Finally, the design of the standardization quality assessment mechanism achieves quantitative evaluation and adaptive optimization of the standardization effect by calculating information entropy and applying the Kolmogorov-Smirnov test.

[0090] According to one aspect of the present application, step S15 is specifically:

[0091] Step S151, receiving the quality-scored data set, constructing a multi-dimensional spatiotemporal feature vector; applying a kernel density estimation method to the constructed feature vector to obtain an empirical distribution of the data; and setting an initial anomaly detection threshold;

[0092] Step S152, obtaining the empirical distribution of step S151, calculating the Wasserstein distance between each data point and the empirical distribution; inputting the calculated Wasserstein distance sequence into the extreme value theory model to estimate the tail behavior of the distance distribution; dynamically adjusting the anomaly detection threshold based on the tail behavior estimation result;

[0093] Step S153, receiving the Wasserstein distance sequence and dynamic threshold of step S152, comparing the distance value of each data point with the threshold; marking the data points whose distance is greater than the threshold as potential outliers; applying the local outlier factor (LOF) algorithm to the marked potential outliers, performing secondary verification, and obtaining the final outlier set;

[0094] Step S154, obtaining the abnormal value set of step S153, and constructing an abnormal pattern library; for each abnormal value, extracting its spatiotemporal context information; using the extracted context information, training an abnormal prediction model based on a long short-term memory network; applying the trained model to the real-time data stream to perform online anomaly detection;

[0095] Step S155, receiving the anomaly prediction result of step S154, and comparing it with the actually observed anomaly; calculating the prediction accuracy, recall rate and F1 score; based on the calculated performance indicators, using the Bayesian optimization method to adjust the relevant parameters of Wasserstein distance calculation and anomaly detection; feeding back the optimized parameters to steps S151-S154 to achieve continuous optimization of the anomaly detection process.

[0096] In this embodiment, first, the introduction of Wasserstein distance solves the problem that traditional anomaly detection methods are difficult to handle high-dimensional distribution differences. By calculating the Wasserstein distance between data points and empirical distributions, this method can effectively capture anomalies at the distribution level and improve the sensitivity to small deviations. The application of extreme value theory models innovatively realizes the dynamic adjustment of anomaly detection thresholds, overcomes the limitation that fixed thresholds are difficult to adapt to changes in data distribution, and enhances the adaptability and robustness of detection. The introduction of the local anomaly factor (LOF) algorithm provides a secondary verification mechanism for potential outliers. By considering the local density of data points, this method can effectively identify local anomalies and reduce the false alarm rate. The construction of an anomaly prediction model based on long short-term memory network (LSTM) solves the problem that traditional methods are difficult to capture time series patterns. By learning the spatiotemporal context information of anomalies, the model can realize the prediction of future anomalies and improve the foresight of detection. The adoption of online anomaly detection strategies enables the model to process data streams in real time and promptly discover and respond to emerging abnormal patterns. The design of performance evaluation and parameter optimization mechanism achieves continuous optimization of the anomaly detection process by calculating accuracy, recall and F1 score and applying Bayesian optimization method, ensuring the efficiency and effectiveness of the model in a dynamic environment.

[0097] In another embodiment of the present application, step S23, semantic trajectory and spatiotemporal pattern extraction, specifically includes the following sub-steps:

[0098] Step S231, obtain the original trajectory data, apply the trajectory segmentation algorithm based on the attention mechanism; identify the semantic stay points and moving segments in the trajectory; extract the spatiotemporal features and duration information for the identified stay points; for the moving segments, calculate the dynamic features such as speed, acceleration and direction change; associate the extracted features with the original trajectory data to form an enhanced trajectory dataset.

[0099] Step S232, receiving the enhanced trajectory dataset of step S231, applying a hierarchical topic model to the stop points therein; constructing a multi-level semantic concept tree, including low-level specific locations and high-level abstract concepts; using word embedding technology, mapping semantic labels to a continuous vector space; based on the mapping results, calculating the similarity between different semantic labels to form a semantic association network.

[0100] Step S233, obtain the mobile segment data of step S231, apply the temporal association rule mining algorithm; extract frequent mobile patterns, including typical paths and recurring trajectory segments; calculate the support and confidence of the extracted mobile patterns; filter out significant mobile patterns according to a preset threshold; associate the filtered patterns with the semantic tags of step S232 to build a semantically enhanced mobile pattern library.

[0101] Step S234, receiving the semantic label of step S232 and the movement mode of step S233, integrating them to form an initial semantic trajectory; applying a sequence alignment algorithm to compare and cluster multiple semantic trajectories; extracting common semantic subsequences and identifying typical activity patterns; using a hidden Markov model to probabilistically model the semantic trajectory and capture the semantic conversion rules; outputting the semantic trajectory model, including a state transition matrix and a transmission probability matrix.

[0102] Step S235, obtain the semantic trajectory model of step S234, apply the sliding time window method, and perform online analysis on the real-time trajectory data; match the real-time data with the historical pattern to identify the current semantic state; predict the next possible semantic position based on the current state and transition probability; combine the spatial constraint information to generate a specific position prediction; compare the prediction result with the actual observation, dynamically update the model parameters, and realize the adaptive optimization of the semantic trajectory model.

[0103] In this embodiment, first, the trajectory segmentation algorithm based on the attention mechanism solves the problem that traditional methods are difficult to identify semantic stay points. By adaptively focusing on key points in the trajectory, the algorithm can effectively distinguish stay points from mobile segments, improving the accuracy of trajectory segmentation and the depth of semantic understanding. The application of the hierarchical topic model innovatively associates trajectory points with abstract semantic concepts. By constructing a multi-level semantic concept tree, the method can capture the fine-grained and coarse-grained semantics of the location and enhance the interpretability of the trajectory. The introduction of word embedding technology enables semantic labels to be represented in a continuous vector space, laying the foundation for calculating semantic similarity and building a semantic association network. The application of the temporal association rule mining algorithm solves the problem that traditional methods are difficult to discover complex mobile patterns. By extracting frequent mobile patterns and calculating their support and confidence, the method can identify statistically significant spatiotemporal behavior patterns. The combination of the sequence alignment algorithm and the hidden Markov model realizes the comparison and probabilistic modeling of multiple semantic trajectories, which can not only extract common semantic subsequences, but also capture the laws of semantic conversion, providing in-depth insights into understanding user activity patterns. The introduction of sliding time windows and online analysis methods enables the model to process streaming trajectory data in real time, and through dynamic matching and prediction, it realizes real-time recognition of the user's current state and prediction of future location. The design of the adaptive optimization mechanism continuously updates the model parameters by comparing the prediction results with the actual observations, thus ensuring the timeliness and accuracy of the semantic trajectory model.

[0104] In another embodiment of the present application, step S24, building a personalized spatiotemporal preference model, specifically includes the following sub-steps:

[0105] Step S241, obtain user historical behavior data, including location check-in, activity participation and evaluation information; construct a multi-dimensional preference tensor, which includes dimensions such as user, location, time and activity type; apply tensor decomposition methods, such as CANDECOMP / PARAFAC (CP) decomposition, to extract latent feature vectors; regularize the extracted feature vectors to obtain an initialized user preference representation.

[0106] Step S242, receiving the user preference representation of step S241, and converting it into a heterogeneous graph structure; the nodes in the graph represent users, locations and activities, and the edges represent interactive relationships; applying a graph attention network (GAT) to learn complex dependencies between nodes; through a multi-head attention mechanism, capturing the importance of different types of relationships; outputting a user preference embedding enhanced by the graph structure.

[0107] Step S243, obtain the user preference embedding of step S242, and construct a temporal preference sequence in combination with time information; apply a long short-term memory network (LSTM) to model the temporal dynamics of user preferences; introduce an attention mechanism based on LSTM to adaptively focus on preference changes at different time scales; and output a dynamic preference representation that captures time dependence.

[0108] Step S244, receiving the dynamic preference representation of step S243, combining it with geographic spatial information; constructing a distance-based spatial influence model to characterize the interaction between locations; applying a spatial attenuation function to simulate the user's preference intensity for locations at different distances; combining the spatial influence model with the dynamic preference representation to obtain a personalized preference model coupled in time and space.

[0109] Step S245, obtain the spatiotemporal coupling preference model of step S244, apply the multi-task learning framework, and simultaneously predict the user's next location, activity type, and time; design a shared underlying representation and task-specific layer to balance generality and specificity; use contrastive learning methods to enhance the model's ability to recognize similar preference patterns; through online learning strategies, continuously update model parameters to adapt to real-time changes in user preferences; output the final personalized spatiotemporal preference model, including prediction functions and dynamic update mechanisms.

