Insurance customer feature vector updating method based on space-time correlation dynamic adjustment

Abstract

The invention belongs to the technical field of feature updating in the insurance field, and particularly relates to an insurance customer feature vector updating method based on space-time relevance dynamic adjustment. According to the method, client core data, external spatio-temporal data and real-time data streams are collected, a spatio-temporal incidence matrix is established, and client feature vectors are enhanced from time and space dimensions through dynamic feature extraction; on the basis, weighting the features by using time attenuation and space attenuation functions, and adjusting the importance of the space-time dimension by using a multi-head attention mechanism to ensure the timeliness and space adaptability of the features; further, through an event-driven and gradually-changed updating mechanism, the updating frequency is controlled, a versioned feature vector is generated, and finally, attenuation parameters are adjusted through an optimization strategy, so that the feature updating effect and the service decision accuracy are improved; the problem of insufficient consideration of space-time relevance in a traditional method is effectively solved, and the real-time performance and the accuracy of insurance decision making are enhanced.

Description

Technical Field

[0001] This application belongs to the technical field of insurance domain feature update, and more specifically, relates to a method for updating insurance customer feature vectors based on dynamic adjustment of spatio-temporal correlation. Background Art

[0002] With the continuous development of technology, the insurance industry is gradually moving towards an intelligent era driven by data. Traditional insurance operations rely on manual evaluation and fixed rules, while modern insurance companies increasingly rely on big data, artificial intelligence, and machine learning technologies to improve efficiency, reduce risks, optimize the customer experience, and promote refined management. The insurance industry is facing a huge influx of information flow, customer data, and event data. How to extract effective information from the vast amount of data for accurate risk assessment and pricing is a key issue in the industry transformation.

[0003] Insurance companies often need to dynamically adjust customers' insurance policies, pricing, and claims processes based on multiple factors such as customers' historical information, real-time dynamics, and external environment. However, customers' behaviors, market environments, and external conditions are constantly changing, which makes traditional analysis methods based on static data unable to meet the needs of refined management and timely decision-making. In particular, the importance of timeliness and spatial factors in the insurance business is being increasingly emphasized. For example, customers' lifestyles, changes in geographical locations, and social hot events all have a profound impact on the insurance business.

[0004] Currently, there are already some data analysis-based feature extraction and update methods in the insurance industry. These methods usually focus on extracting fixed features from customers' historical data for static analysis. For example, some models use traditional machine learning algorithms (such as decision trees, SVMs, etc.) for customer risk assessment and classification. These models analyze customers' past behavior data (such as insurance records, claims records, etc.) to predict customers' future behaviors or risk situations.

[0005] However, traditional methods mostly focus on single-dimensional data analysis (for example, only analyzing based on time or space), and fail to fully consider the spatio-temporal variability of customers' behaviors. In fact, the changes in customers' behaviors in time and space are highly correlated. For example, customers' consumption behaviors and claim frequencies are significantly affected by the external environment (such as holidays, economic changes, geographical locations, etc.). Summary of the Invention

[0006] The present invention provides a method for updating insurance customer feature vectors based on dynamic adjustment of spatio-temporal correlation, aiming to solve the technical problem of insufficient consideration of spatio-temporal correlation in the insurance field.

[0007] A method for updating insurance customer feature vectors based on dynamic adjustment of spatio-temporal correlation includes the following steps:

[0008] Step 1: Collect customer core data, external spatio-temporal data, and real-time data streams, standardize the spatio-temporal coordinates of the collected data, and unify them into a standard grid system; attach spatio-temporal tags to each business event to ensure that each event clearly corresponds to a time and a spatial location, and set timeliness tags for the data to obtain a spatio-temporal correlation matrix between customers and events;

[0009] Step 2: Based on the spatio-temporal correlation matrix and the customer historical feature vector, perform dynamic feature extraction, conduct periodic feature analysis from the time dimension, then calculate the event density change rate to quantify the change trend within a short time window, and extract the entropy value of the movement pattern and the migration of spatial hotspots from the spatial dimension to obtain a spatio-temporal enhanced feature tensor, including spatio-temporal features and corresponding timeliness tags;

[0010] Step 3: Perform attenuation processing on time and space based on the spatio-temporal enhanced feature tensor, use time attenuation and space attenuation functions to weight timeliness and spatial distance respectively, then adopt a multi-head attention mechanism to calculate the importance weights of spatio-temporal dimensions, and dynamically adjust the feature vector to obtain a feature vector with spatio-temporal weights;

[0011] Step 4: Based on the feature vector with spatio-temporal weights, construct an event-driven and gradual update mechanism, set spatio-temporal attenuation thresholds to control the update frequency, and balance the changes of the feature vector and time information through a combined update strategy to generate a versioned feature vector and record the update log;

[0012] Step 5: Monitor the impact of feature updates on core business indicators, analyze the relationship between feature freshness and business indicators, and use a policy optimizer to dynamically adjust the time attenuation parameter and the update strategy to obtain optimized update strategy parameters.

[0013] The present invention collects customer core data, external spatio-temporal data, and real-time data streams, establishes a spatio-temporal correlation matrix, and enhances the customer feature vector from the time and space dimensions through dynamic feature extraction; on this basis, uses time attenuation and space attenuation functions to weight the features, and utilizes a multi-head attention mechanism to adjust the importance of spatio-temporal dimensions to ensure the timeliness and spatial adaptability of the features; further, through an event-driven and gradual update mechanism, controls the update frequency and generates a versioned feature vector, and finally adjusts the attenuation parameter through an optimization strategy to improve the effect of feature updates and the accuracy of business decisions; effectively overcomes the problem of insufficient consideration of spatio-temporal correlation in traditional methods, and enhances the real-time performance and accuracy of insurance decisions.

