User label automatic construction method and system based on multi-source heterogeneous data

Abstract

The application provides a user label automatic construction method and system based on multi-source heterogeneous data, relates to the technical field of data mining, and comprises the following steps: obtaining multi-source behavior data, adopting a bidirectional attention mechanism for feature fusion, dynamically adjusting a clustering center based on time sequence drift, and utilizing a label propagation algorithm considering node time sequence evolution characteristics to construct a multi-level label structure tree. The application can effectively integrate heterogeneous data source information, adaptively capture user behavior changes, improve label accuracy and timeliness, and provide support for precise marketing and personalized recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data mining, and in particular to a user label automatic construction method and system based on multi-source heterogeneous data. BACKGROUND

[0002] With the rapid development of the Internet and mobile Internet, users generate massive behavior data in various platforms and applications, including browsing history, purchase records, social interactions, and other forms. These multi-source heterogeneous data contain rich user interest and feature information, which is of great value for precision marketing, personalized recommendation, and user portrait construction. User labels are a high-level summary of user characteristics, interests, and behavior patterns, and have become a key means of understanding user needs and providing personalized services. Traditional user label construction methods mainly rely on manual rule definition and single data source analysis. With the increasing complexity of data size and complexity, the demand for automatic construction of user labels is increasingly urgent.

[0003] The existing user label automatic construction technology has the following shortcomings:

[0004] The existing technology lacks an effective data fusion mechanism when dealing with multi-source heterogeneous data, often using simple data splicing or weighted average methods, ignoring the internal relevance and complementarity between different data sources, resulting in limited feature expression ability after fusion, and failing to fully mine the multi-dimensional features of users.

[0005] Existing user clustering analysis methods mostly use static clustering algorithms, which fail to effectively cope with the dynamic characteristics of user interests and behaviors over time, resulting in clustering results that cannot timely reflect the evolution trend of user groups, reducing the timeliness and accuracy of labels. SUMMARY

[0006] The embodiments of the present application provide a user label automatic construction method and system based on multi-source heterogeneous data, which can solve the problems in the prior art.

[0007] In a first aspect of the embodiments of the present application, a user label automatic construction method based on multi-source heterogeneous data is provided, comprising:

[0008] Obtaining behavior data generated by a user in multiple different data sources, dividing the behavior data according to a preset time window, and performing standardization processing on the divided behavior data to obtain standardized behavior data;

[0009] Based on the standardized behavior data, an association weight matrix between different data sources is calculated, and a bidirectional attention mechanism is used to perform adaptive feature fusion on the behavior data of different data sources according to the association weight matrix to generate a fusion feature vector;

[0010] According to the fusion feature vector, user clustering analysis is performed, and a clustering center is dynamically adjusted based on a time sequence drift degree of a user group, to obtain a user clustering result;

[0011] A label propagation network is constructed by taking users as nodes and constructing a graph structure, similarity between users is taken as a weight of an edge, a label propagation probability considering a time sequence evolution feature of a node is iteratively calculated, a label is assigned to a user based on a first preset threshold, and a label assignment result is obtained;

[0012] Based on the label assignment result, a corresponding label set is generated for each user, and a multi-level label structure tree is constructed according to a hierarchical relationship and a correlation degree between labels.

[0013] Based on the standardized behavior data, an association weight matrix between different data sources is calculated, and a bidirectional attention mechanism is used to adaptively fuse behavior data of different data sources according to the association weight matrix, to generate a fusion feature vector, including:

[0014] Based on the standardized behavior data, a time sequence correlation feature and a spatial distribution feature of each data source are calculated, and the time sequence correlation feature and the spatial distribution feature are combined to obtain a data source feature matrix;

[0015] A time sequence analysis is performed on the data source feature matrix, and a dynamic change feature between data sources is extracted;

[0016] Based on the dynamic change feature, a propagation time delay and a propagation distance between data sources are calculated, and a structure association feature of the data sources is extracted according to the propagation time delay and the propagation distance;

[0017] The dynamic change feature and the structure association feature are fused to construct an initial association weight matrix;

[0018] Based on the initial association weight matrix, an attention score between data sources is calculated, a global weight matrix is obtained by forward calculation to extract a global dependency relationship between data sources, a local weight matrix is obtained by reverse calculation to extract a local dependency relationship between data sources, and the global weight matrix and the local weight matrix are adaptively weighted and combined to obtain an optimized association weight matrix;

[0019] Based on the optimized association weight matrix, a feature of behavior data of each data source is reconstructed to obtain a reconstructed feature, and a fusion feature vector is generated by performing multi-dimensional feature interaction calculation on the reconstructed feature.

[0020] calculating initial attention scores between data sources based on the initial correlation weight matrix, and sorting the initial attention scores according to the time sequence of the data sources to construct a time sequence attention sequence;

[0021] calculating initial attention scores between data sources based on the initial correlation weight matrix, and sorting the initial attention scores according to the time sequence of the data sources to construct a time sequence attention sequence;

[0022] performing sliding window segmentation on the time sequence attention sequence to obtain a plurality of attention sub-sequences, and respectively calculating distance matrices and similarity matrices between data sources in the attention sub-sequences;

[0023] calculating global dependency relationships between data sources based on the distance matrices to generate a global weight matrix, the global weight matrix reflecting long-range dependency features between data sources under different time windows;

[0024] calculating local dependency relationships between data sources based on the similarity matrices to generate a local weight matrix, the local weight matrix reflecting short-range dependency features between data sources within adjacent time windows;

[0025] setting adaptive weight coefficients for the global weight matrix and the local weight matrix, wherein the adaptive weight coefficients are dynamically adjusted according to the time sequence correlation between data sources, and performing matrix splicing on the weighted global weight matrix and the local weight matrix to obtain an optimized correlation weight matrix.

[0026] performing clustering analysis on users according to the fusion feature vector, and dynamically adjusting the clustering centers based on the time sequence drift degree of the user groups to obtain user clustering results, including:

[0027] calculating a feature distance matrix between users based on the fusion feature vector;

[0028] performing density peak value analysis on the feature distance matrix to identify user points with local maximum density as initial clustering centers;

[0029] calculating Mahalanobis distances between each user and the initial clustering centers;

[0030] constructing a user similarity matrix based on the Mahalanobis distances, dividing the user similarity matrix into a plurality of time segments along the time dimension, and calculating the evolution features of the user groups within each time segment;

[0031] According to the evolution feature, a stability coefficient of the user group is calculated, and the initial clustering center is dynamically adjusted based on the stability coefficient, wherein when the stability coefficient is lower than a second preset threshold, the clustering centers with a distance less than a clustering radius are merged, and when the stability coefficient is higher than the second preset threshold, the clustering centers with a data density greater than an average density of the clustering centers are split;

[0032] The adjusted clustering center is iteratively optimized, the optimal clustering number is determined by calculating the time sequence drift degree of the user group, and a user clustering result considering time sequence evolution is obtained.

[0033] A label propagation network is constructed by taking users as nodes and constructing a graph structure, the similarity between users is taken as the weight of the edge, the label propagation probability considering the time sequence evolution feature of the node is iteratively calculated, the user is labeled based on a first preset threshold, and a label assignment result is obtained, including:

[0034] A graph structure is constructed by taking users as nodes, and the multi-dimensional similarity features between nodes are calculated, and the multi-dimensional similarity features are normalized to obtain the weight value of the edge;

[0035] The graph structure is divided into multiple time windows along the time dimension, and the feature vector of the node is calculated for each time window;

[0036] The degree centrality, betweenness centrality and clustering coefficient of the node are calculated based on the feature vector, and a node importance score matrix is constructed;

[0037] The node importance score matrix is analyzed by using a spectral clustering method to determine a key node group, and the key node group is taken as a source node of label propagation;

[0038] The state transition probability of the node between adjacent time windows is calculated, a time sequence attenuation function is generated in combination with the activity of the node, the time sequence attenuation function and the weight value of the edge are adaptively weighted to obtain a time sequence weight matrix, and the label propagation probability between nodes is iteratively calculated based on the time sequence weight matrix;

[0039] A label propagation confidence matrix is constructed, a label candidate set is recorded to a label assignment result when the label propagation probability is greater than a first preset threshold and has time sequence consistency, and the label candidate set is optimized to obtain the label assignment result.