[0110] In this embodiment, first, the construction of a multi-dimensional preference tensor and the application of CP decomposition solve the problem that traditional methods are difficult to handle multiple types of preferences at the same time. By extracting potential feature vectors and performing regularization processing, this method can effectively capture the multi-dimensional characteristics of user preferences, laying the foundation for subsequent personalized modeling. The introduction of heterogeneous graph structures and the application of graph attention networks (GAT) innovatively encode the complex relationships between users, locations, and activities into graph structures. Through the multi-head attention mechanism, this method can adaptively learn the importance of different types of relationships and enhance the modeling ability of the model for complex dependencies. The combination of long short-term memory networks (LSTMs) and attention mechanisms solves the problem that traditional methods are difficult to capture long-term trends and short-term changes at the same time. By adaptively focusing on preference changes at different time scales, this method can effectively model the temporal dynamics of user preferences. The application of distance-based spatial influence models and spatial attenuation functions overcomes the limitation of traditional methods that ignore spatial correlations, and achieves accurate modeling of spatial factors by simulating the strength of users' preferences for locations at different distances. The design of the multi-task learning framework enables the model to simultaneously predict the user's next location, activity type, and time. By sharing the underlying representation and designing task-specific layers, the framework ensures the versatility of the model while taking into account the specificity of each task. The introduction of the contrastive learning method further enhances the model's ability to recognize similar preference patterns and improves the accuracy and stability of the prediction. The adoption of the online learning strategy enables the model to continuously update parameters and adapt to real-time changes in user preferences, ensuring the timeliness and adaptability of the model.

[0111] Step S25, spatiotemporal graph convolutional neural network anomaly detection, specifically includes the following sub-steps:

[0112] Step S251, receiving user behavior data and environmental context information, and constructing a dynamic spatiotemporal graph; the nodes in the graph represent users and locations, and the edges represent interactive relationships; assigning time-varying feature vectors to each node and edge to capture dynamic changes in behavior and environment; applying a sliding time window method to generate a series of time series graph snapshots as input for anomaly detection.

[0113] Step S252, obtain the snapshot of the timing graph of step S251, and design a multi-layer spatiotemporal graph convolution structure; apply graph convolution operations in the spatial dimension to capture local spatial dependencies; use causal convolutions in the temporal dimension to model timing patterns; introduce skip connections to alleviate the gradient vanishing problem of deep networks; output high-dimensional spatiotemporal feature representations of each node.

[0114] Step S253, receiving the node feature representation of step S252, applying the graph attention pooling layer; adaptively aggregating node information of different scales to form a graph-level representation; designing a multi-head attention mechanism to capture different types of abnormal patterns; retaining the original feature information through residual connections; and outputting graph-level abnormal indication features.

[0115] Step S254, obtain the abnormal indication features of step S253, and construct an abnormal scoring module based on variational autoencoder (VAE); encode the normal behavior pattern into the latent space to form a normal behavior distribution; calculate the degree of deviation between the sample to be detected and the normal distribution, and generate an abnormal score; set an adaptive threshold, and mark the samples above the threshold as potential abnormalities.

[0116] Step S255, receiving the potential anomaly samples of step S254, applying the ensemble learning method for secondary verification; combining traditional anomaly detection algorithms such as isolation forest and one-class SVM to evaluate potential anomalies from multiple angles; using a weighted voting strategy to integrate the results of multiple detectors; for confirmed anomalies, extracting their spatiotemporal context features and updating the anomaly pattern library; through the online learning mechanism, continuously optimizing the detection model to improve the ability to recognize new anomalies; outputting the final anomaly detection results and the updated anomaly detection model.

[0117] In this embodiment, first, the construction of a dynamic spatiotemporal graph and the application of a sliding time window method solve the problem that traditional anomaly detection is difficult to consider both temporal and spatial relationships at the same time. By assigning time-varying feature vectors to nodes and edges, this method can effectively capture the dynamic changes of behavior and environment, and provide rich contextual information for subsequent analysis. The design of a multi-layer spatiotemporal graph convolution structure innovatively combines spatial graph convolution and temporal causal convolution, which can not only capture local spatial dependencies, but also model complex temporal patterns, greatly improving the ability to characterize spatiotemporal data. The introduction of a graph attention pooling layer solves the problem that traditional pooling methods are difficult to handle irregular graph structures. By adaptively aggregating node information of different scales, this method can generate a more representative graph-level representation and enhance the model's ability to recognize global abnormal patterns. The design of an anomaly scoring module based on a variational autoencoder (VAE) overcomes the limitation that traditional methods are difficult to model high-dimensional probability distributions. By encoding normal behavior patterns in the latent space, this method can effectively detect abnormal samples that deviate from the normal distribution, improving the accuracy and sensitivity of detection. The application of ensemble learning methods provides a multi-angle verification mechanism for anomaly detection. By combining multiple traditional anomaly detection algorithms and using a weighted voting strategy, this method can reduce false positives and false negatives and improve the reliability of detection results. The introduction of online learning mechanisms enables the model to be continuously updated and optimized. By continuously learning new anomaly patterns, the effectiveness and adaptability of the detection system in a dynamic environment are guaranteed.

[0118] Step S32, recursive neural topic model and dynamic factor graph model, specifically includes the following sub-steps:

[0119] Step S321, obtain the fused feature vector set and construct a document-word matrix; apply word embedding technology to map words to a continuous vector space; use a sliding time window to divide the data into continuous time segments; for each time segment, initialize the topic-word distribution and document-topic distribution to prepare for the recursive neural topic model.

[0120] Step S322, receiving the initialized distribution of step S321, constructing a recursive neural topic model; designing a recurrent neural network (RNN) structure to capture the temporal dependency of topic evolution; applying variational inference based on RNN to estimate the posterior probability of topic distribution; optimizing model parameters through back propagation algorithm; outputting the dynamic topic distribution of each time segment.

[0121] Step S323, obtain the dynamic topic distribution of step S322, and construct a time-varying feature correlation network; calculate the mutual information between features and quantify the strength of feature correlation; apply the dynamic community detection algorithm to identify highly correlated feature clusters; for each feature cluster, construct factor nodes to form a skeleton structure of the dynamic factor graph.

[0122] Step S324, receiving the factor graph skeleton of step S323, and designing the factor function; encoding the topic distribution information into the factor function to express the conditional dependency between features; applying the message passing algorithm, such as belief propagation, to perform inference on the factor graph; estimating the edge probability distribution of each variable node through iterative updating; and outputting a probabilistic graph model that reflects the dynamic dependency of features.

[0123] Step S325, obtain the dynamic topic model of step S322 and the dynamic factor graph model of step S324, and design a model fusion mechanism; use the attention mechanism to dynamically adjust the weights of the two models in different time and feature dimensions; apply a multi-task learning framework to simultaneously optimize topic evolution and feature dependency modeling; enhance the model's sensitivity to small changes through contrastive learning; design an online update strategy to enable the model to adapt to real-time changes in data distribution; output the final dynamic customer portrait, including a multi-level topic structure and a time-varying feature dependency network.

[0124] In this embodiment, first, the design of the recursive neural topic model solves the problem that traditional topic models are difficult to capture topic evolution. By combining the recurrent neural network (RNN) with the topic model, the method can effectively model the dynamic changes of topics over time, capture long-term dependencies, and provide in-depth insights into the evolution of customer interests. The application of variational inference overcomes the computational bottleneck of traditional methods in processing large-scale data, and achieves efficient optimization of model parameters by approximating the posterior distribution. The introduction of the dynamic factor graph model innovatively combines the probabilistic graph model with time-varying features. By constructing a time-varying feature correlation network and applying a dynamic community detection algorithm, the method can accurately capture the complex time-varying dependencies between features and enhance the model's ability to express dynamic changes in customer features. The application of the message passing algorithm solves the problem of accurate inference on complex graph structures, and achieves accurate estimation of customer status by iteratively updating the edge probability distribution of nodes. The introduction of the attention mechanism further enhances the model's ability to adaptively adjust the importance of different time and feature dimensions, and improves the accuracy and personalization of the portrait. The design of the multi-task learning framework enables the model to simultaneously optimize topic evolution and feature dependency modeling, and improves the generalization ability and efficiency of the model through shared representation learning. The application of contrastive learning methods enhances the model’s sensitivity to small changes, enabling the profile to capture subtle shifts in customer behaviors and preferences.

[0125] Step S41, dynamic heterogeneous graph neural network construction, specifically includes the following sub-steps:

[0126] Step S411, receiving multi-source heterogeneous data, including user behavior logs, social network data and environmental perception information; constructing multiple types of entities (such as users, commodities, locations) and relationships; constructing an initial heterogeneous information network, in which nodes represent entities and edges represent interactions or relationships between entities; assigning an original feature vector to each node and edge as the initial state of the network.

