Customer portrait construction method and system based on time series network evolution analysis

Abstract

The present application relates to the technical field of data processing, and more particularly to a customer portrait construction method and system based on time series network evolution analysis, which obtains customer interaction data containing multiple time period interaction records, constructs a multi-layer time series network with customers as nodes, identifies an evolution sequence formed by arranging the neighbors of each time layer in time sequence to obtain a customer relationship evolution type, sets a different time period weight distribution scheme according to the evolution type, weights the interaction data to obtain evolution adapted interaction data, establishes a connection path between the node and the neighbor node, collects neighbor node interaction data and evolution adapted interaction data, aggregates the data to form a customer feature vector, calculates the conditional probability distribution difference of each feature dimension under different evolution types to select evolution sensitive feature dimensions, uses the evolution sensitive feature dimensions to filter and cluster to obtain a customer group and output a customer portrait, and effectively improves the accuracy of the customer portrait.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for constructing customer profiles based on time-series network evolution analysis. Background Technology

[0002] In the field of customer profiling, existing technologies typically rely on static customer data or behavioral records within a fixed time window for group segmentation. Common practices include collecting customer demographic information, historical transaction records, and clickstream data, extracting feature vectors using statistical methods or machine learning models, and then using clustering algorithms (such as K-means and hierarchical clustering) to generate customer groups. These methods often assume that customer behavioral characteristics remain stable over a period of time, or simply perform simple summation or averaging on data across time dimensions. For the analysis of customer relationship networks, existing technologies mostly use a single-point-in-time social network topology to calculate static indicators such as node centrality and tight centrality to supplement customer characteristics. Some improved solutions introduce time decay weights, assigning higher weights to recent interaction data, but the weight setting lacks an adaptive adjustment mechanism for dynamic changes in customer relationships.

[0003] Existing conventional approaches suffer from two main drawbacks. First, in characterizing customer relationship evolution, most methods treat customer networks as static structures, ignoring the significant changes in customer interactions over time—such as shifting from distant to close cooperation, or from high-frequency interaction to low-frequency contact. This static approach prevents customer profiles from reflecting relationship evolution trends, resulting in poor timeliness of clustering results based on historical data in subsequent operational scenarios. Second, regarding feature fusion and weight allocation, existing technologies employ uniform assumptions about feature importance across different customers, failing to consider the differentiated impact of customer relationship evolution types (such as relationship strengthening versus relationship weakening) on ​​feature discriminative power. The same feature dimension (such as transaction frequency) can contribute drastically to customer segmentation under different evolution types, but existing methods lack mechanisms to identify and filter these evolution-sensitive dimensions, leading to mixed clustering results and difficulty in accurately distinguishing customer groups with different evolutionary patterns. Summary of the Invention

[0004] This invention provides a method and system for constructing customer profiles based on temporal network evolution analysis, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for constructing customer profiles based on temporal network evolution analysis, comprising:

[0006] Acquire customer interaction data, which includes interaction records for multiple time periods;

[0007] A multi-layer time-series network is constructed using customers as nodes. The neighbor nodes of each customer node in each time layer are identified and arranged in chronological order to form a neighbor evolution sequence. The customer relationship evolution type is obtained by performing pattern recognition on the neighbor evolution sequence.

[0008] Different time period weight allocation schemes are set for different customer relationship evolution types, and the customer interaction data is weighted to obtain evolution adaptation interaction data;

[0009] In a multi-layer time-series network, a connection path is established between a customer node and its neighboring nodes at each time layer. Interaction data of neighboring nodes is collected along the connection path, and the interaction data of neighboring nodes is aggregated with the evolutionary adaptation interaction data to obtain a customer feature vector.

[0010] Calculate the conditional probability distribution differences of each feature dimension of the customer feature vector under different customer relationship evolution types, and select feature dimensions whose differences exceed a preset difference threshold as evolution-sensitive feature dimensions.

[0011] The customer feature vectors are filtered using evolution-sensitive feature dimensions. The filtered customer feature vectors are then clustered to obtain customer groups and output customer profiles.

[0012] Pattern recognition of the neighbor evolution sequence yields customer relationship evolution types, including:

[0013] Extract the appearance and disappearance times of each neighbor node in the neighbor evolution sequence to construct the life cycle range of each neighbor node;

[0014] Project the lifecycle intervals of each neighboring node onto the time axis to form lifecycle projection segments, calculate the overlapping area of ​​all lifecycle projection segments, and extract the continuous time period with the longest coverage in the overlapping area as the core relationship time period.

[0015] The ratio of the time span of the core relationship to the total time span of the neighbor evolution sequence is calculated to obtain the proportion of core relationships;

[0016] Extract the marginal time regions outside the core relationship time periods in the neighbor evolution sequence, count the number of new and disappearing neighbor nodes in the marginal time regions, and calculate the change in marginal relationship by subtracting the number of disappearing nodes from the number of new nodes.

[0017] Based on the numerical value of the proportion of core relationships and the numerical sign of the change in peripheral relationships, determine and output the customer relationship evolution type.

[0018] Different time-period weighting schemes are set for different customer relationship evolution types, and the customer interaction data is weighted to obtain evolutionary adaptation interaction data, including:

[0019] Multi-dimensional feature analysis is performed on the customer relationship evolution type to obtain the evolution feature vector, the coupling strength between the components of the evolution feature vector is calculated, and the evolution mode weight matrix is ​​constructed.

[0020] Extract the time span of customer interaction data, determine the number of time segments based on the dimensional structure of the evolution mode weight matrix, and evenly divide the time span into multiple time segments according to the number of time segments.

[0021] Calculate the distribution concentration of customer interaction data within each time segment, associate and map the distribution concentration with the evolution mode weight matrix to obtain the basic weight value of each time segment, and extract the volatility index of customer interaction data within each time segment to dynamically correct the basic weight value of each time segment to obtain the corrected weight value.

[0022] The time segments are combined with their corresponding adjusted weight values ​​to form a time period weight allocation scheme.

[0023] The timestamps of each interaction record in the customer interaction data are matched with the time segments in the time period weight allocation scheme to obtain the corrected weight value corresponding to each interaction record. The customer interaction data is then weighted using the corrected weight value to obtain the evolutionary adaptation interaction data.

[0024] Calculate the distribution concentration of customer interaction data within each time segment, correlate and map the distribution concentration with the evolutionary pattern weight matrix to obtain the basic weight value for each time segment, and extract the volatility index of customer interaction data within each time segment to dynamically adjust the basic weight value for each time segment to obtain the adjusted weight value, including:

[0025] Construct a multi-scale density distribution of customer interaction data within each time segment, calculate the spatial consistency metric of density peaks at different scales, and quantify the distribution concentration based on the spatial consistency metric.

[0026] The distribution concentration of each time segment is used to form a distribution concentration vector. The distribution concentration vector is multiplied by the evolution mode weight matrix to obtain a weight vector. Each component in the weight vector is extracted as the basic weight value of the corresponding time segment.

[0027] Calculate the moving average sequence of customer interaction data within each time segment, calculate the standard deviation of the moving average sequence, and use the product of the standard deviation and the distribution concentration as a volatility index.

[0028] The volatility index for each time segment is multiplied element-wise with the base weight value, and the product result is exponentially normalized to obtain the corrected weight value.

[0029] In a multi-layer time-series network, connection paths are established between customer nodes and their neighboring nodes at each time layer. Interaction data between neighboring nodes is collected along these connection paths. This neighboring node interaction data is then aggregated with evolutionary adaptation interaction data to obtain a customer feature vector, including:

[0030] In a multi-layer time-series network, the neighboring nodes of a client node at each time layer are divided into predecessor neighboring nodes and successor neighboring nodes.

[0031] For each predecessor neighbor node, a predecessor connection path is established from the predecessor neighbor node to the client node; for each successor neighbor node, a successor connection path is established from the client node to the successor neighbor node.

[0032] The preceding and subsequent connection paths are combined to form the connection path between the client node and its neighboring nodes at each time layer.

[0033] Collect interaction data of predecessor neighbor nodes along the predecessor connection path and mark it as influencing factor data; collect interaction data of successor neighbor nodes along the successor connection path and mark it as propagation effect data; combine the influencing factor data and propagation effect data to form neighbor node interaction data.

[0034] The correlation between influencing factor data and evolutionary adaptation interaction data, as well as the correlation between evolutionary adaptation interaction data and propagation effect data, are calculated and multiplied to obtain the transmission coefficient.

