A dynamic tag based decision optimization system and method

Abstract

This invention discloses a decision optimization system and method based on dynamic labels, belonging to the field of decision optimization technology. The system includes: a dynamic label time series construction module, a time series feature discrimination module, an intervention node localization module, and a time series deduplication intervention module. The method collects multi-dimensional time series features at a preset period, generates dynamic labels through normalization and weighted fusion, and constructs a time series set. Stable and abnormal time series features are distinguished by calculating time series deviation and feature fluctuation values. When anomalies are identified, the optimal response period is constructed with the anomaly time as the endpoint. The relative change rate of adjacent labels is calculated, and intervention nodes are determined based on local maxima. Then, the intervention overlap is calculated through a time series deduplication window to avoid duplicate interventions. This invention can improve the accuracy of anomaly identification and the timeliness of intervention, reduce redundant operations, and is suitable for various time series data-driven intelligent decision-making scenarios.

Description

Technical Field

[0001] This invention relates to the field of decision optimization technology, specifically to a decision optimization system and method based on dynamic labels. Background Technology

[0002] In existing intelligent decision-making systems for retail customers, intervention nodes are often triggered by fixed time periods, fixed thresholds, or manually preset rules. These systems typically first label customers with static information and then issue decision instructions at fixed time intervals, or determine whether to trigger an intervention action based on a single numerical value, making it difficult to adapt to real-time changes in the customer's status.

[0003] In actual operation, static tags cannot reflect the continuous changes in customer operations, behavior, and value status in a timely manner. Fixed-time triggering can easily lead to problems such as delayed intervention, premature or ineffective outreach. Although some systems have introduced customer status update mechanisms, they still use single-point status as the basis for judgment, without continuously tracking and quantitatively analyzing tag change trends, resulting in a mismatch between decision-making timing and the customer's actual responsiveness. At the same time, existing solutions lack differentiated handling of different change patterns such as sudden changes, continuous increases, and sustained stability of tags. When customer status fluctuates rapidly, timely intervention cannot be initiated, while repeated outreach is prone to occur when the status is stable, which not only consumes system resources but also reduces the effectiveness of decision execution. Summary of the Invention

[0004] The purpose of this invention is to provide a decision optimization system and method based on dynamic labels to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A decision optimization method based on dynamic labels includes the following steps: S1. Collect multi-dimensional time-series feature data reported by user devices according to a preset cycle, generate dynamic tags based on the multi-dimensional time-series feature data, and construct a dynamic tag time-series set in chronological order; S2. Compare the current dynamic tag with the historical tags in the dynamic tag time series set to calculate the time series deviation and feature fluctuation value of the tag within the continuous window, thereby distinguishing the stable time series features and abnormal time series features of the tag. S3. When an abnormal time series feature is identified, starting from the time when the anomaly occurs, trace back to the time series interval corresponding to the length of the continuous window in S2, and determine this time series interval as the optimal response period; calculate the relative rate of change of dynamic labels of adjacent collection times within the optimal response period, and construct a rate of change sequence; identify local maxima in the rate of change sequence, and take the next time point corresponding to the latest local maxima as the intervention node; if there is no local maxima, take the next time point corresponding to the maximum relative rate of change as the intervention node; when a stable time series feature is identified, no additional intervention operation is performed, and data is collected and anomaly detection is performed according to a fixed period. S4. Perform time-series deduplication detection on intervention nodes. If the same object has been detected to have completed the intervention within the preset time window, cancel the current intervention process. If no duplicate execution record is detected, carry out decision intervention based on the determined intervention nodes.

[0006] Furthermore, S1 includes the following: Multi-dimensional time-series feature data of user terminals and business service systems are collected according to a preset collection cycle. The multi-dimensional time-series feature data includes terminal interaction event features, session behavior quantification features, system resource access features, and device network status features. The multi-dimensional time-series feature data collected at the same time are used to form the original feature vector of the corresponding time. Each type of feature data in the original feature vector is normalized and the values ​​of each type of feature are uniformly mapped to the range of 0 to 1. Each type of feature data that has undergone normalization is assigned a preset weight coefficient. Each feature data is multiplied by its corresponding weight coefficient, and all the product results are summed to obtain the comprehensive feature value at the current time. The sum of all weight coefficients is one, and the weight coefficients are determined by statistical methods based on historical data. The overall range of values ​​of the comprehensive feature value is divided into equal intervals based on the preset total number of dynamic tags, forming a continuous numerical interval with the same number of dynamic tags. The continuous numerical intervals are assigned positive integer numbers starting from 1 in ascending order, and different intervals correspond to different dynamic tags. The dynamic label corresponding to the numerical range into which the comprehensive feature value falls is used as the dynamic label of the current moment. The dynamic labels generated at each moment are associated with and stored with the corresponding timestamps, and arranged in chronological order of collection time to form a dynamic label time series set.