[0127] Step S412, obtain the initial heterogeneous network of step S411, and design a heterogeneous graph attention layer; apply type-specific attention calculations to different types of neighbor nodes; capture the importance of nodes under different semantics through a multi-head attention mechanism; aggregate information from different types of neighbors to generate contextual embeddings of nodes; and output node representations that take heterogeneity into account.

[0128] Step S413, receive the node representation of step S412, and construct a timing graph evolution module; use time encoding technology to inject time information into the representation of nodes and edges; design a graph evolution unit based on gated recurrent units (GRU) to model the dynamic changes of graph structure; apply an adaptive sampling strategy to balance historical information and the latest state; output a node representation sequence that captures the temporal dynamics.

[0129] Step S414, obtain the dynamic node representation of step S413, design a meta-path guided semantic aggregation layer; pre-build a set of meaningful meta-paths to capture complex semantic relationships; for each meta-path, apply a path-specific attention mechanism to aggregate node information on the path; through a multi-task learning framework, simultaneously optimize the node representation under different meta-paths; output node embeddings that incorporate rich semantic information.

[0130] Step S415, receiving the semantically enhanced node embedding of step S414, and constructing a prediction layer that is aware of the graph structure; designing a graph pooling operation to generate a graph-level representation; applying a graph matching network to capture the global similarity of the graph structure; introducing a contrastive learning mechanism to enhance the model's sensitivity to graph structure changes; training the entire network in an end-to-end manner to optimize node representation, graph representation, and downstream task objectives; designing an incremental learning strategy to enable the model to efficiently handle dynamically changing graph structures; and outputting the final dynamic heterogeneous graph neural network model, including trained network parameters and an online update mechanism.

[0131] In this embodiment, first, the fusion of multi-source heterogeneous data and the construction of the initial heterogeneous information network solve the problem that traditional methods are difficult to process multiple types of data at the same time. By encoding different entities and relationships such as users, commodities, and locations into a unified graph structure, this method lays the foundation for subsequent deep learning. The design of the heterogeneous graph attention layer innovatively combines the attention mechanism with graph structure learning. By applying type-specific attention calculations to different types of neighbor nodes, this method can effectively capture the specific semantics of heterogeneous relationships and improve the expressiveness of node representation. The introduction of the multi-head attention mechanism further enhances the model's adaptive learning ability for the importance of nodes under different semantics, and realizes multi-angle analysis of complex user behaviors. The construction of the time-series graph evolution module solves the problem that traditional static graphs are difficult to reflect dynamic changes. By introducing time coding technology and designing a GRU-based graph evolution unit, the model can accurately capture the dynamic changes of the graph structure and realize continuous modeling of the evolution of user interests. The application of the adaptive sampling strategy balances the importance of historical information and the latest state, and improves the model's ability to recognize long-term trends and short-term changes. The design of the meta-path-guided semantic aggregation layer overcomes the limitation of traditional methods in capturing complex semantic relationships. By pre-constructing meaningful meta-paths and applying path-specific attention mechanisms, this method can effectively aggregate rich semantic information and enhance the semantic connotation of node representation. The introduction of the multi-task learning framework enables the model to simultaneously optimize the node representation under different meta-paths, improving the learning efficiency and generalization ability of the model. The design of the graph structure-aware prediction layer solves the problem that traditional methods are difficult to utilize global graph structure information. Through graph pooling operations and graph matching networks, this method can effectively capture the global similarity of graph structures and improve the accuracy and robustness of prediction. The application of the contrastive learning mechanism further enhances the model's sensitivity to changes in graph structure and improves the effects of anomaly detection and pattern recognition. The adoption of the incremental learning strategy enables the model to efficiently handle dynamically changing graph structures, ensuring the real-time nature of online learning and prediction.

[0132] According to another aspect of the present application, a customer portrait key data mining system based on spatiotemporal big data is also provided, comprising:

[0133] at least one processor; and,

[0134] a memory communicatively connected to at least one of the processors; wherein,

[0135] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the customer portrait key data mining method based on spatiotemporal big data described in any of the above technical solutions.

[0136] According to one aspect of the present application, the RMLGF (Riemannian Manifold Learning via Geodesic Flows) method is used in S13 to perform nonlinear normalization of spatial data. RMLGF is implemented by the following steps:

[0137] A k-nearest neighbor graph (k=10) is constructed to represent the local relationship between data points. Geodesic flow is calculated and the shortest path problem is solved using the Dijkstra algorithm. The obtained geodesic distance matrix is ​​used to map the data to a low-dimensional Euclidean space using the multidimensional scaling (MDS) method. The local linear embedding (LLE) algorithm is applied to maintain the local geometric structure and obtain the final normalized representation. Data flow continuity is optimized.

[0138] In another embodiment of the present application, step S2 also includes a data integration and feature fusion process, specifically including: obtaining the spatiotemporal feature matrix F, customer behavior pattern P, time-varying community structure C(t), semantic trajectory ST, spatiotemporal pattern M, personalized spatiotemporal preference model PM, anomaly detection model AD and abnormal behavior set AB.

[0139] Use tensor decomposition methods (such as CP decomposition) to convert each feature matrix and model parameter into a unified tensor representation. Apply Tucker decomposition to perform multimodal feature fusion to obtain a high-order tensor G. Use tensor network (TensorNetwork) methods, such as Matrix Product State (MPS), to compress the high-order tensor G into a low-rank representation. Output the fused feature tensor G and its low-rank representation G_low.

[0140] In another embodiment of the present application, step S24 also includes a behavior evaluation based on a personalized preference model, specifically including:

[0141] Obtain personalized spatiotemporal preference model PM and customer real-time behavior data RB.

[0142] Using the conditional random field (CRF) model, PM is used as prior knowledge to perform sequence labeling on RB. The KL divergence between the behavior sequence and the expected preference is calculated to quantify the degree of behavioral deviation. The sliding window method is used to calculate the behavioral deviation trend at different time scales. The behavioral deviation score BD and deviation trend DT are output.

[0143] In another embodiment of the present application, step S24 is provided with a dynamic attention mechanism to construct a multi-head attention layer, each head corresponding to a preference dimension (such as time, location, activity type, etc.). A gated recurrent unit (GRU) is used to model the temporal changes of user preferences. The weights of each attention head are dynamically adjusted to adapt to the real-time changes of user preferences.

[0144] In another embodiment of the present application, the evaluation indicators include:

[0145] The quality of the spatiotemporal feature matrix is ​​evaluated using the Explained Variance Ratio (EVR).

[0146] Validity of customer behavior patterns: Use the Silhouette Coefficient SC (Silhouette Coefficient) to evaluate clustering effects.

[0147] Stability of time-varying community structure: evaluated using normalized mutual information (NMI).

[0148] According to one aspect of the present application, a specific implementation is provided, including:

[0149] Step S1 first obtains the original spatiotemporal big data from multiple heterogeneous data sources, and uses the Transformer-based feature extraction method to preliminarily process the data to obtain an enhanced feature data set. Then, the enhanced feature data set is classified using a dynamic data source classification method to obtain a classified data set. Subsequently, an adaptive spatiotemporal standardization transformation is applied to the classified data set to obtain a standardized spatiotemporal data set. Next, the standardized data is scored for quality using a data quality assessment algorithm based on graph theory to obtain a quality-scored data set and a data set to be repaired. Then, the quality-scored data is cleaned using an anomaly detection method based on Wasserstein distance to obtain a cleaned data set and a set of outliers. Finally, a tensor completion algorithm based on graph neural networks is applied to complete and enhance the cleaned data to obtain the final high-quality spatiotemporal data set.

[0150] Step S11: Transformer-based feature extraction and enhancement

[0151] Obtain the original spatiotemporal big data D = {D1, D2, ..., Dn} from different sources, where Di represents the data of the i-th data source. For each data source Di, perform the following processing:

[0152] First, the spatiotemporal data is serialized into a format suitable for Transformer processing to obtain the input sequence Xi = [x1, x2, ..., xm], where xj represents a single spatiotemporal data point.

[0153] Then, Xi is input into the pre-trained Transformer model. The model contains a multi-layer self-attention mechanism and a feedforward neural network with the following structure:

[0154] Input embedding layer: maps the original data to a high-dimensional vector space. Position encoding: adds position information to preserve the sequence order. Multi-head self-attention layer: captures the long-range dependencies between data points. Feedforward neural network layer: further processes the output of the attention layer. Layer normalization and residual connection: ensure the stability and training effect of the model. Through the Transformer model, the enhanced feature sequence Yi = [y1, y2, ..., ym] is obtained, where yj is the enhanced feature representation of the corresponding input xj.