[0035] The influencing factor data and propagation effect data are assigned transmission weights using transmission coefficients. The neighbor node interaction data with transmission weights and the evolutionary adaptation interaction data are then aggregated to obtain the customer feature vector.

[0036] The calculation of the differences in conditional probability distributions of each feature dimension of the customer feature vector under different customer relationship evolution types includes:

[0037] Extract the sample value set of each feature dimension of the customer feature vector under each customer relationship evolution type, perform kernel density estimation on the sample value set to obtain the probability density function, calculate the relative entropy between the probability density functions under different customer relationship evolution types, and use the relative entropy as the distribution deviation.

[0038] Identify the transition path between customer relationship evolution types, extract customer samples and transition time nodes along the transition path where the evolution type transition occurs, calculate the transition amplitude of each feature dimension value of the customer sample before and after the transition time node, and multiply the transition amplitude with the distribution deviation of the corresponding feature dimension to obtain the weighted transition amplitude.

[0039] For each feature dimension, the weighted transition amplitudes of all customer samples belonging to the same customer relationship evolution type are grouped into an intra-class set. The mean of the weighted transition amplitudes of each intra-class set is calculated as the intra-class center. The distance between the intra-class centers of different customer relationship evolution types is calculated as the inter-class distance. The ratio of the inter-class distance to the variance of the weighted transition amplitudes of the intra-class set is used as the conditional probability distribution difference of the feature dimension.

[0040] Customer feature vectors are filtered using evolution-sensitive feature dimensions. The filtered customer feature vectors are then clustered to obtain customer groups and output customer profiles, including:

[0041] Construct a correlation matrix among evolution-sensitive feature dimensions, and identify evolution-sensitive feature dimension pairs in the correlation matrix whose correlation coefficients exceed a preset correlation coefficient threshold as redundant feature dimension pairs;

[0042] From each redundant feature dimension pair, the evolution-sensitive feature dimensions whose conditional probability distribution difference does not reach the preset difference lower limit are removed, and the feature values ​​corresponding to the remaining evolution-sensitive feature dimensions in the customer feature vector are retained to form the filtered customer feature vector.

[0043] Extract customer relationship evolution type labels for each customer, and calculate the mean of the filtered customer feature vectors of all customers with the customer relationship evolution type labels for each customer relationship evolution type as the evolution type cluster center corresponding to the customer relationship evolution type;

[0044] Calculate the distance between the filtered customer feature vector of each customer and the cluster center of each evolution type, and classify the customers corresponding to the evolution type cluster center with the smallest distance into the same customer group;

[0045] The feature distribution and customer relationship evolution type distribution of the filtered customer feature vectors within each customer group are statistically analyzed, and the feature distribution and evolution type distribution are used as customer profiles and output.

[0046] A second aspect of this invention provides a customer profile construction system based on temporal network evolution analysis, comprising:

[0047] A data acquisition unit is used to acquire customer interaction data, which includes interaction records for multiple time periods.

[0048] The evolution identification unit is used to construct a multi-layer time-series network with customers as nodes, identify the neighbor nodes of customer nodes in each time layer and arrange them in chronological order to form a neighbor evolution sequence, and perform pattern recognition on the neighbor evolution sequence to obtain the customer relationship evolution type.

[0049] The weight allocation unit is used to set different time period weight allocation schemes for different customer relationship evolution types, and to weight the customer interaction data to obtain evolution adaptation interaction data.

[0050] The feature aggregation unit is used to establish the connection path between the customer node and its neighbor nodes in each time layer in the multi-layer time series network, collect the neighbor node interaction data along the connection path, and aggregate the neighbor node interaction data with the evolutionary adaptation interaction data to obtain the customer feature vector.

[0051] The sensitive selection unit is used to calculate the difference in conditional probability distribution of each feature dimension of the customer feature vector under different customer relationship evolution types, and select the feature dimension whose difference exceeds the preset difference threshold as the evolution sensitive feature dimension.

[0052] The clustering profiling unit is used to filter customer feature vectors using evolution-sensitive feature dimensions, cluster the filtered customer feature vectors to obtain customer groups, and output customer profiles.

[0053] A third aspect of the present invention provides an electronic device, comprising:

[0054] processor;

[0055] Memory used to store processor-executable instructions;

[0056] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0057] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0058] This invention employs a multi-layer temporal network to dynamically model customer interaction data, accurately capturing the evolution patterns of customer relationships over time. By constructing neighbor evolution sequences and identifying relationship evolution types, it can distinguish the relationship change patterns of different customer groups, fundamentally improving the spatiotemporal adaptability of customer profiles. Establishing cross-time layer connection paths and aggregating neighbor interaction data imbues customer feature vectors with rich contextual information, avoiding the drawbacks of static models that lose temporal correlations. Weighting schemes are set for different customer relationship evolution types, weighting customer interaction data to make evolution-adapted interaction data more closely match the actual customer behavior trajectory. Collecting neighbor node interaction data along connection paths and aggregating it with evolution-adapted data significantly enhances the expressive power of feature vectors, especially providing a more refined portrayal of core customer groups with frequently changing relationships. This dynamic weighting and aggregation mechanism greatly improves the robustness of feature extraction. The conditional probability distribution differences of each dimension of the customer feature vector under different relationship evolution types are calculated, automatically selecting discriminative feature dimensions sensitive to evolution types. Using evolution-sensitive feature dimensions to filter and reduce the dimensionality of feature vectors effectively suppresses noise interference and highlights key behavioral indicators. Clustering the filtered feature vectors can form semantically coherent customer groups with clear boundaries, ultimately outputting customer profiles with significantly improved accuracy. Attached Figure Description

[0059] Figure 1 A flowchart illustrating the customer profile construction method based on temporal network evolution analysis;

[0060] Figure 2 Flowchart for generating customer feature vectors. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0063] Figure 1 This is a flowchart illustrating the customer profile construction method based on temporal network evolution analysis, as described in an embodiment of the present invention.

[0064] Customer profiling methods based on temporal network evolution analysis include:

[0065] Acquire customer interaction data, which includes interaction records for multiple time periods;

[0066] A multi-layer time-series network is constructed using customers as nodes. The neighbor nodes of each customer node in each time layer are identified and arranged in chronological order to form a neighbor evolution sequence. The customer relationship evolution type is obtained by performing pattern recognition on the neighbor evolution sequence.

[0067] Different time period weight allocation schemes are set for different customer relationship evolution types, and the customer interaction data is weighted to obtain evolution adaptation interaction data;

[0068] In a multi-layer time-series network, a connection path is established between a customer node and its neighboring nodes at each time layer. Interaction data of neighboring nodes is collected along the connection path, and the interaction data of neighboring nodes is aggregated with the evolutionary adaptation interaction data to obtain a customer feature vector.

[0069] Calculate the conditional probability distribution differences of each feature dimension of the customer feature vector under different customer relationship evolution types, and select feature dimensions whose differences exceed a preset difference threshold as evolution-sensitive feature dimensions.

[0070] The customer feature vectors are filtered using evolution-sensitive feature dimensions. The filtered customer feature vectors are then clustered to obtain customer groups and output customer profiles.

[0071] Pattern recognition of the neighbor evolution sequence yields customer relationship evolution types, including:

[0072] Extract the appearance and disappearance times of each neighbor node in the neighbor evolution sequence to construct the life cycle range of each neighbor node;

[0073] Project the lifecycle intervals of each neighboring node onto the time axis to form lifecycle projection segments, calculate the overlapping area of ​​all lifecycle projection segments, and extract the continuous time period with the longest coverage in the overlapping area as the core relationship time period.

[0074] The ratio of the time span of the core relationship to the total time span of the neighbor evolution sequence is calculated to obtain the proportion of core relationships;

[0075] Extract the marginal time regions outside the core relationship time periods in the neighbor evolution sequence, count the number of new and disappearing neighbor nodes in the marginal time regions, and calculate the change in marginal relationship by subtracting the number of disappearing nodes from the number of new nodes.

[0076] Based on the numerical value of the proportion of core relationships and the numerical sign of the change in peripheral relationships, determine and output the customer relationship evolution type.