[0014] Preferably, step 1 includes the following steps:

[0015] Data acquisition and preprocessing: Collect customer core data, external spatio-temporal data, and real-time data, and clean and preprocess the collected data;

[0016] Space-time coordinate standardization and gridification: Convert each data point into an H3 code:

[0017] Add additional space-time tags: Obtain the timestamp of the event, and convert the geographical location of the event into a unique identifier through the H3 grid code to obtain the space-time tag of the event;

[0018] Set timeliness tags: Add timeliness tags to each piece of data and weight them according to the time decay function:

[0019]

[0020] In the formula: D(t) represents the timeliness decay value; T now represents the current time; T represents the timestamp of the data record; λ represents the decay parameter;

[0021] Among them, the type of timeliness tag is set to three categories according to the timeliness decay value, including static data, quasi-real-time data, and real-time data;

[0022] Generate a space-time correlation matrix: Combine the space-time tags, timeliness tags, and customer historical feature vectors of each event to construct a space-time correlation matrix between customers and events.

[0023] Preferably, step 2 includes the following steps:

[0024] Time dimension feature extraction: Use Fourier transform to process the time series data in the space-time correlation matrix, and then extract the frequency components higher than the preset threshold based on the results of the Fourier transform to obtain periodic features, including daily periodicity, weekly periodicity, and monthly periodicity features;

[0025] Event density change rate: Based on the event occurrence time in the space-time correlation matrix, calculate the density change of events in a short time window to quantify the fluctuation trend of the event occurrence frequency:

[0026]

[0027] In the formula: N(t1) represents the number of events in the time period t1; N(t0) represents the number of events in the time period t0; t1 - t0 represents the length of the time window; R(t) represents the event density change rate;

[0028] Spatial entropy feature extraction: Based on the event space tags in the space-time correlation matrix, divide the spatial distribution into multiple spatial units and calculate the spatial entropy H(S):

[0029]

[0030] In the formula: p iRepresents the occurrence probability of an event in a spatial unit, where N i represents the number of events in the i-th spatial unit; N total represents the total number of events in all spatial units; n represents the number of spatial units;

[0031] Spatial hotspot migration feature extraction: Use a hotspot detection method to identify hotspot regions in space and calculate the migration trajectory of the hotspot regions over time:

[0032] C h (t) = ||C h (t1) - C h (t0)||;

[0033] In the formula: C h (t0) represents the coordinates of the center of the hotspot region at time t0; C h (t1) represents the coordinates of the center of the hotspot region at time t1; C h (t) represents the spatial migration distance of the hotspot center during the time period t;

[0034] By calculating the migration distances between different time points, the migration patterns of spatial hotspots are identified;

[0035] Based on time dimension feature extraction, event density change rate extraction, spatial entropy feature extraction, and spatial hotspot migration feature extraction, a spatio-temporal enhanced feature tensor is obtained.

[0036] Preferably, the step 2 further includes spatio-temporal interaction feature extraction, and the features extracted by spatio-temporal interaction feature extraction are used as part of the spatio-temporal enhanced feature tensor. The spatio-temporal interaction feature extraction includes spatio-temporal co-occurrence pattern extraction and cross-region correlation strength extraction;

[0037] Spatio-temporal co-occurrence pattern extraction: Calculate the co-occurrence degree of events under different spatio-temporal conditions to identify the spatio-temporal interaction relationship of events, that is, for events at different times and spatial positions, calculate the spatio-temporal co-occurrence degree:

[0038]

[0039] In the formula: represents whether event E i and event E j co-occur at time t k and in space S k , if they occur, it is 1, otherwise it is 0; N represents the total number of time windows calculated; C ij (t) represents the correlation between event E i and event E j in space-time;

[0040] Cross-regional association strength extraction: For events E in different regions A and B A and event E B , calculate the cross-regional association strength:

[0041]

[0042] where: N A and N B represent the number of events in regions A and B; represents event E A (i) and E B (j) occur simultaneously; C AB represents the spatio-temporal association strength across regions A and B.

[0043] Preferably, step 3 includes the following steps:

[0044] Time decay function:

[0045]

[0046] where: t0 represents the time when the event occurs; t represents the current time; λ t represents the time decay factor; e represents the base of the natural logarithm;

[0047] Spatial decay function:

[0048]

[0049] where: d represents the spatial distance between events; λ s represents the spatial decay factor;

[0050] Multi-head attention mechanism: Suppose the spatio-temporal enhanced tensor T is divided into h sub-spaces, that is, each head of the multi-head attention mechanism processes a part of T h , and the calculation process is as follows:

[0051] Input queries, keys, and values, calculate the attention scores based on the input queries, keys, and the dimension of the keys, and then obtain the weighted representation based on the calculated attention score weighted value matrix; connect the multi-head outputs of the multi-head attention mechanism to obtain the overall feature representation;

[0052] Spatio-temporal weight adjustment: Obtain the feature vector with spatio-temporal weights based on the overall feature representation output by the multi-head attention mechanism and the values of the time decay function and the spatial decay function:

[0053] T weighted = MultiHeadAttention(T)·f t (t)·f s (d);

[0054] In the formula: MultiHeadAttention(T) represents the overall feature representation output by the multi-head attention mechanism; T weighted represents the feature vector with spatio-temporal weights.