[0040] The state transition probability of the node between adjacent time windows is calculated, a time sequence attenuation function is generated in combination with the activity of the node, the time sequence attenuation function and the weight value of the edge are adaptively weighted to obtain a time sequence weight matrix, and the label propagation probability between nodes is iteratively calculated based on the time sequence weight matrix, including:

[0041] Obtaining multi-dimensional behavior characteristics of nodes in adjacent time windows, constructing a node state feature matrix, and calculating mutual information and conditional entropy between nodes based on the node state feature matrix;

[0042] Constructing a node behavior correlation network using the mutual information and the conditional entropy, extracting topological features and community features of nodes, and combining behavior intensity of nodes to generate a multi-dimensional node activity index;

[0043] Constructing a high-order Markov chain based on the node state feature matrix, calculating a transition probability matrix of the high-order Markov chain at different time scales, regarding a state transition edge with a probability value greater than a third preset threshold in the transition probability matrix as a long-range dependence relationship, and obtaining a state transition probability of nodes;

[0044] Constructing a time series decay function according to the state transition probability and the multi-dimensional node activity index, wherein a decay rate of the time series decay function is dynamically adjusted according to node activity, and the higher the activity, the smaller the decay rate;

[0045] Adaptively combining the time series decay function and a weight value of an edge, dynamically optimizing a combination parameter based on evolution trajectory features of nodes in a historical window, generating a weight matrix fusing time series dynamic characteristics, and iteratively calculating based on the weight matrix fusing time series dynamic characteristics to obtain a label propagation probability between nodes.

[0046] Based on the label assignment result, generating a corresponding label set for each user, and constructing a multi-level label structure tree according to a hierarchical relationship and a correlation degree between labels, including:

[0047] Based on the label assignment result, extracting a label set of each user, and calculating a co-occurrence frequency matrix between pairs of labels in the label set;

[0048] Extracting a semantic feature of the label set to construct a label semantic vector, calculating a usage frequency and a coverage range of a label based on the co-occurrence frequency matrix and the label semantic vector, and generating a label importance score;

[0049] According to the label semantic vector and the label importance score, dividing labels into different levels according to semantic concept granularity, assigning labels with a label importance score greater than a mean value of label importance scores to an upper layer node, and constructing an initial hierarchical tree of labels;

[0050] Calculating an information gain ratio and a semantic inclusion degree between labels at different levels, confirming a hierarchical affiliation relationship between two labels when the information gain ratio and the semantic inclusion degree of an upper layer label to a lower layer label satisfy a first preset condition, and optimizing the initial hierarchical tree based on the confirmed hierarchical affiliation relationship;

[0051] The co-occurrence frequency matrix and the label semantic vector are used to calculate the association strength between labels in the same level, a label pair satisfying a second preset condition is constructed into a horizontal association relationship, and a multi-level label structure tree is generated based on the hierarchical affiliation relationship and the horizontal association relationship.

[0052] In a second aspect, the embodiment of the present application provides a user label automatic construction system based on multi-source heterogeneous data, which comprises:

[0053] A first unit is configured to acquire behavior data generated by a user in multiple different data sources, split the behavior data according to a preset time window, and perform standardization processing on the split behavior data to obtain standardized behavior data.

[0054] A second unit is configured to calculate an association weight matrix between different data sources based on the standardized behavior data, perform adaptive feature fusion on the behavior data of different data sources according to the association weight matrix by using a bidirectional attention mechanism, and generate a fusion feature vector.

[0055] A third unit is configured to perform clustering analysis on the user according to the fusion feature vector, and dynamically adjust a clustering center based on a time sequence drift degree of a user group to obtain a user clustering result.

[0056] A fourth unit is configured to construct a label propagation network of a graph structure by taking the user as a node, take a similarity between users as a weight of an edge, iteratively calculate a label propagation probability considering a time sequence evolution feature of the node, perform label allocation on the user based on a first preset threshold, and obtain a label allocation result.

[0057] A fifth unit is configured to generate a corresponding label set for each user based on the label allocation result, and construct a multi-level label structure tree according to a hierarchical relationship and an association degree between labels.

[0058] In a third aspect, the embodiment of the present application provides an electronic device, which comprises:

[0059] A processor;

[0060] A memory for storing processor-executable instructions;

[0061] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0062] In a fourth aspect, the embodiment of the present application provides a computer-readable storage medium, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0063] The present application has the following beneficial effects:

[0064] The application can comprehensively capture the behavior characteristics of users on different platforms by adopting multi-source heterogeneous data fusion technology, effectively solves the one-sidedness problem of labels brought by single data source, and improves the comprehensiveness and accuracy of user portrait.

[0065] The application introduces a bidirectional attention mechanism and a dynamic clustering center adjustment strategy, can adaptively process the association relationship between different data sources and the time sequence drift of user groups, effectively enhances the timeliness and adaptability in the label construction process, and makes the generated user label better reflect the real-time interest change of the user.

[0066] The label propagation method based on the graph structure of the application fully considers the association influence and time sequence evolution characteristics between users, and by constructing a multi-level label structure tree, the generated user label system has rich semantic information and clear hierarchical relationship, which provides high-quality user feature description for subsequent precision marketing and personalized recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 The figure is a flowchart of the user label automatic construction method based on multi-source heterogeneous data of the embodiment of the application.

[0068] Figure 2 The figure is a precision comparison and analysis column chart of the dependency relationship calculation method of the embodiment of the application.

[0069] Figure 3 The figure is a precision comparison and analysis chart of different label propagation algorithms of the embodiment of the application. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical scheme and advantages of the embodiment of the application clearer, the technical scheme in the embodiment of the application will be described clearly and completely below in combination with the drawings in the embodiment of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0071] The technical scheme of the application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0072] Figure 1 The figure is a flowchart of the user label automatic construction method based on multi-source heterogeneous data of the embodiment of the application, as shown in Figure 1 The method comprises:

[0073] Obtaining behavior data generated by a user in multiple different data sources, dividing the behavior data according to a preset time window, and performing standardization processing on the divided behavior data to obtain standardized behavior data;

[0074] Based on the standardized behavior data, an association weight matrix between different data sources is calculated, and a bidirectional attention mechanism is used to adaptively fuse the behavior data of different data sources according to the association weight matrix to generate a fusion feature vector;

[0075] According to the fusion feature vector, a clustering analysis is performed on the user, and the clustering center is dynamically adjusted based on the time sequence drift degree of the user group to obtain a user clustering result;

[0076] A label propagation network of a graph structure is constructed by taking the user as a node, the similarity between users is taken as the weight of the edge, the label propagation probability considering the time sequence evolution characteristics of the node is iteratively calculated, and the user is assigned a label based on a first preset threshold to obtain a label assignment result;

[0077] Based on the label assignment result, a corresponding label set is generated for each user, and a multi-level label structure tree is constructed according to the hierarchical relationship and the association degree between labels.

[0078] In an optional implementation, based on the standardized behavior data, an association weight matrix between different data sources is calculated, and a bidirectional attention mechanism is used to adaptively fuse the behavior data of different data sources according to the association weight matrix to generate a fusion feature vector, including:

[0079] Based on the standardized behavior data, the time sequence correlation features and the spatial distribution features of each data source are calculated, the time sequence correlation features and the spatial distribution features are combined to obtain a data source feature matrix;

[0080] The data source feature matrix is subjected to time sequence analysis, and the dynamic change features between data sources are extracted;

[0081] Based on the dynamic change features, the propagation time delay and the propagation distance between data sources are calculated, and the structural association features of the data sources are extracted according to the propagation time delay and the propagation distance;

[0082] The dynamic change features and the structural association features are fused to construct an initial association weight matrix;

[0083] The attention score between the data sources is calculated based on the initial correlation weight matrix, a global weight matrix is obtained by extracting the global dependency relationship between the data sources through forward calculation, a local weight matrix is obtained by extracting the local dependency relationship between the data sources through reverse calculation, and the global weight matrix and the local weight matrix are adaptively weighted and combined to obtain an optimized correlation weight matrix;

[0084] The behavior data of each data source is reconstructed based on the optimized correlation weight matrix to obtain reconstructed features, and multi-dimensional feature interaction calculation is performed on the reconstructed features to generate a fusion feature vector.