[0077] After constructing the multi-layer temporal network and identifying neighbor nodes at each time layer, pattern recognition is needed to analyze the neighbor evolution sequence to determine the evolution type of customer relationships. The core idea of ​​pattern recognition is to start with the lifecycle of neighbor nodes, quantify the stability of relationships and the dynamic changes in edge regions, and ultimately form a classification judgment of customer relationship evolution patterns.

[0078] For each neighbor node in the neighbor evolution sequence, extract its first appearance and final disappearance times in the temporal network, defining the interval between these two times as the lifetime interval of that neighbor node. The appearance time corresponds to the earliest time layer in which the neighbor node establishes a connection with the client node in the temporal network, and the disappearance time corresponds to the last time layer in which the connection exists in the temporal network. If a neighbor node maintains a connection with the client node throughout the entire observation window, its lifetime interval is equal to the total time span of the entire neighbor evolution sequence. By performing the above extraction operation on all neighbor nodes, a set of lifetime intervals can be obtained, with each interval representing the duration of a client relationship.

[0079] Each interval in the aforementioned lifecycle interval set is projected onto a unified time axis, forming a corresponding lifecycle projection segment. The time axis uses the start time of the neighbor evolution sequence as the zero point and the timestamps of each time layer as the scale. After obtaining all projection segments, the overlapping areas between all segments are calculated, i.e., the time regions simultaneously covered by multiple lifecycle projection segments are identified. The overlapping area is calculated as follows: for each time point, the number of lifecycle projection segments covering that time point is counted, and the set of consecutive time points with a coverage number exceeding a preset minimum coverage threshold is considered as a candidate overlapping area. Among all candidate overlapping areas, the consecutive time period with the longest coverage length is extracted and defined as the core relationship time period. The core relationship time period represents the time interval in the neighbor evolution sequence where the customer node maintains a stable connection with the most neighbor nodes simultaneously, reflecting the most concentrated and stable core relationships in the customer relationship network.

[0080] After determining the time period of the core relationship, the ratio of its time span to the total time span of the neighboring evolutionary sequences is calculated to obtain the proportion of core relationships. Let the time span of the core relationship time period be denoted as . The total time span of the neighbor evolution sequence is The proportion of core relationships The calculation method is as follows . The range of values ​​is The larger the value, the more stable the customer relationship network is within the entire observation window, and the longer the core relationship occupies the time proportion; the smaller the value, the shorter the stable period of the customer relationship network is, and the relationship structure is more dispersed in the time dimension.

[0081] After determining the core relationship time period, the time regions outside the core relationship time period in the neighbor evolution sequence are extracted and defined as edge time regions. Edge time regions typically include the early time period before the core relationship time period and the late time period after it. Within the edge time regions, the number of newly added and disappeared neighbor nodes are counted. The number of newly added nodes refers to the total number of neighbor nodes that first appear within the edge time region, i.e., whose lifecycle start time falls within the edge time region, denoted as […]. The disappearance count refers to the total number of neighboring nodes that ultimately disappear within the edge time region, i.e., whose lifecycle ends within the edge time region, denoted as . Changes in marginal relationships The calculation method is as follows .when This indicates that more new relationships are created than disappear within the edge time zone, meaning customers are continuously expanding new relationship connections beyond their core relationships; when When this occurs, it indicates that more relationships disappear than are added within the edge time region, and the customer's edge connections outside of core relationships are in a state of contraction; when This indicates that the number of newly added and disappeared relationships is roughly equal, and the relationships in the marginal areas are in a relatively balanced state.

[0082] Based on the proportion of core relationships The magnitude of the value and the change in the marginal relationship The numerical sign is used to determine and output the customer relationship evolution type. Specifically, when... Larger and When the customer relationship network is stable in the long term, and new relationships continue to be added in peripheral areas, it is judged as "stable expansion type"; when Larger and When the customer relationship network is stable overall, but peripheral relationships are gradually shrinking, it is classified as "stable shrinkage type"; when Smaller and When the customer relationship core time is relatively short and there are many new relationships in the peripheral areas, the customer relationship network is in a rapid reconstruction or expansion phase, and is judged as "dynamic expansion type"; when Smaller and When this occurs, it indicates that the core customer relationship period is short, and peripheral relationships disappear simultaneously, indicating that the overall relationship network is in a state of disintegration or migration, and is classified as "dynamic decay type". The above judgment threshold can be adjusted according to the actual business scenario, for example, by adjusting the threshold. The threshold is set at 0.5, meaning that a core relationship ratio exceeding 50% is considered stable, while a ratio below 50% is considered dynamic.

[0083] In practical applications, taking financial customer behavior analysis as an example, a customer's interaction records over the past 12 months constitute a neighbor evolution sequence, with months 3 to 9 representing the core relationship period. , , Within the marginal time zones (months 1-2 and 10-12), , , Therefore, the customer is classified as "stable and expanding". This classification will directly affect the setting of subsequent time period weight allocation schemes, thereby affecting the quality of evolutionary adaptation interaction data generation and ultimately improving the accuracy of customer profiles.

[0084] Through the systematic pattern recognition process of the neighbor evolution sequence described above, the evolutionary characteristics of customer relationships can be quantified from the perspective of temporal structure, providing a reliable type basis for subsequent differentiated weight allocation and feature extraction.

[0085] Different time-period weighting schemes are set for different customer relationship evolution types, and the customer interaction data is weighted to obtain evolutionary adaptation interaction data, including:

[0086] Multi-dimensional feature analysis is performed on the customer relationship evolution type to obtain the evolution feature vector, the coupling strength between the components of the evolution feature vector is calculated, and the evolution mode weight matrix is ​​constructed.

[0087] Extract the time span of customer interaction data, determine the number of time segments based on the dimensional structure of the evolution mode weight matrix, and evenly divide the time span into multiple time segments according to the number of time segments.

[0088] Calculate the distribution concentration of customer interaction data within each time segment, associate and map the distribution concentration with the evolution mode weight matrix to obtain the basic weight value of each time segment, and extract the volatility index of customer interaction data within each time segment to dynamically correct the basic weight value of each time segment to obtain the corrected weight value.

[0089] The time segments are combined with their corresponding adjusted weight values ​​to form a time period weight allocation scheme.

[0090] The timestamps of each interaction record in the customer interaction data are matched with the time segments in the time period weight allocation scheme to obtain the corrected weight value corresponding to each interaction record. The customer interaction data is then weighted using the corrected weight value to obtain the evolutionary adaptation interaction data.

[0091] Developing differentiated time-period weighting schemes for different customer relationship evolution types is a core step in achieving accurate weighting of evolution-adaptive interaction data. Customers with different evolution types exhibit drastically different behavioral patterns over time. For example, the interaction data of stable customers is evenly distributed across time periods, while the interaction data of explosive customers is often highly concentrated within specific time windows. Therefore, it is necessary to perform multi-dimensional feature analysis on customer relationship evolution types, extract feature components that characterize the internal structure of evolution patterns, and then construct an evolution pattern weight matrix that can guide weight allocation.

[0092] When performing multi-dimensional feature analysis on customer relationship evolution types, evolutionary features are extracted from multiple perspectives, including time, intensity, and structure, forming an evolutionary feature vector. Time-dimensional features include evolution cycle length and peak time location; intensity-dimensional features include mean and variance of interaction frequency; and structural-dimensional features include replacement rate of neighbor nodes and distribution pattern of relationship duration. Integrating these features across all dimensions yields an evolutionary feature vector describing the inherent pattern of this evolutionary type. ,in The total number of feature dimensions, and each component It corresponds to a specific evolutionary characteristic index.

[0093] After obtaining the evolutionary eigenvectors, the coupling strength between each component is calculated. The coupling strength reflects the degree of correlation between different evolutionary feature dimensions; a higher coupling strength indicates a stronger synergistic effect between the two feature dimensions in describing the evolutionary pattern. Specifically, the Pearson correlation coefficient is calculated pairwise for each component of the evolutionary eigenvectors to obtain the coupling strength matrix. , of which elements Indicates the first The component and the first The coupling strength value between the components ranges from 0 to 1. arrive Based on the coupling strength matrix, components in the evolutionary feature vectors with coupling strength exceeding a preset coupling threshold are merged to eliminate the interference of redundant dimensions on weight allocation, ultimately constructing an evolutionary pattern weight matrix with simplified dimensions and a clear structure. The rows of the evolutionary pattern weight matrix correspond to the independent evolutionary feature dimensions after merging, and the columns correspond to the time segments to which weights are to be assigned. The matrix elements are initialized with the sensitivity scores of each independent evolutionary feature dimension to each time segment.