[0055] Preferably, the event-driven update mechanism is as follows:

[0056] Set the time trigger update condition to δ(t). When a major event occurs, the feature vector is updated:

[0057] T new = T old + α·δ(t)·ΔT event ;

[0058] In the formula: T new represents the updated feature vector with spatio-temporal weights; T old represents the feature vector with spatio-temporal weights before update; α represents the update intensity factor; δ(t) represents the event trigger flag function, which is 1 if a major event occurs, otherwise 0; ΔT event represents the feature change amount of the spatio-temporal weight brought by the event;

[0059] Among them, the judgment of the major event is based on business rules, and the feature change amount is calculated based on business rules.

[0060] Preferably, the gradual update mechanism is as follows:

[0061] T gradual = T old + γ(t)·ΔT gradual ;

[0062] In the formula: T gradual represents the feature vector with spatio-temporal weights after gradual update; γ(t) represents the gradual change factor, which is calculated according to the product of the time decay function and the space decay function; ΔT gradual represents the feature change of the spatio-temporal weight caused by small changes, which is calculated according to the cumulative effect of small events.

[0063] Preferably, the specific steps for controlling the update frequency based on the spatio-temporal decay threshold are as follows:

[0064] For each feature t i , calculate the timeliness decay value and compare it with the spatio-temporal threshold:

[0065] t i (t) = T(t)·f t (t)·f s (d);

[0066] where: T(t) represents the feature representation output by the multi-head attention mechanism; f t (t) represents the time decay function; f s (d) represents the spatial decay function; t i (t) represents the decay value of feature t with spatio-temporal weights i ;

[0067] When t i (t) is less than the threshold θ i , an update is triggered, and the logic for triggering the update is as follows:

[0068] If there is a major event, execute the event determination update mechanism; if there is no major event, execute the gradual update mechanism.

[0069] Preferably, step 4 further includes feature versioning and historical traceability; a version control mechanism is used to record the changes of the feature vector with spatio-temporal weights. Each feature update generates a new feature version number, and the audit log of the update is recorded. Through version control, historical backtracking and gray release strategies are realized.

[0070] Preferably, step 5 uses Bayesian optimization to optimize the time decay parameter and the update strategy, where Bayesian optimization maximizes the business effect of the model and the freshness of the features, and comprehensively optimizes the time decay factor, the spatial decay factor, and the feature update frequency.

[0071] The beneficial effects of the present invention include:

[0072] By collecting customer core data, external spatio-temporal data, and real-time data streams, the present invention establishes a spatio-temporal association matrix, and enhances the customer feature vector from the time and space dimensions through dynamic feature extraction; on this basis, time decay and spatial decay functions are used to weight the features, and the multi-head attention mechanism is used to adjust the importance of the spatio-temporal dimensions to ensure the timeliness and spatial adaptability of the features; further, through the event-driven and gradual update mechanisms, the update frequency is controlled and a versioned feature vector is generated. Finally, the decay parameter is adjusted through an optimization strategy to improve the effect of feature update and the accuracy of business decisions; effectively overcoming the problem of insufficient consideration of spatio-temporal correlation in traditional methods, and enhancing the real-time and accuracy of insurance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0074] Figure 1 This is the overall step block diagram provided by the embodiment of the present invention.

[0075] Figure 2 This is the specific step block diagram of step 2 provided by the embodiment of the present invention.

[0076] Figure 3 This is the schematic diagram of the steps of step 3 provided by the embodiment of the present invention. Detailed implementation manners

[0077] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0078] See Figure 1 As shown, the dynamic risk pricing optimization method for insurance products with multi-factor integration includes the following steps:

[0079] The method for updating the insurance customer feature vector dynamically adjusted based on spatio-temporal correlation includes the following steps:

[0080] Step 1: Collect customer core data, external spatio-temporal data and real-time data streams, standardize the spatio-temporal coordinates of the collected data, and unify them into a standard grid system; attach spatio-temporal tags to each business event to ensure that each event clearly corresponds to a time and a spatial location, and set timeliness tags for the data to obtain a spatio-temporal correlation matrix between customers and events;

[0081] The said step 1 includes the following steps:

[0082] Data acquisition and preprocessing: Collect customer core data, external spatio-temporal data and real-time data, and clean and preprocess the collected data;

[0083] Among them, the core data includes customer policy information, claim records, customer activity trajectories (such as location changes, device login logs, etc.);

[0084] The external spatio-temporal data includes weather data, traffic flow information, social event data (such as natural disasters, holidays, etc.), and geographic information system data;

[0085] The real-time data streams include Internet of Things device data, mobile terminal behavior event data, location tracking data, etc.;

[0086] Among them, data cleaning and preprocessing include abnormal data processing, data standardization, etc., which are all conventional data processing means, so they will not be elaborated in detail in the embodiments;

[0087] Spatio-temporal coordinate standardization and gridification: Since the spatio-temporal coordinate sources of data are different, all spatio-temporal coordinates need to be unified into the same standard grid system. Considering the accuracy and real-time requirements of geographical locations, the H3 grid system is adopted. Through the H3 grid system, it is divided into multiple hexagonal grid cells of different sizes, which is convenient for multi-level and high-precision representation of spatial data; we convert each data point (such as a geographical location point) into an H3 code:

[0088] H3_ID = H3_geoToH3(lat, lon, resolution);

[0089] where: lat and lon represent the original geographical coordinates; resolution represents the level of the H3 grid; H3_ID represents the unique H3 network identifier corresponding to each geographical location;

[0090] Adding spatio-temporal tags: Obtain the timestamp of the event (for example, assume that an insurance claim event occurs at time T, and this time T represents the timestamp of the event), and convert the geographical location of the event into a unique identifier through the H3 grid code to obtain the spatio-temporal tag of the event. The spatio-temporal identifier of the event includes the timestamp of the event and the H3 network identifier;

[0091] Setting timeliness tags: Add timeliness tags to each piece of data and weight them according to the time decay function:

[0092]