[0085] The application provides an adaptive feature fusion method based on a bidirectional attention mechanism, which calculates the correlation weight matrix between different data sources and uses the bidirectional attention mechanism to adaptively fuse the behavior data of different data sources to generate a fusion feature vector.

[0086] In the implementation process, the time sequence correlation features and the spatial distribution features of each data source are calculated based on the standardized behavior data. The time sequence correlation features are extracted by analyzing the change rule of the behavior data in the time dimension. For example, for user click behavior data, the system divides a day into 24 time periods, counts the click frequency in each time period, and forms a time sequence vector [0.05, 0.03, 0.01, …, 0.08, 0.12]. The spatial distribution features are obtained by analyzing the distribution pattern of the behavior data in the geographical position or virtual space. For example, for position data, the system divides the region into grids, counts the behavior density in each grid, and forms a spatial vector [0.2, 0.1, 0.4, …, 0.3, 0.5]. The time sequence correlation features and the spatial distribution features are combined to obtain a data source feature matrix. Specifically, the system maps the time sequence vector and the spatial vector to a unified dimension space through a fully connected layer after splicing, and obtains a feature matrix with a dimension of MxN, where M represents the number of data sources and N represents the feature dimension.

[0087] The time sequence sequence analysis is performed on the data source feature matrix to extract the dynamic change features between the data sources. The system adopts a sliding window method, and the window size is set to T. The feature sequence of each data source is segmented. In each window, the system calculates the change rate, fluctuation amplitude, trend direction and other statistics of the features. For example, for the feature sequence [0.1, 0.15, 0.2, 0.18, 0.25] in the window, the average change rate is 0.0375, the fluctuation amplitude is 0.15, and the trend direction is upward. The dynamic change feature vector is constructed through these statistics to reflect the evolution pattern of the data source features over time.

[0088] The propagation time delay and the propagation distance between the data sources are calculated based on the dynamic change characteristics, the propagation time delay refers to the time required for the feature change of the data source A to propagate to the data source B, which is determined through cross-correlation analysis of the feature sequence. For example, the feature sequences of the data sources A and B have the maximum correlation at a time offset of 5 units, and the propagation time delay is 5. The propagation distance reflects the attenuation degree in the feature propagation process, which is calculated through the attenuation curve of the feature similarity with time. For example, the time length when the feature similarity attenuates from 0.8 to 0.4 can be used as an indicator of the propagation distance. The structural association features of the data sources are extracted according to the propagation time delay and the propagation distance, and a structural association matrix is constructed, and the element value represents the association strength between the data sources.

[0089] The dynamic change characteristics and the structural association features are fused to construct an initial association weight matrix. The system performs weighted summation of the dynamic change feature matrix and the structural association matrix at the element level, and the weight is adjusted according to the application scenario. For example, in the user behavior analysis scenario, the dynamic feature weight can be set to 0.6 and the structural feature weight can be set to 0.4; in the traffic flow prediction scenario, the dynamic feature weight can be set to 0.4 and the structural feature weight can be set to 0.6. The obtained initial association weight matrix has a dimension of MxM, representing the association strength between the M data sources.

[0090] The association weight matrix is subjected to row normalization to obtain the initial attention score between each pair of data sources. The global dependency relationship between the data sources is extracted through forward calculation, and the system combines the features of each data source with the association weights of all data sources to obtain a global weight matrix. For example, for data source i, its global representation is the weighted sum of all data source features, and the weight comes from the i-th row of the initial attention score. The local dependency relationship between the data sources is extracted through reverse calculation, and the system considers the influence degree of each data source on the current data source to construct a local weight matrix. For example, for data source j, its local representation is the weighted sum of all data sources that affect it, and the weight comes from the j-th column of the initial attention score. The global weight matrix and the local weight matrix are adaptively combined to obtain an optimized association weight matrix, and the adaptive weight is learned through training data, and the importance of global and local information is automatically adjusted according to the characteristics of different data sources.

[0091] The reconstructed features considering the correlation between data sources are obtained by multiplying the correlation weight matrix and the original feature matrix. The reconstructed features are calculated in multiple dimensions, including feature splicing, element-level product, feature difference, etc. For example, for the reconstructed feature vectors F1 = [0.3, 0.5, 0.2] and F2 = [0.4, 0.2, 0.6], the splicing result [0.3, 0.5, 0.2, 0.4, 0.2, 0.6], the element-level product [0.12, 0.1, 0.12], and the feature difference [0.1, -0.3, 0.4] are calculated. The system maps these interactive features to a unified dimensional space through a feature conversion network to generate the final fusion feature vector.

[0092] Through the above method, the application realizes adaptive feature fusion of behavior data of different data sources, can effectively capture the complex relationship between data sources, and improves the data fusion quality and downstream task performance.

[0093] In an optional implementation, the attention score between the data sources is calculated based on the initial correlation weight matrix, the global weight matrix is obtained by extracting the global dependency relationship between the data sources through forward calculation, the local weight matrix is obtained by extracting the local dependency relationship between the data sources through reverse calculation, and the optimized correlation weight matrix is obtained by adaptively weighting and combining the global weight matrix and the local weight matrix, including:

[0094] The initial attention score between the data sources is calculated based on the initial correlation weight matrix, the initial attention score is sorted according to the time sequence of the data sources, and a time sequence attention sequence is constructed;

[0095] The time sequence attention sequence is divided into a plurality of attention subsequences through a sliding window, and the distance matrix and the similarity matrix between the data sources in the attention subsequences are calculated respectively;

[0096] The global dependency relationship between the data sources is calculated based on the distance matrix, and the global weight matrix is generated, which reflects the long-range dependency features between the data sources in different time windows;

[0097] The local dependency relationship between the data sources is calculated based on the similarity matrix, and the local weight matrix is generated, which reflects the short-range dependency features between the data sources in adjacent time windows;

[0098] The adaptive weight coefficients are set for the global weight matrix and the local weight matrix, wherein the adaptive weight coefficients are dynamically adjusted according to the time sequence correlation between the data sources, the weighted global weight matrix and the local weight matrix are spliced to obtain the optimized correlation weight matrix.

[0099] In this embodiment, a method for calculating an optimized correlation weight matrix based on an initial correlation weight matrix is provided. This method realizes accurate modeling of the relationship between data sources through the calculation of global and local dependencies.

[0100] Specifically, first, the initial attention scores between data sources are calculated based on the initial correlation weight matrix. For example, for a system containing 10 data sources, the initial correlation weight matrix can be represented as a 10x10 matrix, where each element represents the degree of correlation between two data sources. By performing row normalization on this matrix, the initial attention scores are obtained. Assuming that the initial correlation weight between data source S1 and S2 is 0.8, and the initial correlation weight between S1 and S3 is 0.4, the attention scores of S1 and S2 after normalization are 0.67 and 0.33, respectively. These attention scores are arranged in the time order of the data sources, such as the collection times of the data sources being t1, t2,..., t10, and the time series attention sequence is constructed as the sequence of attention scores arranged in this time order.

[0101] Next, the time series attention sequence is divided into sliding windows. Assuming that the sliding window size is set to 3 and the step size is 1, for a time series attention sequence with a length of 10, 8 attention sub-sequences can be obtained. For each sub-sequence, the distance matrix and the similarity matrix between the data sources in the sub-sequence are calculated. The distance matrix is obtained by calculating the Euclidean distance between the attention scores, for example, the Euclidean distance between the first and second data sources in the sub-sequence is 0.31, with the attention scores being [0.67, 0.33] and [0.45, 0.55], respectively. The similarity matrix is obtained by calculating the cosine similarity between the attention scores, for example, the cosine similarity between the two data sources is 0.85.

[0102] Based on the distance matrix, the global dependency between the data sources is calculated to generate a global weight matrix. The specific method is to perform weighted averaging on the distance matrices of all sub-sequences, and the weights can be determined according to the time positions of the sub-sequences, with the weights of the sub-sequences closer to the current time being larger. For example, for the 8 sub-sequences, the weights can be set as [0.05, 0.08, 0.10, 0.12, 0.15, 0.16, 0.17, 0.17]. After weighted averaging, the global weight matrix reflects the long-range dependency characteristics between the data sources in different time windows. For example, the element value between data sources S1 and S5 in the global weight matrix is 0.42, indicating that these two data sources have a strong dependency relationship on a long time scale.