[0094] After extracting the time span of customer interaction data, the number of time segments is determined based on the dimensional structure of the evolutionary pattern weight matrix. Specifically, the number of columns in the evolutionary pattern weight matrix corresponds to the number of time segments. This number is adaptively determined by the evolutionary pattern weight matrix during its construction based on the complexity of the evolutionary type. The more complex the evolutionary pattern and the more feature dimensions it has, the greater the corresponding number of time segments. The larger the number of time segments, the better, to ensure that the interaction data within each time segment fully reflects the local characteristics of the evolutionary pattern. This is based on the number of time segments. Divide the total time span of customer interaction data evenly into 100 parts. A series of non-overlapping time segments Each time segment has the same length, which is the total time span divided by 1 / 2. .

[0095] Within each time segment, calculate the distribution concentration of customer interaction data. Distribution concentration measures the degree of clustering of customer interaction behavior within that time segment. Higher concentration indicates that interaction behavior is more concentrated within a specific sub-interval of that time period, while lower concentration indicates that interaction behavior is more evenly distributed within that time segment. When calculating distribution concentration, the time is segmented... The frequency of interactions within the data is statistically analyzed at a fine-grained temporal level to obtain a frequency distribution sequence, from which the Gini coefficient of the sequence is calculated. As an indicator of distribution concentration The range of values ​​is A higher value indicates a higher concentration. The distribution concentration is divided into time segments. Evolutionary pattern weight matrix The association mapping is performed by taking the first element of the evolution mode weight matrix. The elements of each row in the column are weighted and summed, with the component values ​​corresponding to each independent evolutionary feature dimension as the result. As a weight, we obtain the first The basic weight values ​​for each time segment This basic weight value comprehensively reflects the importance of this time segment in the current evolutionary model.

[0096] After determining the basic weight values, volatility indicators of customer interaction data within each time segment are further extracted to dynamically adjust the basic weight values. Volatility indicators characterize the magnitude of change in interaction data within a time segment; the standard deviation of the interaction frequency series can be used. As a volatility indicator, higher volatility indicates significant changes in customer behavior within that time segment, making the data more valuable for depicting evolutionary patterns, and thus the weight of that segment should be appropriately increased. Conversely, lower volatility indicates relatively stable customer behavior within that time segment, and the weight can be appropriately decreased. The specific method for dynamic adjustment is to adjust the normalized volatility indicator... With basic weight value Perform linear fusion to obtain corrected weight values. ,in and Let be the fusion coefficient, satisfying The values ​​of both are preset based on the time sensitivity of the evolutionary type. For time-sensitive evolutionary types, Taking a larger value enhances the correction effect on volatility; for time-stationary evolution types, Take a larger value to maintain the dominance of the base weight value.

[0097] Divide the time into segments With the corresponding corrected weight value Combined, a complete time-time weight allocation scheme is formed. The scheme stores data in key-value pairs, where the key is the start and end timestamps of the time segment, and the value is the corresponding adjusted weight value, facilitating rapid retrieval and matching later.

[0098] During the weighted operation, each interaction record in the customer interaction data is traversed, its timestamp is extracted, and the timestamp is matched with the start and end ranges of each time segment in the time period weighting scheme to determine the time segment to which the interaction record belongs. And obtain the corresponding corrected weight value. The corrected weight value is used as a product factor to weight the original feature values ​​of the interaction record, that is, the original feature values ​​are multiplied by... We obtain the weighted feature values. After performing the above weighting operation on all interaction records, we summarize them to obtain the evolutionary adaptation interaction data. While preserving the integrity of the original interaction information, the evolutionary adaptation interaction data highlights the interaction behaviors within a time period that are more representative of the current evolutionary pattern through differentiated weights, thus providing a higher quality data foundation for the subsequent construction of customer feature vectors.

[0099] In practical applications, the same customer may exhibit different evolutionary types at different historical periods. The above-mentioned weight allocation and weighting process can be performed separately for different periods. Then, the evolutionary adaptation interaction data of each period can be spliced ​​together in chronological order to form an evolutionary adaptation interaction dataset covering the complete historical cycle, ensuring that the customer profile can fully reflect the dynamic changes in customer behavior.

[0100] Calculate the distribution concentration of customer interaction data within each time segment, correlate and map the distribution concentration with the evolutionary pattern weight matrix to obtain the basic weight value for each time segment, and extract the volatility index of customer interaction data within each time segment to dynamically adjust the basic weight value for each time segment to obtain the adjusted weight value, including:

[0101] Construct a multi-scale density distribution of customer interaction data within each time segment, calculate the spatial consistency metric of density peaks at different scales, and quantify the distribution concentration based on the spatial consistency metric.

[0102] The distribution concentration of each time segment is used to form a distribution concentration vector. The distribution concentration vector is multiplied by the evolution mode weight matrix to obtain a weight vector. Each component in the weight vector is extracted as the basic weight value of the corresponding time segment.

[0103] Calculate the moving average sequence of customer interaction data within each time segment, calculate the standard deviation of the moving average sequence, and use the product of the standard deviation and the distribution concentration as a volatility index.

[0104] The volatility index for each time segment is multiplied element-wise with the base weight value, and the product result is exponentially normalized to obtain the corrected weight value.

[0105] After acquiring customer interaction data and segmenting it by time, fine-grained distribution characteristic analysis of the interaction data within each time segment is required to provide a reliable quantitative basis for subsequent weight allocation. For each time segment, a multi-scale density distribution of customer interaction data is constructed. Specifically, for the sequence of interaction events within the same time segment, kernel density estimation is performed at different time window scales (e.g., fine-grained hourly, medium-grained daily, and coarse-grained weekly) to obtain density distribution curves at multiple scales. At each scale, the peak positions of the density curves are located, and the spatial locations of the density peaks at different scales are recorded. The spatial consistency metric of these peak positions is then calculated. The spatial consistency metric reflects whether the concentrated areas of interaction behavior are stable and consistent across different observation scales: if the peak positions at each scale highly overlap, it indicates that the customer interaction data within that time segment has a significant central tendency; if the peak positions are dispersed, it indicates that the interaction behavior within that time segment is relatively uniform or random.

[0106] When quantifying distribution concentration based on spatial consistency metrics, the ratio of the mean of density peak locations at each scale to the average absolute deviation between each peak location is calculated. A larger ratio indicates a more concentrated distribution, meaning higher distribution concentration. Let the... The distribution concentration of each time segment is ,in , This represents the total number of time segments. In the specific implementation, after completing the above multi-scale density distribution construction and spatial consistency calculation for each time segment, the distribution concentration value corresponding to each time segment is obtained. The core advantage of this step is that it avoids the distortion of distribution characteristics caused by improper window selection under a single scale through multi-scale joint analysis, making the quantitative results of distribution concentration more robust to time window parameters.

[0107] Distribution concentration of each time segment Arrangement of distribution concentration vector ,in , No. Each component is... The distribution concentration vector Evolutionary pattern weight matrix Perform matrix multiplication to obtain the weight vector. ,Right now ,in , Extract the weight vector. The first in Each component As the first The basic weight values ​​for each time segment Evolutionary pattern weight matrix The evolutionary relationships between different time segments are encoded. Through matrix multiplication, the basic weight value of each time segment depends not only on its own distribution concentration, but also on the distribution concentration of adjacent or related time segments, thereby realizing the evolution-aware weight allocation across time periods.

[0108] After obtaining the basic weight values ​​for each time segment, a volatility indicator is introduced to dynamically adjust these weight values, capturing the instability of customer interaction data over time. For customer interaction data within each time segment, the moving average is calculated by sliding the window with a fixed step size, resulting in a moving average sequence. The length of the sliding window is adaptively set according to the granularity of the time segment; for example, when the time segment is monthly, the sliding window can be set to 7 days, sliding by 1 day at a time, thus obtaining a sequence reflecting the local mean change trend of interaction frequency within that month. The standard deviation of this moving average sequence is then calculated. This is used to measure the fluctuation range of interactive behavior within a given time segment. The standard deviation is... Distribution concentration of the corresponding time segments The product of these two factors serves as a volatility indicator for that time segment, denoted as... The implication of this design is that if the interaction behavior within a certain time segment is both concentrated (high distribution concentration) and highly volatile (high standard deviation), the volatility index of that segment will be amplified, thus having a more significant impact on the weight of that segment in subsequent correction steps; conversely, if the interaction behavior is concentrated but stable, the volatility index is smaller, and the correction magnitude of the basic weight value is also smaller.