[0093] In the formula: D(t) represents the timeliness decay value; T now represents the current time; T represents the timestamp of the data record; λ represents the decay parameter;

[0094] Among them, the type of timeliness tag is set to three categories according to the timeliness decay value, including static data, quasi-real-time data, and real-time data;

[0095] It should be noted that here we screen out useful data through timeliness. For example, for static data, we can define it as having a smaller D(t), and the data no longer has significant timeliness. That is to say, we can set a threshold interval, and within this threshold interval, it belongs to static data, which can be 0 - 0.2; or 0.1 - 0.2; within this interval range, we classify it as static data; similarly, the calibration methods for quasi-real-time data and real-time data also refer to the calibration method of the static data;

[0096] Generating a spatio-temporal correlation matrix: Through the above steps, we obtain the spatio-temporal tag, timeliness tag, and customer historical feature vector, and combine them to form a spatio-temporal correlation matrix between the customer and the event. Exemplarily, the representation of the spatio-temporal correlation matrix is as follows:

[0097] A i,j = [C i , E j , T j , S j , D j ;

[0098] Wherein: A i,j represents the spatio-temporal association between the i-th customer and the j-th event; Ci represents the historical feature vector of the i-th customer (such as age, gender, purchase history, etc.); E j represents the business feature of the j-th event (such as insurance type, amount, etc.); T j represents the time label of event j; S j represents the spatial label of event j (i.e., network identifier); D j represents the timeliness label of event j (validity calculated through the timeliness decay function);

[0099] In this embodiment, through H3 grid processing, the geographical data is unified into a high-precision grid system. By combining the historical features of customers with the spatio-temporal features of events, a spatio-temporal association matrix of customers and events is constructed, which can not only effectively integrate data from different sources, but also ensure the freshness and timeliness of data, laying a foundation for subsequent spatio-temporal feature extraction and dynamic feature modeling.

[0100] Step 2: Perform dynamic feature extraction based on the spatio-temporal association matrix and the customer historical feature vector. Conduct periodic feature analysis from the time dimension, then calculate the event density change rate to quantify the change trend within a short time window. Extract the entropy value of the movement pattern and the migration of spatial hotspots from the spatial dimension to obtain a spatio-temporal enhanced feature tensor, including spatio-temporal features and corresponding timeliness labels;

[0101] See Figure 2 shown, the said Step 2 includes the following steps:

[0102] Time dimension feature extraction: Use Fourier transform to process the time series data in the spatio-temporal association matrix, and then extract the frequency components higher than the preset threshold based on the result of the Fourier transform to obtain periodic features, including daily periodicity, weekly periodicity, and monthly periodicity features;

[0103] Specifically as follows: Apply Fourier transform to the time series data T j (event time) to obtain the frequency domain representation X(f):

[0104]

[0105] Where: X(f) represents the frequency component in the frequency domain, which is the intensity of the time signal at different frequencies; f represents the frequency, indicating the periodicity of the event time; T represents the total length of the event time series; e -2πift represents the complex exponential function, and e represents the base of the natural logarithm; t represents the time variable, which is a certain moment in the time series;

[0106] Extract the frequency components higher than the preset threshold based on the results of the Fourier transform to obtain periodic features, including daily periodicity, weekly periodicity, and monthly periodicity features;

[0107] Event density change rate: Based on the event occurrence time in the spatio-temporal correlation matrix, calculate the density change of events in a short time window to quantify the fluctuation trend of the event occurrence frequency:

[0108]

[0109] Where: N(t1) represents the number of events within the time period t1; N(t0) represents the number of events within the time period t0; t1 - t0 represents the length of the time window; R(t) represents the density change rate of the event;

[0110] Spatial entropy feature extraction: Based on the event space label S in the spatio-temporal correlation matrix j , divide the spatial distribution into multiple spatial units and calculate the spatial entropy H(S):

[0111]

[0112] Where: p i represents the occurrence probability of events in the spatial unit, where N i represents the number of events in the i-th spatial unit; N total represents the total number of events in all spatial units; n represents the number of spatial units;

[0113] Spatial hot spot migration feature extraction: Use the hot spot detection method to identify the hot spot areas in the space and calculate the migration trajectory of the hot spot areas over time:

[0114] C h (t) = ||C h (t1) - C h (t0)||;

[0115] Where: C h (t0) represents the coordinates of the center of the hot spot area at time t0; C h (t1) represents the coordinates of the center of the hot spot area at time t1; C h (t) represents the spatial migration distance of the hot spot center during the time period t;

[0116] In this embodiment, the calculation of spatial entropy can reveal the regularity of spatial distribution, while the hot spot migration pattern can capture the dynamic interaction between regions, enhancing the timeliness of spatial dimension features;

[0117] Spatiotemporal co-occurrence pattern extraction: Calculate the co-occurrence degree of events under different spatiotemporal conditions to identify the spatiotemporal interaction relationship of events, that is, for events at different times and spatial positions, calculate the spatiotemporal co-occurrence degree:

[0118]

[0119] In the formula: represents event E i and event E j whether they co-occur at time t k and in space S k , if they occur, it is 1, otherwise it is 0; N represents the total number of time windows calculated; C ij (t) represents the correlation between event E i and event E j in spacetime;

[0120] Cross-regional correlation strength extraction: For events E A and event E B in different regions A and B, calculate the cross-regional correlation strength:

[0121]

[0122] In the formula: N A and N B represent the number of events in regions A and B; represents event E A (i) and E B (j) whether they occur simultaneously; C AB represents the spatiotemporal correlation strength across regions A and B;

[0123] Construct a spatiotemporal enhanced feature tensor T based on periodic features, event density changes, spatial entropy, hot spot migration patterns, spatiotemporal co-occurrence degrees, and cross-regional correlation strengths; Exemplarily:

[0124] Spatiotemporal enhanced feature tensor T, with dimensions [N clients , N events , N features , where N clients represents the number of customers; N events represents the number of events; N features represents the number of spatiotemporal features related to each event and the customer;

[0125] The spatio-temporal characteristics of each customer and event will form a three-dimensional tensor, where each dimension represents the customer, the event, and the spatio-temporal characteristics respectively. For example:

[0126] T i,j = concat(X periodic , R(t), H(S), D h (t), C ij (t), C AB );

[0127] In the formula: T i，j represents the spatio-temporal feature vector of the i-th customer and the j-th event; X periodic represents the time periodicity feature;

[0128] In this embodiment, the spatio-temporal co-occurrence features are extracted to reveal the spatio-temporal associations between events, which helps to identify potential business associations; secondly, through the extraction of cross-regional association strength features, the potential associations between different geographical regions are identified, which is very important for understanding how events in different regions affect each other, migrate, and how business activities expand across regions. The cross-regional association strength can not only reveal the dependencies between different regions, but also help to build a more spatio-temporally aware business prediction model.

[0129] Step 3: Perform attenuation processing on time and space based on the spatio-temporal enhanced feature tensor, use the time attenuation and space attenuation functions to weight the timeliness and spatial distance respectively, then adopt the multi-head attention mechanism to calculate the importance weights of the spatio-temporal dimensions, and dynamically adjust the feature vectors to obtain the feature vectors with spatio-temporal weights;

[0130] See Figure 3 shown, the said Step 3 includes the following steps:

[0131] Time attenuation function: Assume that the time when an event occurs is t0, and the current time is t, then the timeliness of this event decays exponentially over time, so the time attenuation function is expressed as:

[0132]

[0133] In the formula: t0 represents the time when the event occurs; t represents the current time; λ t represents the time attenuation factor; e represents the base of the natural logarithm;

[0134] Space attenuation function: Let the spatial distance between two events be d, then the influence of this event weakens as the distance increases, and the space attenuation function is expressed as:

[0135]

[0136] In the formula: d represents the spatial distance between events; λs Represents the spatial attenuation factor;

[0137] Multi-head attention mechanism: Suppose the spatio-temporal enhanced tensor T is divided into h subspaces, that is, each head of the multi-head attention mechanism processes a part of T, and the calculation process is as follows: h For the following:

[0138] Input queries, keys, and values, calculate the attention scores based on the input queries, keys, and the dimension of the keys, and then weight the value matrix based on the calculated attention scores to obtain a weighted representation; connect the multi-head outputs of the multi-head attention mechanism by concatenating the results of each head to obtain an overall feature representation;

[0139] Exemplary:

[0140] Query matrix: Where Is the query weight matrix for each head;

[0141] Key matrix: Where Is the key weight matrix for each head;

[0142] Value matrix: Where Is the value weight matrix for each head;

[0143] For each head h, calculate the attention score A h :

[0144]

[0145] In the formula: d k Represents the dimension of the key; Represents the transpose of the key matrix; the softmax function converts the relevance score of each event to between 0 and 1;

[0146] Weighted summation: According to the attention score weighted matrix V h , obtain the weighted representation Attention h :

[0147] Attention h = A h V h ;

[0148] Multi-head output merging: Finally, the multi-head output channels concatenate the results of each head to obtain an overall feature representation:

[0149] MultiHeadAttention(T) = concat(Attention1, Attention2,..., Attention h )Wo ;

[0150] Where: W o represents the merged output weight matrix;

[0151] Space-time weight adjustment: Based on the overall feature representation output by the multi-head attention mechanism and the values of the time decay function and the space decay function, a feature vector with space-time weights is obtained:

[0152] T weighted = MultiHeadAttention(T)·f t (t)·f s (d);

[0153] Where: MultiHeadAttention(T) represents the overall feature representation output by the multi-head attention mechanism; T weighted represents the feature vector with space-time weights.

[0154] In this embodiment, by introducing the time decay and space decay functions, the adaptability of the model to time and space changes is enhanced, so that the influence of timeliness and space distance can be dynamically reflected in the weighting of the feature vector. And the multi-head attention mechanism is used to process space-time features, which can dynamically adjust the weights of space-time features from multiple angles, improve the expression ability of the model for complex space-time interactions. Finally, by combining the time decay and space decay with the multi-head attention output, it ensures more refined modeling of the space-time features of the model, making the dynamic modeling more accurate.

[0155] Step 4: Based on the feature vector with space-time weights, construct an event-driven and incremental update mechanism, and at the same time set a space-time decay threshold to control the update frequency, and balance the changes of the feature vector and time information through a combined update strategy to generate a versioned feature vector and record the update log;

[0156] The event-driven update mechanism is as follows:

[0157] Set the time trigger update condition as δ(t). When a major event occurs, trigger the update of the feature vector:

[0158] T new = T old + α·δ(t)·ΔT event ;

[0159] Where: T new represents the updated feature vector with space-time weights; T old represents the feature vector with space-time weights before update; α represents the update intensity factor; δ(t) represents the event trigger flag function, which is 1 if a major event occurs, otherwise 0; ΔT eventRepresents the characteristic change amount of the spatio-temporal weight brought by the event;

[0160] Among them, the judgment of the major event is based on business rules (for example, in the insurance industry, possible major events include: major claims, accident occurrence, large transactions, etc.);

[0161] The characteristic change amount is calculated based on business rules, for example:

[0162] For example, if a large claim event occurs, then the characteristic value related to the claim amount may change significantly; the change amount can be calculated in the following way:

[0163] ΔT event = f(claim amount, accident type, time window, etc.);

[0164] It should be noted that f is a function for calculating the change amount according to business logic. For each characteristic, the change amount can be an increment, a proportional change, or a new value after feature update.