[0103] The local dependency relationship between the data sources is calculated based on the similarity matrix to generate a local weight matrix. Unlike the calculation of the global dependency relationship, the local dependency relationship mainly focuses on the short-range features within adjacent time windows. The specific method is to combine the similarity matrices of the two adjacent sub-sequences by weighting. For example, the two adjacent sub-sequences can be assigned weights of 0.6 and 0.4, and the similarity matrices of the two sub-sequences are combined by weighting to obtain the local weight matrix. The local weight matrix reflects the short-range dependency features between the data sources within the adjacent time windows. For example, the element value between the data sources S2 and S3 in the local weight matrix is 0.78, indicating that the two data sources have strong correlation in a short time.

[0104] The adaptive weight coefficients are set for the global weight matrix and the local weight matrix to realize the dynamic balance of the two dependency relationships. The adaptive weight coefficients are dynamically adjusted according to the time sequence correlation between the data sources. The time sequence correlation can be obtained by calculating the autocorrelation function of the data sources at different time points. If the autocorrelation function shows that the data has strong long-term dependency, the weight coefficient of the global weight matrix is increased; if the autocorrelation function shows that the data mainly has short-term correlation, the weight coefficient of the local weight matrix is increased. For example, for rapidly changing sensor data, the weight coefficient of the local weight matrix can be set to 0.7 and the weight coefficient of the global weight matrix can be set to 0.3; and for slowly changing user behavior data, the weight coefficient of the local weight matrix can be set to 0.4 and the weight coefficient of the global weight matrix can be set to 0.6.

[0105] The weighted global weight matrix and the local weight matrix are spliced to obtain an optimized correlation weight matrix. The splicing method is to add the weighted global weight matrix G and the local weight matrix L according to the corresponding element positions to form a new matrix M, that is, M[i,j] = αG[i,j] + βL[i,j], where α and β are the adaptive weight coefficients of the global weight matrix and the local weight matrix, respectively. For example, if the weight of two data sources in the global weight matrix is 0.5 and the weight in the local weight matrix is 0.7, and α = 0.4 and β = 0.6, then the weight between the two data sources in the optimized correlation weight matrix is 0.4 x 0.5 + 0.6 x 0.7 = 0.62.

[0106] The optimized correlation weight matrix obtained by the above method comprehensively considers the global long-range dependency and local short-range dependency between the data sources, and can more accurately describe the complex relationship between the data sources, thereby improving the accuracy of data analysis and prediction based on the correlation weight.

[0107] Figure 2 The bar chart for precision comparison and analysis of the dependency relationship calculation method of the embodiments of the present application is shown in the following table:

[0108] The figure shows the performance of three relationship types (global dependency relationship, local dependency relationship and adaptive weighted combination) under different test durations. The global dependency relationship reflects the end-to-end dependency analysis results of the system, the local dependency relationship represents the local connection between modules, and the adaptive weighted combination is an optimized fusion scheme of the two. From the data, in the short window test, the accuracy of the three schemes is 62.5%, 72.3% and 80.5% respectively; in the medium length window test, it reaches 76.6%, 64.3% and 86.7% respectively; in the long window test, it is 84.5%, 56.4% and 92.3% respectively. With the extension of the test time window, the accuracy of the global dependency relationship shows an upward trend, while the local dependency relationship shows a downward trend, but the adaptive weighted combination scheme always maintains the best performance, fully embodying the adaptability and stability advantages of the scheme.

[0109] In an optional implementation, the user is subjected to cluster analysis according to the fusion feature vector, and the cluster center is dynamically adjusted based on the time series drift degree of the user group to obtain a user clustering result, including:

[0110] A feature distance matrix between users is calculated based on the fusion feature vector;

[0111] Density peak value analysis is performed on the feature distance matrix to identify a user point with a local maximum density as an initial cluster center;

[0112] The Mahalanobis distance between each user and the initial cluster center is calculated;

[0113] A user similarity matrix is constructed based on the Mahalanobis distance, the user similarity matrix is divided into multiple time segments along the time dimension, and the evolution characteristics of the user group in each time segment are calculated;

[0114] The stability coefficient of the user group is calculated according to the evolution characteristics, and the initial cluster center is dynamically adjusted based on the stability coefficient, wherein when the stability coefficient is lower than a second preset threshold, the cluster centers with a distance less than a cluster radius are merged, and when the stability coefficient is higher than the second preset threshold, the cluster centers with a data density greater than an average density of the cluster centers are split;

[0115] The adjusted cluster center is iteratively optimized, the optimal cluster number is determined by calculating the time series drift degree of the user group, and the user clustering result considering the time series evolution is obtained.

[0116] After obtaining the fusion feature vector, first, the feature distance matrix between users is calculated. Specifically, for any two users i and j, the Euclidean distance between their feature vectors is calculated, and an N x N distance matrix D is constructed, where N is the total number of users. For example, for a dataset containing 5000 users, each user's fusion feature vector has a dimension of 128, and the calculated distance matrix D is a 5000 x 5000 matrix, where each element in the matrix represents the feature distance between the corresponding two users.

[0117] Based on the feature distance matrix D, the present application performs density peak analysis to identify initial cluster centers. For each user point, the local density value p is calculated, which is the number of neighbor users within a given radius range. In this embodiment, the local density calculation uses a Gaussian kernel function for smoothing processing, and the distance threshold is set to 20% of the average feature distance. At the same time, the distance δ of each user point to the nearest user point with higher local density is calculated. The user point with the maximum product of local density p and distance δ is identified as the initial cluster center. In the 5000 user sample, 8 initial cluster centers are identified through density peak analysis, with local density values of 125, 118, 110, 105, 98, 92, 87 and 82, respectively, indicating the highest data density in their respective regions.

[0118] After obtaining the initial cluster centers, the Mahalanobis distance between each user and these cluster centers is calculated. The Mahalanobis distance takes into account the covariance structure in the feature space, and can more accurately measure the similarity between users. For each user u and cluster center c, the Mahalanobis distance M(u, c) is calculated. In actual calculation, the covariance matrix of the feature vector is used for standardization to obtain a more reasonable distance measurement result. For example, the Mahalanobis distance of user ID U1001 to the 8 cluster centers is 3.25, 5.12, 7.89, 2.45, 6.34, 8.76, 4.21 and 5.67, respectively.

[0119] Based on the Mahalanobis distance, a user similarity matrix S is constructed, where the element S(i, j) in the matrix represents the similarity between user i and cluster center j, and the larger the value, the higher the similarity. To analyze the user behavior changes in the time dimension, the user similarity matrix is divided into multiple time segments according to the time dimension. In this embodiment, the data of one year is divided into 12 months, and each month constitutes a time segment. For each time segment t, the evolution characteristics of the user group are calculated, including the position offset of the cluster center, the change rate of the cluster member number, and the internal tightness change. For example, for the 3rd time segment, the position offset of the 1st cluster center is 0.35, the member number change rate is +8.5%, and the internal tightness change is -3.2%, indicating that the cluster has a slight expansion trend in this time period.

[0120] According to the calculated evolution characteristics, the stability coefficient σ of the user group is further calculated. The stability coefficient is obtained by the weighted combination of the cluster center displacement, the member change rate and the tightness change, and the value range is 0 to 1, and the larger the value is, the more stable the group is. The second preset threshold is set to 0.6, and when the stability coefficient is lower than 0.6, it is determined that the user group is in an unstable state, and the cluster center needs to be merged. The merging strategy is: if the distance between two cluster centers is less than half of the sum of the respective cluster radii, then the two centers are merged into one. For example, at time segment t = 5, the stability coefficients of cluster centers c2 and c5 are 0.52 and 0.55 respectively, the distance between them is 2.3, and the respective cluster radii are 2.5 and 2.7, which meet the merging condition, so the two cluster centers are merged.

[0121] When the stability coefficient is higher than 0.6, it is determined that the user group is relatively stable, and at this time the data density distribution in the cluster is investigated. If there is a local area in a certain cluster whose data density is greater than 1.5 times the average density of the cluster center, then the area is split into a new cluster. For example, at time segment t = 8, the stability coefficient of cluster c3 is 0.75, and there is a local area whose data density is 105, while the average density of the cluster center is 65, which meets the splitting condition, so a new cluster center is split out from the area.