[0155] Finally, the enhanced feature sequence Yi is aggregated to obtain the overall feature representation Fi of the data source Di. The aggregation method can be a simple average pooling or a more complex attention weighting mechanism.

[0156] Through this processing, the enhanced feature data set D* = {D1, D2, ..., Dn} is obtained, where Di contains the original data Di and its corresponding enhanced features Fi.

[0157] Step S12 (formerly S11): Data source identification and classification

[0158] Obtain the enhanced feature dataset D* = {D1, D2, ..., Dn} output by step S11. For each enhanced feature dataset Di, extract the feature vector Fi = [f1, f2, ..., fk], where fj represents features such as time granularity, spatial precision, and data type. These features now contain high-level semantic information extracted by Transformer. Then, use the dynamic spectral clustering algorithm to cluster the feature vectors. The specific steps are as follows: construct a dynamic similarity matrix S(t), calculate its Laplace matrix L(t), solve the generalized eigenvalue problem to obtain the time-varying feature vector, and finally perform clustering based on the time-varying feature vector. Through this process, the classified dataset D' = {D'1, D'2, ..., D'm} is obtained, where m ≤ n.

[0159] Step S13: Spatiotemporal normalization

[0160] Get the classified data set D' output from step S12, and perform spatiotemporal standardization on each classified data set D'i. First, perform time standardization: detect the timestamp format and convert it to Unix timestamp, then apply wavelet transform to decompose the time series, extract multi-scale time features, and finally resample the time series using adaptive time window method. Next, perform spatial standardization: convert all spatial coordinates into a unified geographic coordinate system (such as WGS84), for non-point data, apply curvature-based shape descriptors to extract feature points, and then use Riemann manifold learning method to perform nonlinear normalization on spatial data. Through these processes, a standardized spatiotemporal data set D'' = {D''1, D''2,..., D''m} is obtained.

[0161] Step S14: Data quality assessment and repair

[0162] Get the standardized spatiotemporal dataset D'' output from step S13, and perform data quality assessment on each dataset D''i. First, construct a spatiotemporal hypergraph H(V,E), where the vertex V represents the data point and the hyperedge E represents the complex spatiotemporal relationship. Then, apply the hypergraph theory to calculate the hypergraph Laplace eigenvalue λ(v) of each vertex. Next, construct the quality score Q(v) = f(λ(v)), where f is a nonlinear mapping function based on deep learning. Mark the data points where Q(v) < Q_threshold is the dataset R to be repaired. Through this process, the quality-scored dataset D''' and the dataset R to be repaired are obtained.

[0163] Step S15: Outlier detection and processing

[0164] Get the quality-scored data set D''' output in step S14, and perform outlier detection on each data point therein. First, construct a spatiotemporal feature vector X = [time feature, space feature, attribute feature]. Then, apply an anomaly detection algorithm based on Wasserstein distance: estimate the empirical distribution P of the data, and for each data point x, calculate its Wasserstein distance W(δ_x, P) with P, and use extreme value theory to dynamically set the anomaly threshold θ. Mark the data points with W(δ_x, P) > θ as outliers and add them to the outlier set A. Through this process, the cleaned data set D'''' and the outlier set A are obtained.

[0165] Step S16: Data completion and enhancement

[0166] Get the cleaned data set D'''', the data set R to be repaired and the outlier set A output by step S15, and perform data completion and enhancement. First, construct a spatiotemporal-attribute fourth-order tensor T, and introduce an additional modal dimension to represent data reliability. Then, apply a tensor completion algorithm based on a graph neural network: represent the tensor T as a heterogeneous graph G, design a graph attention network (GAT) layer to learn node representation, and use a graph autoencoder architecture to reconstruct missing values. For areas with insufficient spatiotemporal density, a conditional generation model based on Gaussian Copula is used to generate new data points. Through this series of processing, a high-quality spatiotemporal data set D* is finally obtained after completion.

[0167] Step S2: Multi-dimensional spatiotemporal feature extraction and customer behavior pattern analysis

[0168] Step S2 first obtains the high-quality spatiotemporal dataset D* output by step S1, applies the adaptive multi-scale singular spectrum analysis method and the topological data analysis method to extract the spatiotemporal features, and obtains the spatiotemporal feature matrix F. Then, based on the spatiotemporal feature matrix F, a customer behavior model is constructed using the graph attention diffusion network to obtain the customer behavior pattern set P and the time-varying community structure C(t). Next, the original trajectory data is subjected to the trajectory segmentation algorithm based on the attention mechanism and the hierarchical topic model to extract the semantic trajectory set ST and the frequent spatiotemporal pattern set M. Subsequently, a personalized spatiotemporal preference model is constructed using the multi-dimensional preference tensor and the graph neural network to obtain the customer personalized spatiotemporal preference model set PM. Finally, an anomaly detection model based on the spatiotemporal graph convolutional neural network is designed to obtain the abnormal behavior detection model AD and the detected abnormal behavior set AB.

[0169] Step S21: Spatiotemporal sequence decomposition and feature extraction

[0170] Get the high-quality spatiotemporal dataset D* output in step S1 and perform the following operations on each customer’s spatiotemporal sequence:

[0171] First, the adaptive multiscale singular spectrum analysis (AMSSA) method is applied. A multiscale trajectory matrix X(τ) is constructed, where τ represents different time scales. Each X(τ) is subjected to singular value decomposition to obtain a feature triple (U_i, σ_i, V_i). The information entropy criterion is used to adaptively select the optimal decomposition scale τ*. The time feature vector T = [periodicity, trend, mutation, entropy] is extracted.

[0172] Then, the topological data analysis (TDA) method is applied to extract spatial features. The persistence homology based on the Vietoris-Rips complex is constructed, and the persistence graph and persistence landscape are calculated. The spatial feature vector S = [connectivity, void, spatial complexity] is extracted.

[0173] Finally, the temporal feature vector T and the spatial feature vector S are combined to obtain the spatiotemporal feature matrix F = [T, S].

[0174] Step S22: Customer behavior pattern discovery

[0175] Obtain the spatiotemporal feature matrix F outputted in step S21, and perform the following steps:

[0176] First, a spatiotemporal behavior graph G(V,E) is constructed, where the vertex V represents the customer and the edge E represents the spatiotemporal behavior similarity between customers.

[0177] Then, a Graph Attention Diffusion Network (GADN) is applied. A multi-head attention mechanism is designed to capture different types of spatiotemporal dependencies, a diffusion convolutional layer is used to aggregate spatiotemporal information at multiple scales, and the customer representation H is learned through end-to-end training.

[0178] Next, a dynamic community detection algorithm is applied to the learned customer representation H. A time-varying modularity matrix Q(t) is constructed, and the optimization problem max Tr(H^TQ(t) H) is solved, subject to H^TH = I. H is clustered using a spectral clustering method.

[0179] Finally, we obtain the customer behavior pattern set P = {P1, P2, ..., Pk} and the time-varying community structure C(t).

[0180] Step S23: Spatiotemporal semantic trajectory extraction

[0181] Get the original trajectory data in step S1 and perform the following operations on each customer’s trajectory:

[0182] First, a trajectory segmentation algorithm based on the attention mechanism is applied. A self-attention matrix A is constructed, and a multi-scale convolutional neural network is used to extract local features. The self-attention and local features are combined to identify semantic stay points and moving segments.

[0183] Then, a hierarchical topic model (HTM) is applied to the identified stay points. A word-document-corpus hierarchy is constructed, and the topic distribution is learned using a variational inference method to extract multi-level semantic labels.

[0184] Then, the temporal association rule mining algorithm is used to extract the mobility pattern. The spatiotemporal item set I = {(l, t, Δt)} is constructed, and a pattern growth algorithm based on the minimum description length (MDL) principle is designed.

[0185] Finally, we obtain the semantic trajectory set ST = {ST1, ST2, ..., STn} and the frequent spatiotemporal pattern set M.

[0186] Step S24: Personalized spatiotemporal preference modeling

[0187] Obtain the semantic trajectory ST and frequent spatiotemporal pattern M output in step S23, and perform the following steps for each customer:

[0188] First, construct a multi-dimensional preference tensor P, whose dimensions include [user, location type, time, context].

[0189] Then, a tensor decomposition method based on graph neural network is applied. The tensor P is represented as a heterogeneous graph G_p, and a relation-aware graph convolutional network (RGCN) is designed to learn the embedding of entities and relations, and the link prediction task is used for model training.

[0190] Next, extract personalized spatiotemporal preference features. Calculate the user-location preference score s(u,l) = f(e_u, e_l), estimate the time preference distribution p(t|u,l) and context preference distribution p(c|u,l,t), and construct a multimodal preference vector V_u = [s(u,l), p(t|u,l), p(c|u,l,t)].