[0109] Volatility indicators for different time segments With the corresponding basic weight value Perform element-wise multiplication to obtain the intermediate product. The intermediate product over all time segments. After performing exponential normalization, the final corrected weight values ​​are obtained. The calculation method for exponential normalization is as follows: ,in The index variable for time segmentation has a value range of 1. arrive The reason for using exponential normalization instead of nonlinear normalization is that the exponential function can amplify the differences in the product values ​​between different time segments, giving more prominent weight to time segments with higher volatility and concentration. This allows for greater attention to these key time segments in the subsequent weighting of customer interaction data, improving the sensitivity of evolutionary adaptation interaction data to nodes of change in customer relationships.

[0110] Adjust weight values The normalization constraint is satisfied, meaning the sum of the corrected weights for all time segments is 1. This ensures that no additional numerical scaling bias is introduced when weighting and aggregating customer interaction data. In practical applications, the corresponding evolutionary pattern weight matrix varies for different customer relationship evolution types. Because they differ, even if two customers have the same distribution concentration vector for their interaction data, the calculated base weight values ​​will differ under different evolution types, leading to different distributions of the corrected weight values. This design ensures that the weight allocation scheme is targeted to the customer relationship evolution type, avoiding the loss of profile accuracy for customers in different evolution modes caused by a "one-size-fits-all" weight allocation method.

[0111] Through the synergistic effect of the three stages—multi-scale distribution centralized quantification, evolutionary pattern-aware basic weight calculation, and volatility-driven dynamic correction—the final corrected weight value is obtained. It can comprehensively reflect the importance of each time segment in the evolution of customer relationships and the stability characteristics of its internal interaction behavior, laying a solid weighting foundation for subsequent accurate weighting of customer interaction data and generation of high-quality evolution-adaptive interaction data.

[0112] like Figure 2 As shown, Figure 2 This is a flowchart illustrating the customer feature vector generation process in an embodiment of the present invention.

[0113] In a multi-layer time-series network, connection paths are established between customer nodes and their neighboring nodes at each time layer. Interaction data between neighboring nodes is collected along these connection paths. This neighboring node interaction data is then aggregated with evolutionary adaptation interaction data to obtain a customer feature vector, including:

[0114] In a multi-layer time-series network, the neighboring nodes of a client node at each time layer are divided into predecessor neighboring nodes and successor neighboring nodes.

[0115] For each predecessor neighbor node, a predecessor connection path is established from the predecessor neighbor node to the client node; for each successor neighbor node, a successor connection path is established from the client node to the successor neighbor node.

[0116] The preceding and subsequent connection paths are combined to form the connection path between the client node and its neighboring nodes at each time layer.

[0117] Collect interaction data of predecessor neighbor nodes along the predecessor connection path and mark it as influencing factor data; collect interaction data of successor neighbor nodes along the successor connection path and mark it as propagation effect data; combine the influencing factor data and propagation effect data to form neighbor node interaction data.

[0118] The correlation between influencing factor data and evolutionary adaptation interaction data, as well as the correlation between evolutionary adaptation interaction data and propagation effect data, are calculated and multiplied to obtain the transmission coefficient.

[0119] The influencing factor data and propagation effect data are assigned transmission weights using transmission coefficients. The neighbor node interaction data with transmission weights and the evolutionary adaptation interaction data are then aggregated to obtain the customer feature vector.

[0120] In multi-layered temporal networks, each client node has connections with several neighboring nodes at different time layers. These neighboring nodes do not have a symmetrical impact on client behavior, but rather exhibit significant causal differences based on the temporal direction of the interaction. Specifically, the client node's neighbors at each time layer are divided into predecessor neighbors and successor neighbors: predecessor neighbors are those that interact with the client node before it in the time dimension, and their interactions have a potential driving or influencing effect on the client node; successor neighbors are those that interact with the client node after it in the time dimension, and their interactions reflect the diffusion or propagation effect of the client node's behavior. This directional division allows subsequent path establishment and data collection to fully capture the temporal causal structure of the interaction relationships.

[0121] For predecessor neighbor nodes, a predecessor connection path is established from the predecessor neighbor node to the customer node. The direction of the path reflects the semantics of "influence inflow," meaning that the behavioral information of the predecessor neighbor node flows along the path to the customer node, representing the process of external factors shaping customer behavior. For successor neighbor nodes, a successor connection path is established from the customer node to the successor neighbor node. The direction of the path reflects the semantics of "effect outflow," meaning that the behavioral information of the customer node flows along the path to the successor neighbor node, representing the process of customer behavior propagation and diffusion in social or business networks. The predecessor and successor connection paths together form a complete set of connection paths between the customer node and its neighbors at each time layer, thus forming a bidirectional path topology in the multi-layer time-series network centered on the customer node and taking into account both time inflow and outflow directions.

[0122] Interaction data of predecessor neighbor nodes is collected along the predecessor connection path and labeled as influencing factor data. This influencing factor data encompasses multi-dimensional information such as interaction frequency, interaction channel type, and interaction content category of predecessor neighbor nodes within the corresponding time layer, reflecting external interaction signals that exert an input influence on the behavior of the target customer node. Interaction data of successor neighbor nodes is collected along the successor connection path and labeled as propagation effect data. This propagation effect data also includes dimensions such as interaction frequency, channel distribution, and behavior category of successor neighbor nodes within the corresponding time layer, reflecting the responsive interaction signals triggered by the customer node's behavior in the network. Combining the influencing factor data and the propagation effect data forms complete neighbor node interaction data, providing a data foundation for subsequent propagation coefficient calculations.

[0123] The calculation of the transmission coefficient is the core step in this process. It quantifies the strength of the mutual transmission between influencing factor data and propagation effect data through the client node's own interaction data (i.e., evolutionary adaptation interaction data). Specifically, it calculates the correlation between influencing factor data and evolutionary adaptation interaction data, denoted as [missing information]. This correlation reflects the degree of association between the interaction behavior of the predecessor neighbor node and the client's own interaction behavior. A higher value indicates a more significant driving effect of the predecessor neighbor node on customer behavior. The correlation between evolutionary adaptation interaction data and propagation effect data is calculated and denoted as... This relevance reflects the degree of correlation between the customer's own interactive behavior and the response behavior of subsequent neighboring nodes. A higher value indicates a stronger propagation effect of customer behavior within the network. Multiplying the two correlation coefficients yields the propagation coefficient. ,Right now Conductivity The overall strength of the transmission chain, from "influence inflow - customer node - effect outflow," was comprehensively measured. The larger the value, the more prominent the mediating role of the customer node in the network, and the higher the weight should be given to the contribution of its neighbor node interaction data to the customer feature vector.

[0124] The correlation above can be calculated using methods such as Pearson correlation coefficient or cosine similarity, with the choice made flexibly based on the distribution characteristics of the data. In practical applications, the influencing factor data, propagation effect data, and evolutionary adaptation interaction data have all undergone time-time weighting (i.e., evolutionary adaptation processing). Therefore, the correlation calculation can fully reflect the time-series weight differences under different customer relationship evolution types, avoiding correlation estimation bias caused by uneven distribution of time-time weights.

[0125] Using the conductivity coefficient Transmission weights are assigned to the influencing factor data and the propagation effect data, respectively. Specifically, the influencing factor data is multiplied by... After obtaining the weighted influencing factor data, multiply the propagation effect data by... By obtaining weighted propagation effect data, neighbor node interaction data with higher propagation link strength is given a larger proportion in subsequent aggregation, while neighbor node interaction data with weaker propagation links has a correspondingly reduced impact on the aggregation result. This mechanism effectively avoids the information noise problem caused by equally weighted aggregation of all neighbor node interaction data, making the customer feature vector more accurately reflect the true role and behavioral characteristics of customer nodes in the network.