[0122] The adjusted cluster centers are iteratively optimized, and the attribution cluster of all users is calculated and the cluster center position is updated each time, until the cluster center position changes less than the preset convergence threshold 0.01 or reaches the maximum iteration number 100 times. In the optimization process, the optimal cluster number is determined by calculating the time sequence drift degree of the user group. The time sequence drift degree is measured by the change amount of user cluster distribution between consecutive time segments. When the cluster number is set too large, the drift degree will abnormally increase; when it is set too small, it cannot capture the internal differences of the user group. By testing different cluster numbers (from 5 to 15), it is finally determined that when the cluster number is 7, the best balance between maintaining time sequence stability and group differentiation can be achieved.

[0123] Finally, the user clustering results considering the time sequence evolution are obtained, including 7 user groups with different behavior characteristics. These groups represent: high-activity long-term users (accounting for 23.5%), seasonal high-consumption users (accounting for 18.7%), stable growth users (accounting for 15.3%), emerging rapid development users (accounting for 12.6%), fluctuating users (accounting for 11.8%), slow loss users (accounting for 10.2%) and low-activity temporary users (accounting for 7.9%). Each group has a unique time sequence behavior pattern and characteristic combination, which can provide decision basis for subsequent precision marketing and user management.

[0124] In an alternative embodiment, a label propagation network is constructed by taking users as nodes of a graph structure, similarity between users is taken as weight of edges, label propagation probability considering time evolution characteristics of nodes is iteratively calculated, label assignment is performed on users based on a first preset threshold to obtain a label assignment result, including:

[0125] A graph structure is constructed by taking users as nodes, multi-dimensional similarity characteristics between nodes are calculated, and the multi-dimensional similarity characteristics are normalized to obtain weight values of edges.

[0126] The graph structure is divided into multiple time windows along the time dimension, and a feature vector of a node is calculated for each time window.

[0127] Degree centrality, betweenness centrality and clustering coefficient of the node are calculated based on the feature vector, and a node importance score matrix is constructed.

[0128] The node importance score matrix is analyzed by using a spectral clustering method to determine a key node group, and the key node group is taken as a source node of label propagation.

[0129] State transition probability of the node between adjacent time windows is calculated, a time sequence attenuation function is generated in combination with activity of the node, a time sequence weight matrix is obtained by adaptively weighting the time sequence attenuation function and the weight values of the edges, and label propagation probability between nodes is iteratively calculated based on the time sequence weight matrix.

[0130] A label propagation confidence matrix is constructed, a label record with label propagation probability greater than a first preset threshold and having time sequence consistency is recorded to a label candidate set, and the label assignment result is obtained by optimizing the label candidate set.

[0131] The present application provides a user label propagation method, which constructs a label propagation network by taking users as nodes of a graph structure, calculates multi-dimensional similarity between users as weight of edges, and iteratively calculates label propagation probability considering time evolution characteristics of nodes, so as to realize accurate user label assignment.

[0132] The multi-dimensional similarity features include user behavior similarity, interest similarity, and social relationship similarity. The user behavior similarity is calculated by analyzing the co-occurrence frequency of user clicks, browsing, purchases, etc.; the interest similarity is calculated by analyzing the overlap degree of user label sets; and the social relationship similarity is calculated by analyzing the direct and indirect social connections between users. For example, for user A and user B, assuming that the number of common browsed goods is 15, and the total number of browsed goods by each is 45 and 60 respectively, the behavior similarity can be calculated as the number of common browsed goods divided by the union set of browsed goods by both, resulting in 0.167. Similarly, the interest similarity is 0.35, and the social relationship similarity is 0.28. These similarity features are normalized by linear combination and assigned weights of 0.4, 0.35, and 0.25 respectively, resulting in a final edge weight value of 0.26.

[0133] Next, the graph structure is divided into multiple time windows along the time dimension, with each window spanning 7 days and adjacent windows overlapping by 3 days. For each time window, a feature vector of the node is calculated. The feature vector includes the user's activity, interaction frequency, and content preference in the time window. For example, a user's daily active duration in a time window is 45 minutes, the daily interaction frequency is 23 times, and the content preference distribution is 60% video, 30% image-text, and 10% audio. Based on this, the feature vector of the node in this time window is generated.

[0134] Based on the feature vector, the degree centrality, betweenness centrality, and clustering coefficient of the node are calculated to construct a node importance score matrix. The degree centrality reflects the number of direct connections of the node, the betweenness centrality reflects the degree of the node acting as a "bridge" in the network, and the clustering coefficient reflects the connection tightness between the node's neighbors. For example, a node's degree centrality is 0.72 (indicating a direct connection with 72% of the nodes), betweenness centrality is 0.31, and clustering coefficient is 0.65, then the node's importance score can be calculated by weighted average as 0.58. The constructed importance score matrix includes the importance scores of all nodes in each time window.

[0135] The node importance score matrix is analyzed using the spectral clustering method to determine the key node group. Spectral clustering first constructs a similarity matrix, calculates its Laplacian matrix, performs eigenvalue decomposition on the Laplacian matrix, selects the first k eigenvectors to form a feature space, and applies the K-means clustering algorithm in the feature space. The optimal number of clusters is determined to be 5 by calculating the silhouette coefficient, resulting in 5 node clusters. The top 20% nodes with the highest importance scores in each cluster are selected as key nodes, a total of 78 key nodes are determined as source nodes for label propagation.

[0136] The state transition probability of the node between adjacent time windows is calculated, and a time sequence decay function is generated in combination with the node activity. The state transition probability is calculated by analyzing the feature vector change of the node in adjacent time windows. For example, the feature vector similarity of a node in time windows t1 and t2 is 0.83, indicating that the probability of stable state maintenance is 0.83. The node activity is defined as the proportion of active days of a user on the platform, and a user is active for 24 days in a 30-day observation period, with an activity of 0.8. Based on this, the time sequence decay function is defined as a function that exponentially decays with the increase of time interval, and the decay rate is inversely proportional to the activity. The time sequence decay function is adaptively weighted with the edge weight value, and the weighting coefficient is dynamically adjusted according to the node importance. The higher the importance of the node, the greater the weighting coefficient, which ranges from 0.6 to 0.9, to obtain the time sequence weight matrix.

[0137] The label propagation probability between nodes is iteratively calculated based on the time sequence weight matrix. In the initial state, the label confidence of the source node is set to 1.0, and the label confidence of the non-source node is set to 0. In each iteration, the node receives label information from its neighbors and updates its own label confidence according to the time sequence weight matrix. For example, a node has three neighbors with time sequence weights of 0.65, 0.42, and 0.38, and the label confidence of the neighbors is 0.8, 0.7, and 0.5, respectively. The updated label confidence of the node is 0.65x0.8+0.42x0.7+0.38x0.5=1.01, and after normalization processing, it is 0.67. The iteration process continues until the label confidence changes less than 0.001 or reaches the maximum iteration number of 100 times.

[0138] The label records with label propagation probability greater than the first preset threshold 0.75 and time sequence consistency are recorded to the label candidate set. The time sequence consistency requires that the average confidence of the label fluctuates by no more than 20% in consecutive 3 time windows. For example, the confidence of a user's "sports enthusiast" label in consecutive three time windows is 0.78, 0.82 and 0.76, the average value is 0.79, and the fluctuation is 7.6%, which meets the time sequence consistency requirement, and the confidence is greater than 0.75, so the label is included in the candidate set.

[0139] The label candidate set is optimized, the mutually exclusive labels are removed, the highly related labels are merged, and the final label assignment result is generated. For example, the mutually exclusive labels such as "budget limited consumer" and "luxury goods enthusiast" of the same user are removed; the highly related labels such as "soccer fan" and "sports event follower" are merged into "sports enthusiast". Finally, 1-5 labels are assigned to each user, with an accuracy of 92.3%, which is 18.7% higher than the traditional method.