[0191] Finally, the customer personalized spatiotemporal preference model set PM = {PM1, PM2, ..., PMn} is obtained.

[0192] Step S25: Spatiotemporal context-aware abnormal behavior detection

[0193] Obtain the customer behavior pattern P, the time-varying community structure C(t) output in step S22, and the personalized preference model PM output in step S24, and perform the following steps:

[0194] First, we design an anomaly detection model based on spatiotemporal graph convolutional neural network (ST-GCN). We construct a dynamic spatiotemporal graph G(t) = (V, E(t), A(t)), design a spatiotemporal attention layer to capture long-range dependencies, and use a graph differential pooling layer to learn hierarchical representations.

[0195] Then, the contrastive learning model is trained. The positive sample pairs (x_i, x_j) are constructed as normal behaviors in the same community, and the negative samples are selected using the hard negative sample mining strategy to optimize the contrastive loss function.

[0196] Finally, for each new observed behavior x_new, the similarity with the known pattern is calculated and the probability of the behavior in the current context is estimated, and whether it is abnormal is determined according to the set threshold.

[0197] Obtain an abnormal behavior detection model AD and a detected abnormal behavior set AB.

[0198] Step S3: Dynamic customer profile construction and personalized recommendation

[0199] Step S3 first obtains various types of customer-related data output by step S2, and applies the meta-path-guided graph attention mechanism and hierarchical contrastive learning method to perform multi-dimensional customer feature fusion to obtain the fused customer feature vector set F*. Then, based on F*, a recursive neural topic model and a dynamic factor graph model are used to construct a dynamic customer portrait to obtain a dynamic customer portrait set DP. Next, using the dynamic customer portrait DP, spectral graph theory and incremental singular value decomposition are applied to perform dynamic community detection, and continuous homology analysis of group evolution is used to obtain the dynamic customer group G(t) and evolution pattern E.

[0200] Step S31: Multi-dimensional customer feature fusion

[0201] Obtain various customer-related data output in step S2 and perform the following operations:

[0202] First, a heterogeneous information network (HIN) G = (V, E, φ, ψ) is constructed, where V is the node set, including entities such as customers, locations, and behaviors, E is the edge set, representing the relationship between entities, and φ and ψ are the type mapping functions of nodes and edges, respectively.

[0203] Then, the meta-path guided graph attention mechanism (MPGA) is applied. A set of key meta-paths P = {P1,P2, ..., Pk} is constructed, and for each meta-path Pi, the semantic specific embedding h_v^i of node v is calculated. An adaptive attention mechanism is designed to integrate multi-path information: h_v = Σ_i α_i * h_v^i.

[0204] Finally, a hierarchical contrastive learning method is used to optimize the node representation. Node-level contrast, meta-path-level contrast, and global structure-level contrast are performed to maximize the mutual information of relevant information.

[0205] The fused customer feature vector set F* = {f_1, f_2, ..., f*_n} is obtained.

[0206] Step S32: Dynamic customer profile generation

[0207] Get the fusion feature F* output in step S31 and build a dynamic profile for each customer:

[0208] First, a recursive neural topic model (RNTM) is designed. A variational recursive autoencoder is used to capture the temporal topic evolution, and for each time step t, a topic distribution θ_t and a topic word distribution β_t are generated. The historical information is integrated through the attention mechanism: h_t =Attention(h_1, h_2, ..., h_t-1).

[0209] Then, a dynamic factor graph model is applied to capture the time-varying dependencies between features. The factor graph G_f = (V_f, E_f) is constructed, and the particle MCMC method is used to infer the time-varying conditional probability distribution p(f_i|pa(f_i), t). The weights of the edges are dynamically adjusted through L1 regularization to update the factor graph structure.

[0210] Finally, the time-series topics and dynamic factor graphs are integrated to generate multi-granular customer portraits, including the macro-level topic evolution trajectory T, the meso-level key features and their time-varying distribution K, and the micro-level instantaneous feature vector I_t.

[0211] Get the dynamic customer portrait set DP = {DP_1, DP_2, ..., DP_n}.

[0212] Step S33: Customer group discovery and evolution analysis

[0213] Obtain the dynamic customer profile DP output in step S32 and perform the following steps:

[0214] First, construct a time-varying customer similarity graph G_s(t) = (V, E(t), W(t)). Use the multi-kernel learning method to calculate the similarity between customers W_ij(t) = sim(DP_i(t), DP_j(t)).

[0215] Then, spectral graph theory and incremental singular value decomposition are applied to dynamic community detection. The graph Laplacian matrix L(t) = D(t) - W(t) is calculated, and the incremental SVD decomposition of L(t) is maintained. The first k singular vectors are clustered using k-means++ to obtain the dynamic community C(t).

[0216] Next, we designed a population evolution analysis method based on persistent homology, constructed the community evolution cluster flow {F_t}, calculated the persistent homology group PH_*(F_t) and persistent barcode, and extracted key evolutionary events.

[0217] Finally, the dynamic customer group G(t) = {G_1(t), G_2(t), ..., G_k(t)} and the evolution pattern E are obtained.

[0218] Step S4: Based on the dynamic customer profile DP and group analysis results G(t), a dynamic heterogeneous graph neural network and multi-task reinforcement learning framework are designed to generate a personalized recommendation model RM and an interaction strategy set IS. Finally, to solve the cold start problem, a cross-domain meta-learning framework and a graph neural network distillation method are designed to obtain a cold start processing model CM and a cross-domain knowledge base CK.

[0219] Step S41: Personalized recommendation and interaction strategy generation

[0220] Obtain the dynamic customer profile DP outputted in step S32 and the group analysis result G(t) outputted in step S33, and perform the following steps:

[0221] First, a dynamic heterogeneous graph neural network (DH-GNN) is designed. A multi-relation graph G_r = (V, {E_r}) is constructed, a temporal attention mechanism is used to capture long-term and short-term interests, and a relation-specific graph convolution layer is designed.

[0222] Then, a multi-task reinforcement learning framework is applied to optimize the recommendation strategy. The state space S = {DP, G(t), context information} is constructed, and the action space A = {recommended items, interaction methods, recommendation timing} is designed. Using hierarchical reinforcement learning, the top-level strategy selects the recommendation type, and the bottom-level strategy selects the specific items.

[0223] Next, causal inference is integrated to improve the robustness of recommendations. A temporal causal graph G_c(t) is constructed, dual machine learning is used to estimate treatment effects, and counterfactual reasoning is applied to optimize recommendation decisions.

[0224] Finally, we obtain the personalized recommendation model RM and interaction strategy set IS = {IS_1, IS_2, ..., IS_n}.

[0225] Step S42: Cross-domain knowledge transfer and cold start problem solving

[0226] Obtain the output results of steps S31 to S41, and perform the following steps to solve the cold start problem for new users or new scenarios:

[0227] First, a cross-domain meta-learning framework (CMF) is designed. Each user is regarded as a task, and a meta-task set T ={T_1, T_2, ..., T_m} is constructed. Domain adversarial training is used to align the feature distributions of the source domain and the target domain, and a fast adaptation mechanism based on the prototype network is designed.

[0228] Then, apply graph neural network distillation for knowledge transfer. Train the teacher GNN model M_teacher in the source domain, design a lightweight student model M_student, and use the attention mechanism to selectively transfer knowledge. Optimize the distillation loss: L = λ_1L_task + λ_2 L_distill + λ_3 L_structure.

[0229] Finally, few-shot learning and active learning strategies are integrated. A matching network is used to learn few-shot user representations, and a sampling strategy based on information entropy and representation uncertainty is designed. Dynamically update user representations: f_u new = α * f_u old +(1-α) * f_u update .

[0230] Get the cold start processing model CM and the cross-domain knowledge base CK.