[0126] The neighbor node interaction data (including weighted influencing factor data and weighted propagation effect data), after being assigned propagation weights, is aggregated with the evolutionary adaptation interaction data to obtain a customer feature vector. The aggregation method can be concatenation or weighted summation, depending on the input format requirements of the downstream task (such as clustering or classification). In the concatenation method, the evolutionary adaptation interaction data, weighted influencing factor data, and weighted propagation effect data are concatenated as independent feature blocks into a high-dimensional vector, preserving the independent semantic information of each part. In the weighted summation method, the three parts are linearly combined according to their respective weight coefficients to generate a low-dimensional, compact feature vector, suitable for subsequent efficient clustering processing. Regardless of the aggregation method used, the resulting customer feature vector simultaneously includes the customer's own temporal weighted interaction information, influencing factor information from predecessor neighbor nodes, and propagation effect information flowing to successor neighbor nodes. This achieves a comprehensive and accurate characterization of customer behavior characteristics within the context of a multi-layer temporal network structure, providing a high-quality feature representation foundation for subsequent evolution-sensitive feature dimension selection and customer group segmentation.

[0127] The calculation of the differences in conditional probability distributions of each feature dimension of the customer feature vector under different customer relationship evolution types includes:

[0128] Extract the sample value set of each feature dimension of the customer feature vector under each customer relationship evolution type, perform kernel density estimation on the sample value set to obtain the probability density function, calculate the relative entropy between the probability density functions under different customer relationship evolution types, and use the relative entropy as the distribution deviation.

[0129] Identify the transition path between customer relationship evolution types, extract customer samples and transition time nodes along the transition path where the evolution type transition occurs, calculate the transition amplitude of each feature dimension value of the customer sample before and after the transition time node, and multiply the transition amplitude with the distribution deviation of the corresponding feature dimension to obtain the weighted transition amplitude.

[0130] For each feature dimension, the weighted transition amplitudes of all customer samples belonging to the same customer relationship evolution type are grouped into an intra-class set. The mean of the weighted transition amplitudes of each intra-class set is calculated as the intra-class center. The distance between the intra-class centers of different customer relationship evolution types is calculated as the inter-class distance. The ratio of the inter-class distance to the variance of the weighted transition amplitudes of the intra-class set is used as the conditional probability distribution difference of the feature dimension.

[0131] When calculating the differences in conditional probability distributions of each feature dimension of a customer feature vector, it is necessary to proceed sequentially from three levels: probability density estimation, transition amplitude calculation, and construction of inter-class and intra-class statistics.

[0132] For each feature dimension in the customer feature vector, customer samples are grouped according to customer relationship evolution type. All sample values ​​for that feature dimension under each evolution type are extracted, forming the corresponding sample value set. Since sample values ​​in real-world scenarios often exhibit non-parametric distribution characteristics, it is not suitable to directly apply parametric assumptions such as Gaussian distribution. Therefore, a kernel density estimation method is used to fit the probability density function to each sample value set. Kernel density estimation places a kernel function (usually a Gaussian kernel) centered on each sample point. The smoothness is controlled by a bandwidth parameter. After superimposing all kernel functions and normalizing, a continuously differentiable probability density function is obtained. The bandwidth selection can be adaptively determined according to the Silverman criterion to avoid overfitting due to too small a bandwidth or obscuring distribution details due to too large a bandwidth. After performing the above kernel density estimation for each customer relationship evolution type, the probability density function of that feature dimension under each evolution type can be obtained.

[0133] After obtaining the probability density functions for different evolutionary types, the relative entropy between any two evolutionary types is calculated to quantify the degree of deviation between the two probability distributions. Let the evolutionary types be... Lower feature dimension The probability density function is Evolutionary type The probability density function is as follows Then from arrive The relative entropy is ,in Indicates evolution type and Between in feature dimension The relative entropy on, For feature value variables, and These are the probability density functions for that feature dimension under the corresponding evolutionary type. Since relative entropy lacks symmetry, in practice, we take... and The mean value serves as the deviation of the symmetric distribution between the two evolutionary types. To eliminate directional influences, when multiple evolutionary types exist, the mean or maximum value of the symmetric relative entropy for all type pairs is taken to obtain the feature dimension. Overall distribution deviation The larger the value, the more significant the difference in the distribution of this feature dimension among different evolutionary types.

[0134] Based on the calculation of distribution deviation, we further identify the conversion paths between customer relationship evolution types. A conversion path refers to a directed path by which a customer changes from one evolution type to another, such as from a "stable relationship" to a "rapid expansion" relationship. Along each conversion path, we select customer samples that have undergone evolution type conversion along that path and record the time point of each customer's conversion. For each customer who converted, extract the values ​​of their feature dimensions prior to the conversion time point. Feature dimension values ​​after the transformation time point Calculate the absolute value of the difference between the two as the customer's feature dimension. The jump amplitude ,in This represents a single customer sample in terms of feature dimensions. The magnitude of the transition. The magnitude of the transition reflects the degree of change in the value of this feature dimension when a customer undergoes an evolutionary type transformation. The larger the magnitude of the transition, the more sensitive the feature dimension is to the evolutionary type transformation.

[0135] The weighted transition magnitude is obtained by multiplying the transition magnitude by the distribution deviation of the corresponding feature dimension. ,in Indicates the weighted transition magnitude. For feature dimensions The significance of introducing distribution deviation as a weight lies in the fact that if the probability distribution of a certain feature dimension differs very little under different evolutionary types, even if its transition amplitude is large, it may just be random fluctuation rather than a true evolution-sensitive signal; conversely, if the distribution deviation is large, the magnitude of the transition amplitude has a stronger distinguishing significance. Weighted transition amplitude integrates the two aspects of information, making subsequent feature selection more statistically reliable.

[0136] For each feature dimension, the weighted transition magnitudes of all customer samples belonging to the same customer relationship evolution type are aggregated to form the within-class set under that evolution type. ,in Indicates the evolution type index, Indicates the feature dimension index. For collections within a class. The average of all weighted transition amplitudes is taken to obtain the feature dimension of this evolution type. In-class center on ,in Indicates the number of samples in the collection within the class. Indicates the in-class center. The in-class center represents the feature dimension of the customer sample under this evolution type. The average transition strength level on.

[0137] After obtaining the intra-class centers of each evolutionary type, the intra-class center distances between different evolutionary types are calculated, i.e., the inter-class distances. For feature dimensions... If it exists Evolutionary types, interclass distance Defined as the mean of the absolute values ​​of the intra-class central differences among all type pairs: ,in The total number of evolutionary types, and For indexes of different evolutionary types, Representing feature dimension The inter-class distance. The larger the inter-class distance, the more significant the difference in the weighted transition magnitude of customers of different evolutionary types on this feature dimension, and the more capable this feature dimension is of distinguishing different evolutionary types.

[0138] At the same time, calculate the collection within each class. Weighted transition amplitude variance Then, the mean of the variances of all evolution types is taken to obtain the feature dimension. within-class variance mean ,in Indicates evolution type Lower feature dimension within-class set variance This represents the mean of the intra-class variance. The mean of the intra-class variance reflects the dispersion of the weighted transition amplitude of customer samples within the same evolutionary type. The smaller the intra-class variance, the more consistent the behavior of customers of the same type.

[0139] Finally, the ratio of inter-class distance to the mean of intra-class variance is defined as the feature dimension. Conditional probability distribution difference score ,in Representing feature dimension The conditional probability distribution difference score, To prevent extremely small positive numbers with a denominator of zero. The physical meaning is similar to the ratio of inter-class divergence to intra-class divergence in Fisher's discriminant criterion: a larger numerator indicates higher distinguishability between different evolutionary types, while a smaller denominator indicates more compactness within the same class. Both factors together determine the ability of this feature dimension to distinguish customer relationship evolutionary types. Compare with a preset difference threshold and select Feature dimensions exceeding the threshold are used as evolution-sensitive feature dimensions for subsequent screening and clustering of customer feature vectors, thereby improving the accuracy of customer profiles.

[0140] Customer feature vectors are filtered using evolution-sensitive feature dimensions. The filtered customer feature vectors are then clustered to obtain customer groups and output customer profiles, including:

[0141] Construct a correlation matrix among evolution-sensitive feature dimensions, and identify evolution-sensitive feature dimension pairs in the correlation matrix whose correlation coefficients exceed a preset correlation coefficient threshold as redundant feature dimension pairs;

[0142] From each redundant feature dimension pair, the evolution-sensitive feature dimensions whose conditional probability distribution difference does not reach the preset difference lower limit are removed, and the feature values ​​corresponding to the remaining evolution-sensitive feature dimensions in the customer feature vector are retained to form the filtered customer feature vector.