[0140] In an optional embodiment, the state transition probability of the node between adjacent time windows is calculated, a time sequence decay function is generated in combination with the activity of the node, the time sequence decay function is adaptively weighted with the weight value of the edge to obtain a time sequence weight matrix, and the label propagation probability between nodes is iteratively calculated based on the time sequence weight matrix, including:

[0141] The multi-dimensional behavior features of the node in the adjacent time windows are obtained, a node state feature matrix is constructed, and the mutual information and conditional entropy between nodes are calculated based on the node state feature matrix;

[0142] The node behavior correlation network is constructed using the mutual information and the conditional entropy, the topological features and community features of the node are extracted, and the multi-dimensional node activity index is generated in combination with the behavior intensity of the node;

[0143] The high-order Markov chain is constructed based on the node state feature matrix, the transition probability matrix of the high-order Markov chain is calculated under different time scales, the state transition edge with a probability value greater than a third preset threshold in the transition probability matrix is taken as a long-range dependence relationship, and the state transition probability of the node is obtained.

[0144] The time sequence decay function is constructed according to the state transition probability and the multi-dimensional node activity index, wherein the decay rate of the time sequence decay function is dynamically adjusted according to the node activity, and the higher the activity, the smaller the decay rate.

[0145] The time sequence decay function is adaptively weighted and combined with the weight value of the edge, the combination parameters are dynamically optimized based on the evolution trajectory features of the node in the historical window, a weight matrix fusing time sequence dynamic characteristics is generated, and the label propagation probability between nodes is obtained by iteratively calculating based on the weight matrix fusing time sequence dynamic characteristics.

[0146] The embodiment provides a method for calculating the label propagation probability between nodes, which calculates the state transition probability of the node between adjacent time windows, generates a time sequence decay function in combination with the activity of the node, and accurately calculates the label propagation probability between nodes.

[0147] The method first acquires the multi-dimensional behavior characteristics of the nodes in adjacent time windows, and constructs a node state feature matrix. Specifically, for a user node in a social network, its behavior characteristics can include the number of posts, the number of comments, the number of likes, the number of shares, and the browsing time, etc. For example, for a certain user node within a time window t, the number of posts is recorded as 15, the number of comments is 32, the number of likes is 47 times, the number of shares is 8 times, and the browsing time is 127 minutes, which constitute the behavior characteristic vector of the node in the time window. The behavior characteristics of all nodes in each time window are normalized to form a node state feature matrix with dimensions N x T x D, where N is the number of nodes, T is the number of time windows, and D is the dimension of the behavior characteristics.

[0148] Based on the node state feature matrix, the mutual information and conditional entropy between nodes are calculated. Mutual information represents the degree of correlation between the behavior patterns of two nodes, and conditional entropy represents the uncertainty of the state of another node given the state of a node. The calculation method uses frequency statistics to calculate the joint probability and marginal probability after discretizing the behavior characteristics. For example, the posting behavior of two user nodes can be divided into high, medium and low three levels, and the mutual information value is calculated by counting the frequency of the two users in each level combination.

[0149] The node behavior correlation network is constructed using mutual information and conditional entropy. The specific method is to establish a connection between the corresponding nodes when the mutual information between the node pairs is greater than the preset threshold value 0.3 and the conditional entropy is less than the preset threshold value 0.7. Based on this network, the topological features of the nodes such as degree centrality, betweenness centrality and closeness centrality, as well as the community features such as community membership and community activity are extracted. Combined with the behavior intensity of the nodes such as posting frequency and interaction frequency, a multi-dimensional node activity index is generated. For example, the degree centrality of a certain node is 0.75, the community activity is 0.82, and the posting frequency is 0.63, and the comprehensive calculation obtains the activity index of 0.74.

[0150] Based on the node state feature matrix, a high-order Markov chain is constructed to consider the state dependence relationship of the node in multiple time windows. The transition probability matrix of the high-order Markov chain is calculated under different time scales, including short-term (1-3 time windows), medium-term (4-7 time windows) and long-term (8-12 time windows). Through statistical analysis, when the transition probability value is greater than the preset threshold value 0.4, the state transition edge is marked as a long-range dependence relationship, and the state transition probability of the node is obtained. For example, the probability of node A from the behavior state "high post-low comment" to "medium post-high comment" is 0.53, which is greater than the threshold value 0.4, and a long-range dependence relationship is established between the two states.

[0151] A time-decay function is constructed according to the state transition probability and the multi-dimensional node activity index. The decay rate of the function is dynamically adjusted according to the node activity, and the higher the activity, the smaller the decay rate. The specific implementation is to design a basic decay function, and the decay parameter is negatively correlated with the node activity. For example, for a node with an activity of 0.74, the decay parameter is set to 0.15, and for a node with an activity of 0.35, the decay parameter is set to 0.42. In this way, the historical behavior of a high-activity node has a longer impact on the current, while the historical behavior of a low-activity node has a faster impact.

[0152] The time-decay function is adaptively combined with the weight value of the edge to generate a weight matrix that integrates the time-dynamic characteristics. In the adaptive weighting process, the combination parameters are dynamically optimized based on the evolution trajectory features of the nodes in the historical window. Specifically, the stability of the state change of the node in the historical window is analyzed, and the node with high stability is given higher weight to historical information, and the node with low stability is given higher weight to current information. For example, the state change stability of node B is 0.68, and the historical information weight of node B is set to 0.65, and the current information weight of node B is set to 0.35; while the state change stability of node C is 0.31, and the historical information weight of node C is set to 0.25, and the current information weight of node C is set to 0.75.

[0153] Based on the weight matrix integrating the time-dynamic characteristics, iterative calculation is performed to obtain the label propagation probability between nodes. The maximum number of iterations is set to 100, and the convergence threshold is set to 0.001. In each iteration, the node propagates its label to adjacent nodes according to the weight matrix, and updates its state according to the received label information. After iteration to convergence or reaching the maximum number of iterations, the final label propagation probability matrix between nodes is obtained. For example, the probability of node D propagating label 1 to node E is 0.72, the probability of node D propagating label 2 to node E is 0.23, and the probability of node D propagating label 3 to node E is 0.05.

[0154] Through the above method, the label propagation probability between nodes is accurately calculated, which provides effective support for subsequent community discovery, information propagation analysis and other tasks.

[0155] Figure 3 The figure shows the accuracy comparison of different label propagation algorithms in the embodiment of the application.

[0156] The figure shows the performance trend of three different label propagation schemes in the iteration process. In the legend, the traditional label propagation represents the basic label propagation algorithm, the time-aware label propagation represents the improved algorithm with the introduction of time features, and the technical solution is the optimization algorithm that integrates multi-dimensional features. From the performance curve, it can be observed that in the early stage of iteration (100 times), the label propagation accuracy of the three schemes is 68.2%, 76.4% and 82.3% respectively; with the increase of iteration times, the performance of each scheme shows an upward trend, and when the iteration reaches 500 times, the accuracy is increased to 77.3%, 82.5% and 89.8% respectively; by the time the iteration reaches 1000 times, the final accuracy of the three schemes is 80%, 85.6% and 93.4% respectively. Throughout the entire iteration process, the technical solution always maintains the optimal performance and the fastest convergence speed, which is 13.4 and 7.8 percentage points higher than the traditional method and the time-aware method respectively, fully proving the superiority of the scheme in the label propagation task.

[0157] In an optional implementation, a corresponding label set is generated for each user based on the label assignment result, and a multi-level label structure tree is constructed according to the hierarchical relationship and the correlation degree between labels, including:

[0158] Based on the label assignment result, the label set of each user is extracted, and the co-occurrence frequency matrix between the labels in the label set is calculated;

[0159] The label set is subjected to semantic feature extraction to construct a label semantic vector, and based on the co-occurrence frequency matrix and the label semantic vector, the usage frequency and coverage of the label are calculated to generate a label importance score;

[0160] According to the label semantic vector and the label importance score, the labels are divided into different levels according to the semantic concept granularity, and the labels with a label importance score greater than the average label importance score are assigned to the upper node to construct an initial hierarchical tree of labels;

[0161] The information gain ratio and semantic inclusion degree between labels of different levels are calculated, and when the information gain ratio and semantic inclusion degree of the upper label to the lower label satisfy a first preset condition, the hierarchical affiliation relationship between the two labels is confirmed, and the initial hierarchical tree is optimized based on the confirmed hierarchical affiliation relationship;

[0162] The co-occurrence frequency matrix and the label semantic vector are used to calculate the correlation strength between labels of the same level, and the labels that satisfy a second preset condition are constructed to have a horizontal correlation relationship, and based on the hierarchical affiliation relationship and the horizontal correlation relationship, a multi-level label structure tree is generated.