[0165] The gradient-based update mechanism is as follows:

[0166] T gradual = T old + γ(t)·ΔT gradual ;

[0167] In the formula: T gradual represents the feature vector with spatio-temporal weight after gradient-based update; γ(t) represents the gradient factor, which is calculated according to the product of the time decay function and the space decay function; ΔT gradual represents the change in the feature with spatio-temporal weight caused by small changes, and is calculated according to the cumulative effect of small events.

[0168] The specific steps for controlling the update frequency based on the spatio-temporal decay threshold are as follows:

[0169] For each feature t i , calculate the timeliness decay value and compare it with the spatio-temporal threshold:

[0170] t i (t) = T(t)·f t (t)·f s (d);

[0171] In the formula: T(t) represents the feature representation output by the multi-head attention mechanism; f t (t) represents the time decay function; f s (d) represents the space decay function; t i (t) represents the decay value of the feature t with spatio-temporal weight i ;

[0172] When t i (t) is less than the threshold θ i an update is triggered, and the logic for triggering the update is as follows:

[0173] If there is a major event, an event determination update mechanism is executed; if there is no major event, a gradual update mechanism is executed.

[0174] Step 4 further includes feature versioning and historical traceability; a version control mechanism is adopted to record the changes of the feature vectors with spatio-temporal weights. Each feature update generates a new feature version number, and an audit log of the update is recorded. Through version control, historical backtracking and gray release strategies are implemented;

[0175] Exemplarily, each feature update generates a new feature version number V n , and an audit log of the update is recorded, including information such as the update time, update event, attenuation factor, etc. Through version control, historical backtracking and gray release strategies can be supported;

[0176] For each feature version V n , the historical state is represented by the following formula:

[0177] V n = V n-1 + ΔF update with timestamp t n ;

[0178] In the formula: V n represents the nth version of the feature; ΔF update represents the change of the feature in this update; t n represents the timestamp of the update;

[0179] An audit log is generated each time a feature is updated, recording information such as the timeliness and spatial impact update strategy during the update. The audit log can help analyze the impact of feature changes on the model in the later stage. Especially in the case of feature drift, it can timely backtrack and take corrective measures;

[0180] The representation of the audit log record is as follows:

[0181] L n = <V n , t n , α, λ t , λ s , update_reason>;

[0182] Where: L n represents the nth audit log; α is the update intensity factor; λ trepresents the time decay factor; λ s represents the space decay factor; update_reason represents the trigger reason for this update;

[0183] In this embodiment, by combining event-driven and gradual update mechanisms, and cooperating with spatio-temporal decay threshold control, the flexibility and real-time performance of dynamic feature updates are achieved; version control and audit log recording ensure that the system can trace historical features and support gray release and rollback operations; thus, both the timeliness of the feature vector is ensured, and overly frequent updates are avoided, improving the stability and performance of the system

[0184] Step 5: Monitor the impact of feature updates on core business metrics, analyze the relationship between feature freshness and business metrics, and use a policy optimizer to dynamically adjust the time decay parameter and update policy to obtain optimized update policy parameters;

[0185] In step 5, Bayesian optimization is used to optimize the time decay parameter and update policy, where Bayesian optimization comprehensively optimizes the time decay factor, space decay factor, and feature update frequency to maximize the business effect of the model and the freshness of features;

[0186] Exemplarily:

[0187] Construct an objective function based on the time decay factor, space decay factor, and feature update frequency;

[0188] Assume that the goal is to maximize the business effect of the model (e.g., underwriting accuracy rate, claim approval efficiency, etc.) and the freshness of features, then the objective function is as follows:

[0189] f(λ t , λ s , frequency) = γ1·Δf accuracy (λ t , λ s , frequency) + γ2·Δf freshness (λ t , λ s , frequency) - δ·penalty(λ t , λ s , frequency);

[0190] In the formula: λ t and λ s respectively represent the time decay factor and the space decay factor; frequency represents the feature update frequency; Δf accuracy represents the improvement of the business effect after feature update, and the specific index can be the improvement of the underwriting accuracy rate; Δf freshnessIndicates the improvement in feature freshness, reflecting the improvement in model effectiveness after feature update; penalty represents the penalty term for over-updating or under-updating of features; δ represents the penalty factor;

[0191] Exemplarily: the said Δf freshness Consider the following factors:

[0192] Data novelty: Measures the similarity between newly collected data and existing data. The new feature data can cover more unseen customer behaviors or events;

[0193] Model response: Observe the prediction performance of the model on new data after using new features, such as the improvement in metrics such as prediction accuracy, precision, recall, etc. of the model;

[0194] Time window: Regularly evaluate the impact of updated features on model performance within a certain time window to ensure that features are updated in a timely manner and reflect the latest trends;

[0195] The calculation formula is:

[0196] Δf freshness = Freshness Score(t new ) - Freshness Score(t old );

[0197] In the formula: Freshness Score represents the feature freshness score calculated based on the above factors; t new and t old respectively represent the timestamps of the new feature data and the old feature data;

[0198] The said Δf accuracy is measured based on the following business metrics:

[0199] Underwriting accuracy: Whether the accuracy of underwriting decisions has improved after the introduction of new features;

[0200] Claims approval efficiency: Whether the new features have accelerated the claims approval process and reduced the approval time;

[0201] Customer satisfaction: Evaluate the impact of new features on the customer experience through customer feedback or satisfaction scores;

[0202] Δf accuracy = Accuracy Score new - Accuracy Score old ;