[0143] Extract customer relationship evolution type labels for each customer, and calculate the mean of the filtered customer feature vectors of all customers with the customer relationship evolution type labels for each customer relationship evolution type as the evolution type cluster center corresponding to the customer relationship evolution type;

[0144] Calculate the distance between the filtered customer feature vector of each customer and the cluster center of each evolution type, and classify the customers corresponding to the evolution type cluster center with the smallest distance into the same customer group;

[0145] The feature distribution and customer relationship evolution type distribution of the filtered customer feature vectors within each customer group are statistically analyzed, and the feature distribution and evolution type distribution are used as customer profiles and output.

[0146] After screening the evolution-sensitive feature dimensions, it's not feasible to directly include all of them in the clustering process. This is because some evolution-sensitive feature dimensions may exhibit strong linear correlations. Retaining these highly correlated dimensions simultaneously would lead to redundant representations in the feature space, causing the clustering results to favor these redundant directions and reducing the accuracy of customer profile differentiation. Therefore, it's necessary to first perform redundancy removal on the set of evolution-sensitive feature dimensions.

[0147] For the selected set of evolution-sensitive feature dimensions, a correlation matrix is ​​constructed between each pair of these dimensions. Each element in the correlation matrix represents the Pearson correlation coefficient between the corresponding two feature dimensions across the entire customer sample. When the absolute value of the correlation coefficient between two evolution-sensitive feature dimensions exceeds a preset correlation coefficient threshold, the dimension pair consisting of these two dimensions is marked as a redundant feature dimension pair. The preset correlation coefficient threshold can be flexibly set according to actual business needs, typically between 0.85 and 0.95, to ensure that only highly linearly correlated redundant dimensions are removed, while dimensions with independent information content are retained.

[0148] For each identified redundant feature dimension pair, it is necessary to determine which dimension should be removed. The criterion is the difference score of the conditional probability distribution calculated in the previous step for each dimension. In a redundant feature dimension pair, if the difference score of the conditional probability distribution of a certain evolution-sensitive feature dimension does not reach a preset lower limit, that dimension is considered relatively redundant and removed; if the scores of both dimensions do not reach the preset lower limit, both are removed; if the scores of both dimensions reach the preset lower limit, the dimension with the higher score is retained, and the dimension with the lower score is removed. This processing logic ensures that while eliminating redundancy, it maximizes the retention of feature dimensions that can distinguish customer relationship evolution types. After redundancy removal, the feature values ​​corresponding to the retained set of evolution-sensitive feature dimensions are extracted from the original customer feature vector and concatenated to form a filtered customer feature vector. This vector has lower dimensionality and higher information density, making it more suitable for subsequent cluster analysis.

[0149] The clustering process incorporates customer relationship evolution type labels as prior information to achieve guided cluster initialization. Customer relationship evolution type labels obtained during the multi-layer time-series network analysis phase are extracted for each customer. For each customer relationship evolution type, the filtered customer feature vectors of all customers with that type label are collected. The mean of these vectors is calculated dimension-wise to obtain the cluster centers corresponding to that evolution type. Let the total number of evolution types be... Then a total of This initialization method utilizes known evolutionary type cluster centers. Compared to traditional random initialization, it fully leverages known evolutionary type information, making the initial positions of cluster centers closer to the actual distribution of customer groups and effectively avoiding the instability of clustering results caused by random initialization.

[0150] In obtaining After identifying the cluster centers for each evolutionary type, calculate the Euclidean distance between the filtered customer feature vector and each evolutionary type cluster center for each customer. Let the customer... The filtered customer feature vector is , No. The cluster centers for each evolutionary type are Then the customer To the The distance between the cluster centers is For each customer, find out what makes... Cluster center index that yields the minimum value This customer was classified as the first The system divides customers into customer groups. Once all customers have been grouped, the mean of the customer feature vectors within each group is recalculated, and the positions of the cluster centers for each evolution type are updated. This distance calculation and center update process is repeated iteratively until the displacement of each cluster center is lower than a preset convergence threshold or the maximum number of iterations is reached, thus obtaining a stable customer group partitioning result.

[0151] It is worth noting that during the iteration process, the evolutionary type label of some customers may not be entirely consistent with the customer group they ultimately belong to. This phenomenon is reasonable because the customer feature vector contains aggregated information of interaction data from neighboring nodes, and its position in the feature space may deviate from the initial cluster centers based purely on the evolutionary type label. The clustering iteration process is actually a process of adaptively adjusting the boundaries of customer groups under the guidance of prior information on evolutionary types. The final customer groups reflect both the evolutionary patterns of customer relationships and the actual distribution characteristics of customers in the multidimensional feature space.

[0152] After customer groups are segmented, statistical analysis is performed on each group to generate customer profiles. For each customer group, the distribution of the filtered customer feature vectors across all feature dimensions is statistically analyzed, including descriptive statistics such as mean, variance, and quantile intervals for each dimension, forming a feature distribution description for that group. Simultaneously, the frequency and proportion of each customer relationship evolution type tag within the group are statistically analyzed, forming a customer relationship evolution type distribution description for that group. The feature distribution and evolution type distribution are combined to construct the customer profile for that customer group and output in a structured format.

[0153] The output customer profiles can characterize customer groups from two dimensions: First, the feature distribution reveals the typical characteristic ranges of this customer group in terms of interaction behavior and preference tendencies, providing direct feature basis for downstream applications such as precision marketing and personalized recommendations; second, the evolution type distribution reveals the relationship evolution patterns of this customer group, helping to predict the future behavioral trends of the customer group and providing a dynamic perspective for the formulation of customer relationship management strategies. The combination of these two types of information means that the customer profile not only describes the current static characteristics of customers, but also embeds the dynamic patterns of customer relationship evolution over time, significantly improving the accuracy and practical value of the customer profile.

[0154] A second aspect of this invention provides a customer profile construction system based on temporal network evolution analysis, comprising:

[0155] A data acquisition unit is used to acquire customer interaction data, which includes interaction records for multiple time periods.

[0156] The evolution identification unit is used to construct a multi-layer time-series network with customers as nodes, identify the neighbor nodes of customer nodes in each time layer and arrange them in chronological order to form a neighbor evolution sequence, and perform pattern recognition on the neighbor evolution sequence to obtain the customer relationship evolution type.

[0157] The weight allocation unit is used to set different time period weight allocation schemes for different customer relationship evolution types, and to weight the customer interaction data to obtain evolution adaptation interaction data.

[0158] The feature aggregation unit is used to establish connection paths between a client node and its neighbor nodes at each time layer in a multi-layer time-series network, collect neighbor node interaction data along the connection paths, and aggregate the neighbor node interaction data with evolutionary adaptation interaction data to obtain a client feature vector.

[0159] The sensitive selection unit is used to calculate the difference in conditional probability distribution of each feature dimension of the customer feature vector under different customer relationship evolution types, and select the feature dimension whose difference exceeds the preset difference threshold as the evolution sensitive feature dimension.

[0160] The clustering profiling unit is used to filter customer feature vectors using evolution-sensitive feature dimensions, cluster the filtered customer feature vectors to obtain customer groups, and output customer profiles.