[0163] Based on the label assignment result, the label set of each user is extracted, and a co-occurrence frequency matrix is constructed by analyzing the co-occurrence relationship of labels in user data. Taking the user label data of an e-commerce platform as an example, for user A, its label set may contain labels such as "digital products", "mobile phone accessories", "charging equipment", etc. By statistical analysis of the label sets of all users, the co-occurrence frequency between label pairs is calculated. For example, the co-occurrence frequency of "digital products" and "mobile phone accessories" is 0.78, the co-occurrence frequency of "mobile phone accessories" and "charging equipment" is 0.65, and the co-occurrence frequency of "digital products" and "charging equipment" is 0.42. These co-occurrence frequency values constitute the co-occurrence frequency matrix, which reflects the degree of association between labels.

[0164] Semantic feature extraction of label set is a key step to construct label semantic vector. This step uses natural language processing technology to analyze the text meaning of labels and extract semantic features. In specific implementation, word embedding technology is used to map each label to a high-dimensional vector space. For example, the semantic vector of "digital products" may be [0.23, 0.45, -0.12, 0.67,...], usually with a dimension of 100 to 300. These vectors capture the semantic information of labels, so that labels with similar semantics are closer in vector space.

[0165] Based on the co-occurrence frequency matrix and label semantic vector, the usage frequency and coverage of the label are calculated, and the label importance score is generated. The usage frequency of a label refers to the proportion of the number of times the label appears in all user label sets, and the coverage represents the proportion of the number of users using the label to the total number of users. For example, the usage frequency of "digital products" is 0.25, the coverage is 0.32, and the importance score is 0.78; the usage frequency of "mobile phone accessories" is 0.18, the coverage is 0.24, and the importance score is 0.65; the usage frequency of "charging equipment" is 0.12, the coverage is 0.18, and the importance score is 0.47. The calculation of label importance score considers the usage frequency, coverage and association degree with other labels of the label, reflecting the importance of the label in the entire label system.

[0166] According to the label semantic vector and the label importance score, the labels are divided into different levels according to the semantic concept granularity. The semantic concept granularity is determined by analyzing the abstract degree of the label semantic vector. The label concept granularity is coarse when the abstract degree is high, and should be located at the upper level of the hierarchical structure. The mean value of all label importance scores is calculated, for example, the mean value is 0.55, and the labels with importance scores greater than the mean value are assigned to the upper layer. In the above example, the importance scores of "digital products" (0.78) and "mobile phone accessories" (0.65) are greater than the mean value, and are assigned to the upper layer; while the score of "charging equipment" (0.47) is less than the mean value, and is assigned to the lower layer. In this way, an initial hierarchical tree of labels is constructed.

[0167] The information gain ratio and semantic inclusion degree between labels at different levels are calculated to confirm the hierarchical subordinate relationship between labels. The information gain ratio measures the degree of information provided by one label to another label, and the semantic inclusion degree reflects whether the semantics of one label contains the semantics of another label. When the information gain ratio of the upper label to the lower label is greater than 0.3 and the semantic inclusion degree is greater than 0.5 (the first preset condition), it is confirmed that there is a hierarchical subordinate relationship between the two. For example, the information gain ratio of "digital products" to "charging equipment" is 0.42, and the semantic inclusion degree is 0.68, which satisfies the first preset condition, so it is confirmed that "charging equipment" is a lower level label of "digital products". While the information gain ratio of "mobile phone accessories" to "charging equipment" is 0.56, and the semantic inclusion degree is 0.75, which also satisfies the condition, and it is confirmed that "charging equipment" is also a lower level label of "mobile phone accessories". Based on these confirmed hierarchical subordinate relationships, the initial hierarchical tree is optimized and adjusted.

[0168] The association strength between labels at the same level is calculated using the co-occurrence frequency matrix and the label semantic vector. The association strength considers both the co-occurrence frequency and the semantic similarity of the labels. When the association strength of a label pair is greater than 0.6 (the second preset condition), a horizontal association relationship is constructed between the two labels. For example, the association strength of "digital products" and "mobile phone accessories" at the same upper level is 0.72, which satisfies the second preset condition, so a horizontal association relationship is established between them. Finally, based on the determined hierarchical subordinate relationship and horizontal association relationship, a complete multi-level label structure tree is generated. In this structure tree, "digital products" and "mobile phone accessories" are located at the upper level and have a horizontal association, and "charging equipment" is located at the lower level and has a hierarchical subordinate relationship with both upper level labels.

[0169] This multi-level label structure tree reflects the hierarchical relationship and association degree between labels through a scientific method, providing a structured label system for user portrait analysis, personalized recommendation and other applications, effectively improving the organization efficiency and application value of labels. The entire construction process fully utilizes the semantic information and usage characteristics of labels, ensuring the rationality and practicality of the label structure.

[0170] The embodiment of the application is based on a user label automatic construction system of multi-source heterogeneous data, comprising:

[0171] The first unit is configured to acquire behavior data generated by a user in multiple different data sources, split the behavior data according to a preset time window, and perform standardization processing on the split behavior data to obtain standardized behavior data.

[0172] The second unit is configured to calculate an association weight matrix between different data sources based on the standardized behavior data, perform adaptive feature fusion on the behavior data of different data sources according to the association weight matrix by using a bidirectional attention mechanism, and generate a fusion feature vector.

[0173] The third unit is configured to perform clustering analysis on the user according to the fusion feature vector, and dynamically adjust a clustering center based on a time sequence drift degree of a user group to obtain a user clustering result.

[0174] The fourth unit is configured to construct a label propagation network of a graph structure by taking the user as a node, take a similarity between users as a weight of an edge, iteratively calculate a label propagation probability considering a time sequence evolution feature of the node, perform label assignment on the user based on a first preset threshold, and obtain a label assignment result.

[0175] The fifth unit is configured to generate a corresponding label set for each user based on the label assignment result, and construct a multi-level label structure tree according to a hierarchical relationship and an association degree between labels.

[0176] The third aspect of the embodiment of the application provides an electronic device, comprising:

[0177] a processor;

[0178] a memory for storing processor-executable instructions;

[0179] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0180] The fourth aspect of the embodiment of the application provides a computer-readable storage medium having computer program instructions stored thereon, and the computer program instructions are executed by a processor to implement the method described above.

[0181] The application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for executing various aspects of the application.

[0182] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for automatic construction of user labels based on multi-source heterogeneous data, characterized in that, The method comprises the following steps: acquiring behavior data generated by a user in multiple different data sources, dividing the behavior data according to a preset time window, and performing standardization processing on the divided behavior data to obtain standardized behavior data; based on the standardized behavior data, calculating a correlation weight matrix between different data sources, and using a bidirectional attention mechanism to perform adaptive feature fusion on the behavior data of different data sources according to the correlation weight matrix to generate a fusion feature vector; performing clustering analysis on the user according to the fusion feature vector, and dynamically adjusting the clustering center based on the time sequence drift degree of the user group to obtain a user clustering result; constructing a label propagation network of a graph structure with the user as a node, taking the similarity between users as the weight of the edge, iteratively calculating the label propagation probability considering the time sequence evolution characteristics of the node, and performing label assignment on the user based on a first preset threshold to obtain a label assignment result; based on the label assignment result, generating a corresponding label set for each user, and constructing a multi-level label structure tree according to the hierarchical relationship and correlation degree between the labels; based on the standardized behavior data, calculating a correlation weight matrix between different data sources, and using a bidirectional attention mechanism to perform adaptive feature fusion on the behavior data of different data sources according to the correlation weight matrix to generate a fusion feature vector, comprising: based on the standardized behavior data, calculating the time sequence correlation feature and the spatial distribution feature of each data source, combining the time sequence correlation feature and the spatial distribution feature to obtain a data source feature matrix; performing time sequence analysis on the data source feature matrix to extract the dynamic change feature between the data sources; based on the dynamic change feature, calculating the propagation time delay and the propagation distance between the data sources, and extracting the structural correlation feature of the data sources according to the propagation time delay and the propagation distance; performing feature fusion on the dynamic change feature and the structural correlation feature to construct an initial correlation weight matrix; based on the initial correlation weight matrix, calculating the attention score between the data sources, extracting the global dependency relationship between the data sources to obtain a global weight matrix through forward calculation, extracting the local dependency relationship between the data sources to obtain a local weight matrix through reverse calculation, and adaptively combining the global weight matrix and the local weight matrix to obtain an optimized correlation weight matrix, comprising: based on the initial correlation weight matrix, calculating the initial attention score between the data sources, sorting the initial attention score according to the time sequence of the data sources, and constructing a time sequence attention sequence; performing sliding window segmentation on the time sequence attention sequence to obtain multiple attention sub-sequences, and respectively calculating the distance matrix and the similarity matrix between the data sources in the attention sub-sequences; based on the distance matrix, calculating the global dependency relationship between the data sources to generate a global weight matrix, wherein the global weight matrix reflects the long-range dependency feature between the data sources under different time windows; calculating local dependency relationships between the data sources based on the similarity matrix, generating a local weight matrix, the local weight matrix reflecting short-range dependency characteristics between the data sources within adjacent time windows; setting adaptive weight coefficients for the global weight matrix and the local weight matrix, wherein the adaptive weight coefficients are dynamically adjusted according to the time sequence correlation between the data sources, and performing matrix splicing on the weighted global weight matrix and the local weight matrix to obtain an optimized correlation weight matrix; based on the optimized correlation weight matrix, reconstructing features of behavior data of each data source to obtain reconstructed features, and performing multi-dimensional feature interaction calculation on the reconstructed features to generate a fusion feature vector.