[0231] In another embodiment of the present application, embodiment 2 is provided as follows:

[0232] Step S1, obtain the original spatiotemporal big data from multiple heterogeneous data sources, use the Transformer-based feature extraction method to preliminarily process the data, and obtain an enhanced feature data set; then, use the dynamic data source classification method to classify the enhanced feature data set to obtain a classified data set; then, apply adaptive spatiotemporal standardization transformation to the classified data set to obtain a standardized spatiotemporal data set; then, use the graph theory-based data quality assessment algorithm to perform quality scoring on the standardized data to obtain a quality-scored data set and a data set to be repaired; then, use the Wasserstein distance-based anomaly detection method to clean the quality-scored data to obtain a cleaned data set and an outlier set; finally, use the graph neural network-based tensor completion algorithm to complete and enhance the cleaned data to obtain a final high-quality spatiotemporal data set;

[0233] Step S2: First, obtain the high-quality spatiotemporal data set D* output by step S1, apply the adaptive multi-scale singular spectrum analysis method and the topological data analysis method to extract the spatiotemporal features, and obtain the spatiotemporal feature matrix F; then, based on the spatiotemporal feature matrix F, use the graph attention diffusion network to build a customer behavior model, and obtain the customer behavior pattern set P and the time-varying community structure C(t); then, apply the trajectory segmentation algorithm based on the attention mechanism and the hierarchical topic model to the original trajectory data to extract the semantic trajectory set ST and the frequent spatiotemporal pattern set M; then, use the multi-dimensional preference tensor and the graph neural network to build a personalized spatiotemporal preference model, and obtain the customer personalized spatiotemporal preference model set PM; finally, design an anomaly detection model based on the spatiotemporal graph convolutional neural network, and obtain the abnormal behavior detection model AD and the detected abnormal behavior set AB;

[0234] Step S21, obtaining a high-quality spatiotemporal data set, applying an adaptive multi-scale singular spectrum analysis method and a dynamic manifold learning method to sequentially perform feature extraction on the spatiotemporal sequence of each customer, obtaining and combining a time feature vector and a space feature vector to form a spatiotemporal feature matrix;

[0235] Step S22, obtaining a spatiotemporal feature matrix and constructing a spatiotemporal behavior graph; using a graph attention diffusion network, capturing different types of spatiotemporal dependencies through a multi-head attention mechanism, using a diffusion convolutional layer to perform multi-scale aggregation of spatiotemporal information, and learning customer representations through end-to-end training; for the learned customer representations, constructing a time-varying modularity matrix according to a dynamic community detection algorithm, solving an optimization problem to maximize the modularity, clustering the customer representations using a spectral clustering method, and obtaining a set of customer behavior patterns and a time-varying community structure; wherein the vertices of the spatiotemporal behavior graph represent customers, and the edges represent the spatiotemporal behavior similarities between customers;

[0236] Step S23, obtaining original trajectory data, and performing spatiotemporal semantic trajectory extraction on the trajectory of each customer;

[0237] First, a trajectory segmentation algorithm based on the attention mechanism is applied; a self-attention matrix is ​​constructed, a multi-scale convolutional neural network is used to extract local features, and the self-attention and local features are combined to identify semantic stay points and moving segments;

[0238] Then, a hierarchical topic model is applied to the identified stop points; a word-document-corpus hierarchy is constructed, and the topic distribution is learned using variational inference methods to extract multi-level semantic labels;

[0239] Then, the temporal association rule mining algorithm is used to extract the mobility pattern; the spatiotemporal itemsets are constructed, and the pattern growth algorithm based on the minimum description length principle is designed;

[0240] Finally, we obtain a set of semantic trajectories and a set of frequent spatiotemporal patterns;

[0241] Step S24, obtaining the semantic trajectory and frequent spatiotemporal patterns outputted in step S23, and performing personalized spatiotemporal preference modeling for each customer;

[0242] First, a multi-dimensional preference tensor is constructed, with dimensions including user, location type, time and context;

[0243] Then, a tensor decomposition method based on graph neural network is applied; the tensor is represented as a heterogeneous graph, a relationship-aware graph convolutional network is designed to learn the embedding of entities and relations, and the link prediction task is used for model training;

[0244] Next, extract personalized spatiotemporal preference features; calculate user-location preference scores, estimate time preference distribution and context preference distribution, and construct a multimodal preference vector;

[0245] Finally, a set of customer personalized spatiotemporal preference models is obtained;

[0246] Step S25, obtaining the customer behavior pattern and time-varying community structure outputted in step S22 and the personalized preference model outputted in step S24, and performing spatiotemporal context-aware abnormal behavior detection;

[0247] First, we design an anomaly detection model based on spatiotemporal graph convolutional neural network. We build a dynamic spatiotemporal graph, design a spatiotemporal attention layer to capture long-range dependencies, and use a graph differential pooling layer to learn hierarchical representations.

[0248] Then, the contrastive learning model is trained; positive sample pairs are constructed as normal behaviors in the same community, negative samples are selected using the hard negative sample mining strategy, and the contrastive loss function is optimized;

[0249] Finally, for each new observed behavior, the similarity with the known pattern is calculated and the probability of the behavior in the current context is estimated, and whether it is abnormal is determined according to the set threshold;

[0250] Obtain an abnormal behavior detection model and a set of detected abnormal behaviors.

[0251] Step S3: First, various types of customer-related data output by step S2 are obtained, and the meta-path-guided graph attention mechanism and hierarchical contrastive learning method are used to fuse multi-dimensional customer features to obtain the fused customer feature vector set F*; then, based on the customer feature vector set F*, a recursive neural topic model and a dynamic factor graph model are used to construct a dynamic customer portrait to obtain a dynamic customer portrait set DP; then, using the dynamic customer portrait DP, spectral graph theory and incremental singular value decomposition are used to perform dynamic community detection, and continuous homology analysis of group evolution is used to obtain a dynamic customer group G(t) and an evolution pattern E; then, based on the dynamic customer portrait DP and the group analysis result G(t), a dynamic heterogeneous graph neural network and a multi-task reinforcement learning framework are designed to generate a personalized recommendation model RM and an interaction strategy set IS; finally, in order to solve the cold start problem, a cross-domain meta-learning framework and a graph neural network distillation method are designed to obtain a cold start processing model CM and a cross-domain knowledge base CK.

[0252] Step S31, obtaining various types of customer-related data outputted in step S2, and performing multi-dimensional customer feature fusion;

[0253] First, a heterogeneous information network is constructed, which includes entities such as customers, locations, behaviors, and the relationships between entities;

[0254] Then, a meta-path-guided graph attention mechanism is applied; a set of key meta-paths is constructed, and for each meta-path, the semantic-specific embedding of nodes is calculated; an adaptive attention mechanism is designed to integrate multi-path information;

[0255] Finally, a hierarchical contrastive learning method is used to optimize node representation; node-level comparison, meta-path-level comparison, and global structure-level comparison are performed to maximize the mutual information of relevant information;

[0256] Obtain the fused customer feature vector set;

[0257] Step S32: Obtain the fusion features outputted in step S31 and construct a dynamic profile for each customer;

[0258] First, we design a recursive neural topic model. We use a variational recursive autoencoder to capture the temporal topic evolution and generate topic distribution and topic word distribution for each time step. We also integrate historical information through the attention mechanism.

[0259] Then, a dynamic factor graph model is applied to capture the time-varying dependencies between features. A factor graph is constructed and the particle MCMC method is used to infer the time-varying conditional probability distribution. The weights of the edges are dynamically adjusted through L1 regularization to update the factor graph structure.

[0260] Finally, the time-series topics and dynamic factor graphs are integrated to generate multi-granular customer portraits, including the evolution trajectory of the macro-level topics, the key features and their time-varying distribution at the meso-level, and the instantaneous feature vectors at the micro-level. A dynamic customer portrait set is obtained.

[0261] Step S33, obtaining the dynamic customer portrait outputted in step S32, and performing customer group discovery and evolution analysis;

[0262] First, a time-varying customer similarity graph is constructed; the similarity between customers is calculated using a multi-kernel learning method;

[0263] Then, spectral graph theory and incremental singular value decomposition are applied to perform dynamic community detection; the graph Laplacian matrix is ​​calculated and its incremental SVD decomposition is maintained; k-means++ is used to cluster the first k singular vectors to obtain dynamic communities;

[0264] Next, we designed a population evolution analysis method based on dynamic manifold learning; constructed a community evolution manifold, applied the dynamic manifold learning algorithm to capture the continuous changes in population structure, and extracted key evolution events;

[0265] Finally, dynamic customer groups and evolution patterns are obtained.

[0266] Obtain dynamic customer portraits and group analysis results, and perform personalized recommendations and interactive strategy generation;

[0267] First, a dynamic heterogeneous graph neural network is designed; a multi-relationship graph is constructed, a temporal attention mechanism is used to capture long-term and short-term interests, and a relationship-specific graph convolution layer is designed; then, a multi-task reinforcement learning framework is applied to optimize the recommendation strategy; a state space is constructed including dynamic customer portraits, group structure, and contextual information, and an action space is designed including recommended items, interaction methods, and recommendation timing; hierarchical reinforcement learning is used, with the top-level strategy selecting the recommendation type and the bottom-level strategy selecting the specific item; then, causal inference is integrated to improve the robustness of the recommendation; a temporal causal graph is constructed, dual machine learning is used to estimate the treatment effect, and counterfactual reasoning is applied to optimize the recommendation decision; finally, a personalized recommendation model and a set of interaction strategies are obtained;

[0268] Based on all the above output results, a meta-task set is constructed as the basis for cross-domain meta-learning; a domain adversarial training mechanism and a rapid adaptation mechanism based on the prototype network are designed to realize knowledge transfer between the source domain and the target domain; a graph neural network is trained in the source domain as a teacher model, and a lightweight student model is designed for rapid deployment in the target domain; effective knowledge transfer is achieved through attention mechanism and comprehensive loss function optimization; a matching network and a sampling strategy based on information entropy are designed to achieve rapid user representation learning and dynamic updating; a meta-learning framework, knowledge distillation method, few-shot learning and active learning strategy are integrated to build a complete cold start processing model; finally, a cold start processing model and a structured cross-domain knowledge base are output to provide solutions for personalized recommendations for new users or new scenarios.