[0161] A third aspect of the present invention provides an electronic device, comprising:

[0162] processor;

[0163] Memory used to store processor-executable instructions;

[0164] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0165] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0166] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A customer profile construction method based on temporal network evolution analysis, characterized in that, include: Acquire customer interaction data, which includes interaction records for multiple time periods; A multi-layer time-series network is constructed using customers as nodes. The neighbor nodes of each customer node in each time layer are identified and arranged in chronological order to form a neighbor evolution sequence. The customer relationship evolution type is obtained by performing pattern recognition on the neighbor evolution sequence. Different time period weight allocation schemes are set for different customer relationship evolution types, and the customer interaction data is weighted to obtain evolution adaptation interaction data; In a multi-layer time-series network, a connection path is established between a customer node and its neighboring nodes at each time layer. Interaction data of neighboring nodes is collected along the connection path, and the interaction data of neighboring nodes is aggregated with the evolutionary adaptation interaction data to obtain a customer feature vector. Calculate the conditional probability distribution differences of each feature dimension of the customer feature vector under different customer relationship evolution types, and select feature dimensions whose differences exceed a preset difference threshold as evolution-sensitive feature dimensions. The customer feature vectors are filtered using evolution-sensitive feature dimensions, and the filtered customer feature vectors are clustered to obtain customer groups and output customer profiles. Pattern recognition of the neighbor evolution sequence yields customer relationship evolution types, including: Extract the appearance and disappearance times of each neighbor node in the neighbor evolution sequence to construct the life cycle range of each neighbor node; Project the lifecycle intervals of each neighboring node onto the time axis to form lifecycle projection segments, calculate the overlapping area of ​​all lifecycle projection segments, and extract the continuous time period with the longest coverage in the overlapping area as the core relationship time period. The ratio of the time span of the core relationship to the total time span of the neighbor evolution sequence is calculated to obtain the proportion of core relationships; Extract the marginal time regions outside the core relationship time periods in the neighbor evolution sequence, count the number of new and disappearing neighbor nodes in the marginal time regions, and calculate the change in marginal relationship by subtracting the number of disappearing nodes from the number of new nodes. Based on the numerical value of the proportion of core relationships and the numerical sign of the change in peripheral relationships, determine and output the customer relationship evolution type; Different time-period weighting schemes are set for different customer relationship evolution types, and the customer interaction data is weighted to obtain evolutionary adaptation interaction data, including: Multi-dimensional feature analysis is performed on the customer relationship evolution type to obtain the evolution feature vector, the coupling strength between the components of the evolution feature vector is calculated, and the evolution mode weight matrix is ​​constructed. Extract the time span of customer interaction data, determine the number of time segments based on the dimensional structure of the evolution mode weight matrix, and evenly divide the time span into multiple time segments according to the number of time segments. Calculate the distribution concentration of customer interaction data within each time segment, associate and map the distribution concentration with the evolution mode weight matrix to obtain the basic weight value of each time segment, and extract the volatility index of customer interaction data within each time segment to dynamically correct the basic weight value of each time segment to obtain the corrected weight value. The time segments are combined with their corresponding adjusted weight values ​​to form a time period weight allocation scheme. The timestamps of each interaction record in the customer interaction data are matched with the time segments in the time period weight allocation scheme to obtain the corrected weight value corresponding to each interaction record. The customer interaction data is then weighted using the corrected weight value to obtain the evolutionary adaptation interaction data.

2. The method according to claim 1, characterized in that, Calculate the distribution concentration of customer interaction data within each time segment, correlate and map the distribution concentration with the evolutionary pattern weight matrix to obtain the basic weight value for each time segment, and extract the volatility index of customer interaction data within each time segment to dynamically adjust the basic weight value for each time segment to obtain the adjusted weight value, including: Construct a multi-scale density distribution of customer interaction data within each time segment, calculate the spatial consistency metric of density peaks at different scales, and quantify the distribution concentration based on the spatial consistency metric. The distribution concentration of each time segment is used to form a distribution concentration vector. The distribution concentration vector is multiplied by the evolution mode weight matrix to obtain a weight vector. Each component in the weight vector is extracted as the basic weight value of the corresponding time segment. Calculate the moving average sequence of customer interaction data within each time segment, calculate the standard deviation of the moving average sequence, and use the product of the standard deviation and the distribution concentration as a volatility index. The volatility index for each time segment is multiplied element-wise with the base weight value, and the product result is exponentially normalized to obtain the corrected weight value.

3. The method according to claim 1, characterized in that, In a multi-layer time-series network, connection paths are established between customer nodes and their neighboring nodes at each time layer. Interaction data between neighboring nodes is collected along these connection paths. This neighboring node interaction data is then aggregated with evolutionary adaptation interaction data to obtain a customer feature vector, including: In a multi-layer time-series network, the neighboring nodes of a client node at each time layer are divided into predecessor neighboring nodes and successor neighboring nodes. For each predecessor neighbor node, a predecessor connection path is established from the predecessor neighbor node to the client node; for each successor neighbor node, a successor connection path is established from the client node to the successor neighbor node. The preceding and subsequent connection paths are combined to form the connection path between the client node and its neighboring nodes at each time layer. Collect interaction data of predecessor neighbor nodes along the predecessor connection path and mark it as influencing factor data; collect interaction data of successor neighbor nodes along the successor connection path and mark it as propagation effect data; combine the influencing factor data and propagation effect data to form neighbor node interaction data. The correlation between influencing factor data and evolutionary adaptation interaction data, as well as the correlation between evolutionary adaptation interaction data and propagation effect data, are calculated and multiplied to obtain the transmission coefficient. The influencing factor data and propagation effect data are assigned transmission weights using transmission coefficients. The neighbor node interaction data with transmission weights and the evolutionary adaptation interaction data are then aggregated to obtain the customer feature vector.

4. The method according to claim 1, characterized in that, The calculation of the differences in conditional probability distributions of each feature dimension of the customer feature vector under different customer relationship evolution types includes: Extract the sample value set of each feature dimension of the customer feature vector under each customer relationship evolution type, perform kernel density estimation on the sample value set to obtain the probability density function, calculate the relative entropy between the probability density functions under different customer relationship evolution types, and use the relative entropy as the distribution deviation. Identify the transition path between customer relationship evolution types, extract customer samples and transition time nodes along the transition path where the evolution type transition occurs, calculate the transition amplitude of each feature dimension value of the customer sample before and after the transition time node, and multiply the transition amplitude with the distribution deviation of the corresponding feature dimension to obtain the weighted transition amplitude. For each feature dimension, the weighted transition amplitudes of all customer samples belonging to the same customer relationship evolution type are grouped into an intra-class set. The mean of the weighted transition amplitudes of each intra-class set is calculated as the intra-class center. The distance between the intra-class centers of different customer relationship evolution types is calculated as the inter-class distance. The ratio of the inter-class distance to the variance of the weighted transition amplitudes of the intra-class set is used as the conditional probability distribution difference of the feature dimension.

5. The method according to claim 1, characterized in that, Customer feature vectors are filtered using evolution-sensitive feature dimensions. The filtered customer feature vectors are then clustered to obtain customer groups and output customer profiles, including: Construct a correlation matrix among evolution-sensitive feature dimensions, and identify evolution-sensitive feature dimension pairs in the correlation matrix whose correlation coefficients exceed a preset correlation coefficient threshold as redundant feature dimension pairs; From each redundant feature dimension pair, the evolution-sensitive feature dimensions whose conditional probability distribution difference does not reach the preset difference lower limit are removed, and the feature values ​​corresponding to the remaining evolution-sensitive feature dimensions in the customer feature vector are retained to form the filtered customer feature vector. Extract customer relationship evolution type labels for each customer, and calculate the mean of the filtered customer feature vectors of all customers with the customer relationship evolution type labels for each customer relationship evolution type as the evolution type cluster center corresponding to the customer relationship evolution type; Calculate the distance between the filtered customer feature vector of each customer and the cluster center of each evolution type, and classify the customers corresponding to the evolution type cluster center with the smallest distance into the same customer group; The feature distribution and customer relationship evolution type distribution of the filtered customer feature vectors within each customer group are statistically analyzed, and the feature distribution and evolution type distribution are used as customer profiles and output.

6. A customer profiling system based on temporal network evolution analysis, used to implement the method as described in any one of claims 1-5, characterized in that, include: A data acquisition unit is used to acquire customer interaction data, which includes interaction records for multiple time periods; The evolution identification unit is used to construct a multi-layer time-series network with customers as nodes, identify the neighbor nodes of customer nodes in each time layer and arrange them in chronological order to form a neighbor evolution sequence, and perform pattern recognition on the neighbor evolution sequence to obtain the customer relationship evolution type. The weight allocation unit is used to set different time period weight allocation schemes for different customer relationship evolution types, and to weight the customer interaction data to obtain evolution adaptation interaction data. The feature aggregation unit is used to establish the connection path between the customer node and its neighbor nodes in each time layer in the multi-layer time series network, collect the neighbor node interaction data along the connection path, and aggregate the neighbor node interaction data with the evolutionary adaptation interaction data to obtain the customer feature vector. The sensitive selection unit is used to calculate the difference in conditional probability distribution of each feature dimension of the customer feature vector under different customer relationship evolution types, and select the feature dimension whose difference exceeds the preset difference threshold as the evolution sensitive feature dimension. The clustering profiling unit is used to filter customer feature vectors using evolution-sensitive feature dimensions, cluster the filtered customer feature vectors to obtain customer groups, and output customer profiles.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 5.

Images