2. The method of claim 1, wherein, performing clustering analysis on users according to the fusion feature vector, and dynamically adjusting clustering centers based on a time sequence drift degree of user groups to obtain user clustering results, including: calculating a feature distance matrix between users based on the fusion feature vector; performing density peak value analysis on the feature distance matrix to identify user points with local density maximum values as initial clustering centers; calculating Mahalanobis distances between each user and the initial clustering centers; based on the Mahalanobis distances, constructing a user similarity matrix, dividing the user similarity matrix into multiple time segments along a time dimension, and calculating evolution characteristics of user groups within each time segment; calculating a stability coefficient of the user groups according to the evolution characteristics, and dynamically adjusting the initial clustering centers based on the stability coefficient, wherein when the stability coefficient is lower than a second preset threshold, clustering centers with a distance less than a clustering radius are merged, and when the stability coefficient is higher than the second preset threshold, clustering centers with a data density greater than an average density of the clustering centers are split; iteratively optimizing the adjusted clustering centers to determine an optimal clustering number by calculating a time sequence drift degree of the user groups, and obtaining user clustering results considering time sequence evolution.

3. The method of claim 1, wherein, constructing a label propagation network of a graph structure with users as nodes, taking similarity between users as weights of edges, iteratively calculating label propagation probabilities considering time sequence evolution characteristics of nodes, and performing label allocation on users based on a first preset threshold to obtain a label allocation result, including: constructing a graph structure with users as nodes, calculating multi-dimensional similarity characteristics between nodes, and obtaining weight values of edges by normalizing the multi-dimensional similarity characteristics; dividing the graph structure into multiple time windows along a time dimension, and calculating feature vectors of nodes for each time window; calculating degree centrality, betweenness centrality and clustering coefficient of the nodes based on the feature vectors, and constructing a node importance score matrix; analyzing the node importance score matrix by using a spectral clustering method to determine a key node group, and taking the key node group as source nodes of label propagation; calculating state transition probabilities of nodes between adjacent time windows, combining activity degrees of the nodes to generate a time sequence decay function, adaptively weighting the time sequence decay function and the weight values of the edges to obtain a time sequence weight matrix, and iteratively calculating label propagation probabilities between nodes based on the time sequence weight matrix; and The label propagation confidence matrix is constructed, label records with label propagation probability greater than a first preset threshold and time sequence consistency are recorded to a label candidate set, and the label candidate set is optimized to obtain a label assignment result.

4. The method of claim 3, wherein, The state transition probability of the node between adjacent time windows is calculated, the time sequence decay function is generated by combining the activity of the node, the time sequence decay function is adaptively weighted with the weight value of the edge to obtain a time sequence weight matrix, and the label propagation probability between nodes is iteratively calculated based on the time sequence weight matrix, including: The multi-dimensional behavior characteristics of the node in the adjacent time windows are obtained, the node state feature matrix is constructed, and the mutual information and conditional entropy between nodes are calculated based on the node state feature matrix; The node behavior correlation network is constructed using the mutual information and the conditional entropy, the topological features and community features of the node are extracted, and the multi-dimensional node activity index is generated by combining the behavior intensity of the node; The high-order Markov chain is constructed based on the node state feature matrix, the transition probability matrix of the high-order Markov chain is calculated under different time scales, the state transition edges with probability values greater than a third preset threshold in the transition probability matrix are taken as long-range dependence relationships, and the state transition probability of the node is obtained. The time sequence decay function is constructed according to the state transition probability and the multi-dimensional node activity index, wherein the decay rate of the time sequence decay function is dynamically adjusted according to the node activity, and the higher the activity, the smaller the decay rate. The time sequence decay function is adaptively weighted and combined with the weight value of the edge, the combination parameters are dynamically optimized based on the evolution trajectory features of the node in the historical window, a weight matrix fusing time sequence dynamic characteristics is generated, and the label propagation probability between nodes is obtained by iteratively calculating based on the weight matrix fusing time sequence dynamic characteristics.

5. The method of claim 1, wherein, Based on the label assignment result, a corresponding label set is generated for each user, and a multi-level label structure tree is constructed according to the hierarchical relationship and correlation degree between labels, including: Based on the label assignment result, the label set of each user is extracted, and the co-occurrence frequency matrix between label pairs in the label set is calculated. The label semantic vector is constructed by performing semantic feature extraction on the label set, the usage frequency and coverage of the label are calculated based on the co-occurrence frequency matrix and the label semantic vector, and the label importance score is generated. According to the label semantic vector and the label importance score, the labels are divided into different levels according to the semantic concept granularity, the labels with label importance scores greater than the average label importance score are assigned to the upper nodes, and an initial hierarchical tree of the labels is constructed. The information gain ratio and semantic inclusion degree between labels at different levels are calculated, when the information gain ratio and semantic inclusion degree of the upper label to the lower label satisfy a first preset condition, the hierarchical affiliation relationship between the two labels is confirmed, and the initial hierarchical tree is optimized based on the confirmed hierarchical affiliation relationship. The association strength between labels at the same level is calculated using the co-occurrence frequency matrix and the label semantic vector, the label pairs with association strength satisfying a second preset condition are constructed to have a horizontal correlation relationship, and a multi-level label structure tree is generated based on the hierarchical affiliation relationship and the horizontal correlation relationship.

6. A user tag automated construction system based on multi-source heterogeneous data, for implementing the method according to any one of claims 1-5, characterized in that, including: The first unit is configured to acquire behavior data generated by a user in multiple different data sources, split the behavior data according to a preset time window, and perform standardization processing on the split behavior data to obtain standardized behavior data. The second unit is configured to calculate an association weight matrix between different data sources based on the standardized behavior data, perform adaptive feature fusion on the behavior data of different data sources according to the association weight matrix by using a bidirectional attention mechanism, and generate a fusion feature vector. The third unit is configured to perform clustering analysis on the user according to the fusion feature vector, dynamically adjust a clustering center based on a time sequence drift degree of a user group, and obtain a user clustering result. The fourth unit is configured to construct a label propagation network of a graph structure by taking the user as a node, take a similarity between users as a weight of an edge, iteratively calculate a label propagation probability considering a time sequence evolution feature of the node, perform label assignment on the user based on a first preset threshold, and obtain a label assignment result. The fifth unit is configured to generate a corresponding label set for each user based on the label assignment result, and construct a multi-level label structure tree according to a hierarchical relationship and an association degree between labels.

7. An electronic device, comprising: The processor is configured to call the instructions stored in the memory to execute the method in any one of claims 1 to 5. The computer program instructions are executed by the processor to implement the method in any one of claims 1 to 5. ​ ​ 8. A computer-readable storage medium having stored thereon computer program instructions, wherein, ​

Images