[0269] It should be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction; in order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A customer portrait key data mining method based on spatiotemporal big data, characterized in that: The steps include: Step S1, obtain multi-source heterogeneous spatiotemporal big data, and obtain high-quality spatiotemporal data sets through Transformer feature extraction, dynamic classification, adaptive normalization, graph theory quality assessment, Wasserstein distance anomaly detection and graph neural network tensor completion; Step S2, based on the spatiotemporal data set, extracting the spatiotemporal feature matrix through multi-scale singular spectrum analysis and topological analysis; Use graph attention diffusion network to build customer behavior model, obtain behavior pattern and time-varying community structure; apply attention mechanism trajectory segmentation and hierarchical topic model to extract semantic trajectory and spatiotemporal pattern; combine multi-dimensional preference tensor and graph neural network to build personalized spatiotemporal preference model; finally, design spatiotemporal graph convolutional neural network, generate anomaly detection model and abnormal behavior set, and build output data set; The output data set includes spatiotemporal feature matrix F, customer behavior pattern P, time-varying community structure C(t), semantic trajectory ST, spatiotemporal pattern M, personalized spatiotemporal preference model PM, anomaly detection model AD and abnormal behavior set AB; Step S3: Based on the output data set, the customer features are integrated to obtain a vector set through the meta-path-guided graph attention mechanism and hierarchical contrastive learning; based on the vector set, a recursive neural topic model and a dynamic factor graph are used to construct a customer portrait set; spectral graph theory and dynamic manifold learning are used to obtain dynamic customer groups and evolution patterns; Output dynamic customer portraits and group analysis results; Step S4: Based on the dynamic customer portrait and group analysis results, the constructed dynamic heterogeneous graph neural network and multi-task reinforcement learning framework are used to generate a recommendation model and strategy set; The step S1 is specifically as follows: Step S11, obtaining different types of original spatiotemporal big data from multiple heterogeneous data sources, and for each data source, using a pre-trained Transformer model to perform feature extraction, including: serializing the spatiotemporal data into a format suitable for Transformer processing to obtain an input sequence; inputting the input sequence into the Transformer model, and processing it through a multi-layer self-attention mechanism and a feedforward neural network to obtain an enhanced feature sequence; aggregating the enhanced feature sequence to obtain an overall feature representation of the data source, and obtaining and outputting an enhanced feature data set, including original data and its corresponding enhanced features, and the original data includes original trajectory data; Step S12, obtaining an enhanced feature data set, first extracting feature vectors for each enhanced feature data set, constructing a dynamic similarity matrix, calculating its Laplace matrix, solving the generalized eigenvalue problem to obtain time-varying feature vectors, clustering based on the time-varying feature vectors, completing the process of classifying the feature vectors using a dynamic spectral clustering algorithm, and obtaining a classified data set; the time-varying feature vectors include time granularity, spatial precision, and data type; Step S13, obtaining the classified data set, performing time standardization, detecting the timestamp format and uniformly converting it into a Unix timestamp, applying wavelet transform to decompose the time series, extracting multi-scale time features, resampling the time series using an adaptive time window method, and then performing spatial standardization, converting all spatial coordinates into a unified geographic coordinate system, for non-point data, applying a curvature-based shape descriptor to extract feature points, and using the Riemann manifold learning method to perform nonlinear normalization on the spatial data, and then completing the spatiotemporal standardization of each classified data set to obtain a standardized spatiotemporal data set; Step S14: for the standardized spatiotemporal data set, call to build a spatiotemporal hypergraph, apply the hypergraph theory, calculate the hypergraph Laplace eigenvalue of each vertex, call the threshold of the quality score, and mark the data points below the threshold as the data set to be repaired; obtain the data set with quality score and the data set to be repaired, and perform data quality assessment on each data set; Among them, the vertices of the spatiotemporal hypergraph represent data points, and the hyperedges represent spatiotemporal relationships; the quality score is the result of the processing of the eigenvalue by the nonlinear mapping function based on deep learning; Step S15, obtaining the quality-scored data set, constructing a spatiotemporal feature vector, applying an anomaly detection algorithm based on Wasserstein distance, and estimating the empirical distribution of the data; for each data point, calculating its Wasserstein distance with the empirical distribution, and dynamically setting the anomaly threshold using extreme value theory; marking the data point with a distance greater than the threshold as an outlier and adding it to the outlier set; obtaining the cleaned data set and the outlier set; Among them, the spatiotemporal feature vector includes time feature, space feature and attribute feature; Step S16: obtain the cleaned data set, the data set to be repaired, and the set of outliers, construct a spatiotemporal-attribute fourth-order tensor, and introduce an additional modal dimension to represent data reliability. A tensor completion algorithm based on a graph neural network is used to represent the tensor as a heterogeneous graph, and a graph attention network layer is constructed to learn node representations. A graph autoencoder architecture is used to reconstruct missing values. For areas with insufficient spatiotemporal density, a conditional generation model based on Gaussian Copula is used to generate new data points. The data is completed and enhanced to obtain a high-quality spatiotemporal data set after completion. Step S3 is specifically as follows: Step S31, obtain the customer-related data outputted in step S2, construct a heterogeneous information network; construct and apply a set of key meta-paths, and use a graph attention mechanism to calculate the semantic-specific embedding of nodes; design an adaptive attention mechanism, integrate multi-path information to obtain a comprehensive embedding of nodes; further optimize node representation through a hierarchical contrastive learning task; and finally output a fused customer feature vector set; Step S32, obtain a fused feature vector set, apply a recursive neural topic model to capture the evolution of the topic over time; construct a dynamic factor graph model to capture the time-varying dependencies between features; integrate the outputs of the two models to generate a multi-granular customer portrait, including the topic evolution trajectory at the macro level, the time-varying distribution of key features at the meso level, and the instantaneous feature vector at the micro level; output a dynamic customer portrait set; Step S33, obtain a dynamic customer portrait set and construct a time-varying customer similarity graph; apply spectral graph theory and incremental singular value decomposition to perform dynamic community detection; construct a community evolution manifold and use a dynamic manifold learning algorithm to capture the continuous changes in group structure; analyze the manifold trajectory and extract key events in the customer group evolution process; finally output a dynamic customer group structure and evolution pattern set; Step S4 is specifically as follows: Step S41, obtaining dynamic customer portraits and group analysis results; constructing a multi-relationship graph and a dynamic heterogeneous graph neural network to capture the user's long-term and short-term interest changes and the specific semantics of heterogeneous relationships; Step S42: Based on the dynamic heterogeneous graph neural network, a hierarchical reinforcement learning framework is constructed to construct a state space including customer portraits, group structure and context information, and an action space including recommended items, interaction methods and recommendation timing; a personalized recommendation strategy is generated by training a reinforcement learning agent; Step S43, construct a temporal causal graph and apply dual machine learning methods, and use counterfactual reasoning technology to optimize recommendation decisions; integrate dynamic heterogeneous graph neural networks, hierarchical reinforcement learning frameworks, personalized recommendation strategies, temporal causal graphs, dual machine learning methods and counterfactual reasoning technology to form a complete personalized recommendation model and interactive strategy set.