[0203] In the formula: Accuracy Score new and Accuracy Score oldrespectively represent the new accuracy score and the old accuracy frequency score calculated based on business metrics;

[0204] The specific calculation formula of the penalty is as follows:

[0205] penalty = α·(Update Frequency Penalty) + β·(Stability Penalty) + γ·(Cost Penalty);

[0206] In the formula: α, β, and γ all represent weight factors; Update Frequency Penalty represents the update frequency; Stability Penalty represents the model stability; Cost Penalty represents the update cost;

[0207] Use a Gaussian process to build a surrogate model to approximate the objective function f(λ t , λ s , frequency), learn the potential laws of the objective function through historical data, and give the prediction results and uncertainties of each set of parameters. The surrogate model is as follows:

[0208]

[0209] Where: μ represents the mean function; k represents the covariance function; λ t ′, λ s ′, frequency′ are the neighboring points of the currently selected λ t , λ s , frequency parameter points;

[0210] Adopt the expected improvement as the acquisition function to guide the exploration process of Bayesian optimization. Then the goal is to maximize:

[0211]

[0212] In the formula: f * represents the current optimal objective value; f(λ t , λ s , frequency) represents the predicted value of the surrogate model;

[0213] In each iteration, Bayesian optimization selects a new set of parameters (λ t , λ s , frequency), and updates the surrogate model by calculating the objective function, which can intelligently select the optimal parameter configuration in the high-dimensional parameter space.

[0214] In this embodiment, through Bayesian optimization, the spatio-temporal decay factor and the feature update frequency can be intelligently adjusted to maximize the timeliness, freshness, and business effect of features.

[0215] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for updating insurance customer feature vectors based on dynamic adjustment of spatiotemporal correlation, characterized in that: The following steps are involved: Step 1: Collect customer core data, external spatiotemporal data, and real-time data streams, standardize the spatiotemporal coordinates of the collected data, and unify them into a standard grid system; attach spatiotemporal tags to each business event to ensure that each event clearly corresponds to the time and space location, and set timeliness tags for the data to obtain the spatiotemporal association matrix between customers and events; Step 2: Extract dynamic features based on the spatiotemporal association matrix and the customer's historical feature vector, perform periodic feature analysis from the time dimension, calculate the event density change rate, quantify the change trend within a short time window, extract the entropy value of the mobility pattern and the migration of spatial hotspots from the spatial dimension, and obtain a spatiotemporal enhanced feature tensor, including spatiotemporal features and corresponding timeliness labels; Step 3: Based on the spatiotemporal enhanced feature tensor, time and space are attenuated, and time and space attenuation functions are used to weight timeliness and spatial distance respectively. Then, the multi-head attention mechanism is used to calculate the importance weights of the spatiotemporal dimensions, and the feature vector is dynamically adjusted to obtain a feature vector with spatiotemporal weights. Step 4: Based on the feature vector with spatiotemporal weights, an event-driven and gradual update mechanism is constructed. At the same time, a spatiotemporal attenuation threshold is set to control the update frequency. The changes of feature vectors and time information are balanced through a combined update strategy. A versioned feature vector is generated and an update log is recorded. Step 5: Monitor the impact of feature updates on core business indicators, analyze the relationship between feature freshness and business indicators, and use the strategy optimizer to dynamically adjust the time decay parameters and update strategy to obtain the optimized update strategy parameters.

2. The insurance customer feature vector updating method based on dynamic adjustment of spatiotemporal correlation according to claim 1 is characterized in that: The step 1 comprises the following steps: Data acquisition and preprocessing: Collect customer core data, external spatiotemporal data, and real-time data, and clean and preprocess the collected data; Standardization and gridding of time and space coordinates: Convert each data point into H3 code: Adding spatiotemporal labels: Get the timestamp of the event and convert the geographic location of the event into a unique identifier through H3 grid coding to obtain the spatiotemporal label of the event; Set timeliness labels: Add timeliness labels to each piece of data and weight them according to the time decay function: Where: D(t) represents the time-dependent attenuation value; T now represents the current time; T represents the timestamp of the data record; λ represents the attenuation parameter; The types of timeliness tags are set to three categories according to the timeliness decay value, including static data, quasi-real-time data, and real-time data; Generate a spatiotemporal association matrix: Combine the spatiotemporal label, timeliness label and customer history feature vector of each event to construct a spatiotemporal association matrix between customers and events.

3. The insurance customer feature vector updating method based on dynamic adjustment of spatiotemporal correlation according to claim 1 is characterized in that: The step 2 comprises the following steps: Time dimension feature extraction: Fourier transform is used to process the time series data in the spatiotemporal correlation matrix, and then the frequency components above the preset threshold are extracted based on the results of Fourier transform to obtain periodic features, including daily periodicity, weekly periodicity and monthly periodicity features; Event density change rate: Based on the event occurrence time of the spatiotemporal correlation matrix, calculate the density change of events in a short time window and quantify the fluctuation trend of event frequency: Where: N(t1) represents the number of events in the time period t1; N(t0) represents the number of events in the time period t0; t1-t0 represents the length of the time window; R(t) represents the rate of change of the density of events; Spatial entropy feature extraction: Based on the event space labels in the spatiotemporal association matrix, the spatial distribution is divided into multiple spatial units, and the spatial entropy H(S) is calculated: Where: p i represents the probability of an event occurring in a spatial unit, Where N i represents the number of events in the ith spatial unit; N total represents the total number of events in all spatial units; n represents the number of spatial units; Spatial hotspot migration feature extraction: Use hotspot detection methods to identify hotspot areas in space and calculate the migration trajectory of hotspot areas over time: C h (t)=||C h (t1)-C h (t0)||; Where: C h (t0) represents the coordinates of the center of the hot spot area at time t0; c h (t1) represents the coordinates of the center of the hot spot area at time t1; C h (t) represents the spatial migration distance of the hotspot center in time period t; By calculating the migration distances between different time points, the migration patterns of spatial hot spots are identified; The spatiotemporal enhanced feature tensor is obtained based on time dimension feature extraction, event density change rate extraction, spatial entropy feature extraction and spatial hotspot migration feature extraction.