2. The method for mining key data of customer portraits based on spatiotemporal big data according to claim 1, characterized in that: The step S2 is specifically as follows: Step S21, obtaining a high-quality spatiotemporal data set, applying an adaptive multi-scale singular spectrum analysis method and a dynamic manifold learning method to sequentially perform feature extraction on the spatiotemporal sequence of each customer, obtaining and combining a time feature vector and a space feature vector to form a spatiotemporal feature matrix; Step S22: Obtain a spatiotemporal feature matrix and construct a spatiotemporal behavior graph; apply a graph attention diffusion network to capture different types of spatiotemporal dependencies and perform multi-scale aggregation to learn customer representation; Perform dynamic community detection on the learned customer representations, build and optimize the time-varying modularity matrix; use spectral clustering to cluster the optimized customer representations; and finally output a set of customer behavior patterns and a time-varying community structure. Step S23, obtaining the original trajectory data, applying the trajectory segmentation algorithm based on the attention mechanism to identify the semantic stay points and movement segments; applying the hierarchical topic model to the stay points to extract the multi-level semantic labels; using the temporal association rule mining algorithm to extract the frequent movement patterns for the movement segments; integrating the processed stay point and movement segment information to generate a semantic trajectory set; summarizing the analysis results of all trajectories to form a frequent spatiotemporal pattern set; Step S24, obtaining semantic trajectories and frequent spatiotemporal patterns, and constructing a multi-dimensional preference tensor; The preference tensor is transformed into a heterogeneous graph and a relation-aware graph convolutional network is applied to learn the embedding representation of entities and relations. Based on the learned embedding, the user-location preference score is calculated and the time and context preference distribution is estimated. The time and context preference distribution and user-location preference score are integrated to construct a multimodal preference vector and generate a personalized spatiotemporal preference model for each user. Finally Output a set of customer personalized spatiotemporal preference models; Step S25, obtaining customer behavior patterns and time-varying community structures, as well as personalized preference models; Construct a dynamic spatiotemporal graph and apply the spatiotemporal attention mechanism and graph differential pooling to learn the hierarchical representation of user behavior; Based on the hierarchical representation, train the spatiotemporal graph convolutional neural network as the basic model for anomaly detection; Enhance the model's discriminative ability through comparative learning methods; For newly observed user behaviors, anomaly judgment is performed by combining the learned behavior patterns and personalized preference models; finally, an abnormal behavior detection model and a set of detected abnormal behaviors are output.

3. The method for mining key data of customer portraits based on spatiotemporal big data as claimed in claim 2, characterized in that: In step S21, the process of applying the adaptive multi-scale singular spectrum analysis method includes: constructing a multi-scale trajectory matrix, which includes different time scales; performing singular value decomposition on each trajectory matrix to obtain feature triplets; adaptively selecting the best decomposition scale using the information entropy criterion; extracting time feature vectors, including periodicity, trend, mutation and entropy; The process of extracting spatial features using the dynamic manifold learning method is as follows: using local linear embedding or isometric mapping algorithms to reduce the high-dimensional spatiotemporal data to a low-dimensional manifold space, obtaining a low-dimensional embedding of the spatiotemporal data, and capturing the dynamic changes of the data structure in the low-dimensional manifold space by sliding the time window; calculating the geometric and topological features on the manifold; extracting spatial feature vectors; the geometric and topological features include curvature, geodesic distance, and spectral features; the spatial feature vectors include local connectivity, global structure, and complexity.

4. The method for mining key data of customer portraits based on spatiotemporal big data as claimed in claim 2, characterized in that: The step S12 is specifically as follows: Step S121, receiving an enhanced feature data set, constructing a distance matrix using Euclidean distance calculation; generating a similarity matrix based on the distance matrix by applying a Gaussian kernel function; calculating a normalized Laplace matrix based on the similarity matrix; solving a generalized eigenvalue problem to obtain eigenvalues ​​and eigenvectors; selecting an eigenvector corresponding to the minimum non-zero eigenvalue to form a time-varying feature matrix; Step S122, obtaining the time-varying feature matrix of step S121, setting the maximum number of clusters; executing the k-means algorithm for each possible number of clusters to obtain clustering results; calculating the silhouette coefficient and the Calinski-Harabasz index based on the clustering results; drawing curves of the above two indicators changing with the number of clusters; analyzing the curves using the inflection point method to determine the optimal number of clusters; Step S123, receiving the time-varying feature matrix of step S121 and the optimal number of clusters of step S122; calculating the cosine similarity between the data points in the matrix; constructing the adjacency matrix using the k-nearest neighbors algorithm; calculating the graph Laplacian matrix based on the adjacency matrix; solving the eigenvector of the graph Laplacian matrix; Perform l2 normalization on the feature vector; use the k-means++ algorithm to perform clustering on the normalized feature vector to obtain the clustering result; Step S124, obtaining the clustering result of step S123, receiving a new data point; extracting a time-varying feature vector of the new data point; calculating the distance between the vector and the existing cluster center; and assigning the new data point to the nearest cluster; Update the center position of the affected cluster; determine whether the cluster structure change exceeds a preset threshold, and if so, trigger the re-execution of steps S122 and S123; Step S125, receiving the clustering results and original data set of step S123 or S124; calculating the Davies-Bouldin index and Dunn index as internal evaluation indicators; if there is labeled data, calculating the adjusted Rand index and normalized mutual information as external evaluation indicators; combining internal and external indicators to generate an overall quality indicator; comparing the overall quality indicator with a preset threshold, if it is lower than the threshold, adjusting the kernel function parameters in step S121, and re-executing steps S121 to S124; outputting the final optimized clustering results and quality indicators.

5. The method for mining key data of customer portraits based on spatiotemporal big data as claimed in claim 2, characterized in that: The step S13 is specifically as follows: Step S131, obtaining original spatiotemporal data, detecting the timestamp format therein; converting the detected timestamps uniformly into Unix timestamp format; applying wavelet transform to decompose the converted time series; extracting multi-scale time features from the decomposition results; The extracted time series are resampled using the adaptive time window method to obtain standardized time features; Step S132, receiving the original spatial data, converting all the spatial coordinates therein into the WGS84 geographic coordinate system; for non-point type spatial data, extracting feature points using a curvature-based shape descriptor algorithm; inputting the extracted feature points into a Riemann manifold learning model, performing a nonlinear normalization operation, and outputting standardized spatial features; Step S133, obtaining the standardized time features of step S131 and the standardized space features of step S132; constructing a spatiotemporal tensor, integrating the time and space features into the same data structure; applying the Tucker decomposition method to the constructed spatiotemporal tensor to perform feature fusion; introducing an attention mechanism, dynamically adjusting the weights of the time and space features, and outputting the fused spatiotemporal features; Step S134, receiving the fused spatiotemporal features of step S133, and calculating the local and global statistical features thereof; Based on the calculated statistical features, dynamically adjust the scale parameters of spatiotemporal analysis; apply the adjusted scale parameters to multi-scale analysis methods to capture spatiotemporal patterns at different scales and output spatiotemporal features at adaptive scales; Step S135, obtaining the original spatiotemporal data and the adaptive scale spatiotemporal features of step S134; Calculate the information entropy of the data before and after standardization to evaluate the degree of information retention; use the Kolmogorov-Smirnov test method to test the effect of standardization; compare the test result with the preset threshold. If it does not meet the requirements, adjust the following parameters of steps S131 to S134: timestamp format detection parameters, wavelet transform parameters, adaptive time window parameters, spatial coordinate transformation parameters, shape descriptor extraction parameters, Riemann manifold learning parameters, Tucker decomposition parameters, attention mechanism parameters and multi-scale analysis parameters, and re-execute steps S131 to S134 until the standardization quality requirements are met and the final high-quality standardized spatiotemporal data is output.

6. The method for mining key data of customer portraits based on spatiotemporal big data as claimed in claim 2, characterized in that: Step S15 is specifically as follows: Step S151, receiving the quality-scored data set, constructing a multi-dimensional spatiotemporal feature vector; applying a kernel density estimation method to the constructed feature vector to obtain an empirical distribution of the data; and setting an initial anomaly detection threshold; Step S152, obtaining the empirical distribution of step S151, calculating the Wasserstein distance between each data point and the empirical distribution; inputting the calculated Wasserstein distance sequence into the extreme value theory model to estimate the tail behavior of the distance distribution; dynamically adjusting the anomaly detection threshold based on the tail behavior estimation result; Step S153, receiving the Wasserstein distance sequence and dynamic threshold of step S152, comparing the distance value of each data point with the threshold; marking the data points whose distance is greater than the threshold as potential outliers; applying the local outlier factor algorithm to the marked potential outliers, performing secondary verification, and obtaining the final outlier set; Step S154, obtaining the abnormal value set of step S153, and constructing an abnormal pattern library; for each abnormal value, extracting its spatiotemporal context information; using the extracted context information, training an abnormal prediction model based on a long short-term memory network; applying the trained model to the real-time data stream to perform online anomaly detection; Step S155, receiving the abnormality prediction result of step S154, and comparing it with the actually observed abnormality; Calculate the prediction accuracy, recall rate and F1 score; Based on the calculated performance indicators, use the Bayesian optimization method to adjust the parameters related to Wasserstein distance calculation and anomaly detection; The optimized parameters are fed back to steps S151-S154 to achieve continuous optimization of the anomaly detection process.

7. A customer portrait key data mining system based on spatiotemporal big data, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the customer portrait key data mining method based on spatiotemporal big data as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Risk management method and system for cross-border e-commerce transaction behavior

    CN118469715A

  • Business recommendation method, device and system based on big data and customer portrait

    CN118585703A