4. The method for updating the insurance customer feature vector based on dynamic adjustment of spatiotemporal correlation according to claim 1, characterized in that: The step 2 also includes spatiotemporal interaction feature extraction, and the features extracted from the spatiotemporal interaction features are used as part of the spatiotemporal enhanced feature tensor, wherein the spatiotemporal interaction feature extraction includes spatiotemporal co-occurrence pattern extraction and cross-regional correlation strength extraction; Spatiotemporal co-occurrence pattern extraction: Calculate the co-occurrence of events under different spatiotemporal conditions and identify the spatiotemporal interaction relationship of events, that is, calculate the spatiotemporal co-occurrence of events at different times and spatial locations: Where: Indicates event E i and event E j At time t k and space S k If co-occurrence occurs, it is 1, otherwise it is 0; N represents the total number of time windows calculated; C ij (t) represents event E i and event E j correlations in space and time; Cross-region correlation strength extraction: for events E in different regions A and B A and event E B , calculate the cross-regional association strength: Where: N A and N B represents the number of events in region A and region B; Indicates event E A (i) and E B (j) Whether they occur simultaneously; C AB Represents the strength of spatiotemporal correlation across regions A and B.

5. The method for updating the insurance customer feature vector based on dynamic adjustment of spatiotemporal correlation according to claim 1, characterized in that: The step 3 comprises the following steps: Time decay function: Where: t0 represents the time when the event occurs; t represents the current time; λ t represents the time decay factor; e represents the base of the natural logarithm; Spatial attenuation function: Where: d represents the spatial distance between events; λ s represents the spatial attenuation factor; Multi-head attention mechanism: Assume that the spatiotemporal enhanced tensor T is divided into h subspaces, that is, each head of the multi-head attention mechanism processes T h The calculation process is as follows: Input query, key and value, calculate attention score based on the input query, key and key dimension, and then get weighted representation based on the calculated attention score weight value matrix; connect the multi-head output of the multi-head attention mechanism by connecting the results of each head to get the overall feature representation; Temporal and spatial weight adjustment: Based on the overall feature representation output by the multi-head attention mechanism and the values ​​of the time decay function and the space decay function, a feature vector with temporal and spatial weights is obtained: T weighted =MultiHeadAttention(T)·f t (t)·f s (d); Where: MultiHeadAttention(T) represents the overall feature representation of the output of the multi-head attention mechanism; T weighted Represents a feature vector with spatiotemporal weights.

6. The insurance customer feature vector updating method based on dynamic adjustment of spatiotemporal correlation according to claim 1, characterized in that: The event-driven update mechanism is as follows: Set the time trigger update condition to δ(t). When a major event occurs, the feature vector update is triggered: T new =T old +α·δ(t)·ΔT event ; Where: T new represents the updated feature vector with spatiotemporal weights; T old represents the feature vector with spatiotemporal weights before updating; α represents the update intensity factor; δ(t) represents the event trigger flag function, which is 1 if a major event occurs, otherwise 0; ΔT event Represents the characteristic change of spatiotemporal weight caused by the event; The judgment of the major event is based on business rules, and the characteristic change amount is calculated based on business rules.

7. The method for updating insurance customer feature vectors based on dynamic adjustment of spatiotemporal correlation according to claim 1, characterized in that: The gradual update mechanism is as follows: T gradual =T old +γ(t)·ΔT gradual ; Where: T gradual represents the feature vector with spatiotemporal weights after gradual update; γ(t) represents the gradual factor, which is calculated based on the product of the time decay function and the space decay function; ΔT gradual Represents the changes in features with spatiotemporal weights caused by small changes, calculated based on the cumulative effect of small events.

8. The method for updating insurance customer feature vectors based on dynamic adjustment of spatiotemporal correlation according to claim 1, characterized in that: The specific steps of controlling the update frequency based on the spatiotemporal attenuation threshold are as follows: For each feature t i , calculate the time-effect decay value and compare it with the spatiotemporal threshold: t i (t)=T(t)·f t (t)·f s (d); Where: T(t) represents the feature representation of the output of the multi-head attention mechanism; f t (t) represents the time decay function; f s (d) represents the spatial attenuation function; t i (t) represents feature t with spatiotemporal weights i The attenuation value of When t i (t) is less than the threshold value θ i The update is triggered when the update is triggered, and the logic of triggering the update is as follows: If there is a major event, the event-determined update mechanism is executed; if there is no major event, the gradual update mechanism is executed.

9. The method for updating insurance customer feature vectors based on dynamic adjustment of spatiotemporal correlation according to claim 1, characterized in that: The step 4 also includes feature versioning and historical tracing; a version control mechanism is used to record changes in feature vectors with spatiotemporal weights. Each feature update generates a new feature version number and records the updated audit log. Through version control, historical tracing and grayscale release strategies are implemented.

10. The insurance customer feature vector updating method based on dynamic adjustment of spatiotemporal correlation according to claim 1, characterized in that: The step 5 uses Bayesian optimization to optimize the time decay parameters and update strategy, wherein Bayesian optimization is used to maximize the business effect of the model and the freshness of the features, and the time decay factor, space decay factor and feature update frequency are comprehensively optimized.

Images