[0007] Furthermore, S2 includes the following: Calculate the temporal deviation between each historical dynamic tag and the current dynamic tag within the continuous window. The temporal deviation is equal to the absolute value of the difference between the value of the historical dynamic tag and the value of the current dynamic tag. The length of the continuous window is greater than 1, and the continuous window contains the dynamic tag at the current acquisition time and several historical dynamic tags before the current acquisition time. The number of historical dynamic tags is equal to the length of the continuous window minus one. The temporal deviations corresponding to all historical dynamic tags within the continuous window are summed to obtain the total deviation. The total deviation is then divided by the number of historical dynamic tags within the continuous window to obtain the average temporal deviation, where the number of historical dynamic tags is equal to the length of the continuous window minus one. Calculate the difference between each time series deviation and the average time series deviation within the continuous window, sum the squares of each difference to obtain the sum of squares; divide the sum of squares by the number of historical dynamic tags, and then take the square root of the division result to obtain the characteristic fluctuation value. A preset deviation threshold and a fluctuation threshold are set, both of which are determined by statistical analysis of historical dynamic tag time-series data; When the average time series deviation is less than or equal to the deviation threshold and the feature fluctuation value is less than or equal to the fluctuation value threshold, the tag time series feature within the continuous window is determined to be a stable time series feature; when the average time series deviation is greater than the deviation threshold or the feature fluctuation value is greater than the fluctuation value threshold, the tag time series feature within the continuous window is determined to be an abnormal time series feature.

[0008] Furthermore, S3 includes the following: When the timing characteristics of the tags within a continuous window are determined to be abnormal timing characteristics, the acquisition time when the abnormality occurs is taken as the endpoint, and a timing interval is formed by tracing back the window length by acquisition time. This timing interval is determined as the optimal response period. The window length is the same as the length of the continuous window defined in S2. Traverse the dynamic tag sequence within the optimal response period and calculate the relative change rate of dynamic tags at adjacent acquisition times. The relative change rate is obtained by taking the absolute value of the difference between adjacent tags and dividing it by the larger of the two adjacent tags. Arrange the calculated relative change rates in chronological order to form a relative change rate sequence. The relative rate of change sequence is compared point by point to identify the numerical points that meet the local maximum condition. The criterion for determining the local maximum is that the current relative rate of change value is greater than the relative rate of change values ​​of the adjacent previous time and the adjacent next time. The beginning and end points of the relative rate of change sequence are not involved in the local maximum determination. Only the points in the middle of the sequence are judged for maximum value. Record the time information corresponding to all numerical points that satisfy the local maximum condition, select the one with the largest time value from the recorded results, and use the next data collection time point calculated according to the fixed collection cycle as the intervention node. When there are no numerical points in the relative rate of change sequence that satisfy the local maximum condition, all values ​​in the sequence are compared to determine the time information corresponding to the largest value, and the next data collection time point calculated according to the fixed collection cycle is used as the intervention node. When stable temporal features are identified, the original collection cycle remains unchanged, and the multi-dimensional temporal feature collection and temporal feature comparison operations are continued. No intervention node determination or decision-making intervention related operations are performed.

[0009] Furthermore, S4 includes the following: Traverse all historical acquisition times within the time-series deduplication window, and calculate the intervention overlap degree corresponding to each time. The intervention overlap degree is equal to the absolute value of the time difference between the intervention node time and the corresponding historical time, divided by the length of the time-series deduplication window. The length of the time-series deduplication window is greater than 1, and the time-series deduplication window contains several historical acquisition times before the intervention node time. The number of historical acquisition times is equal to the length of the time-series deduplication window. All intervention overlap rates are arranged in chronological order of data collection time to form an overlap rate sequence; the overlap rate sequence is numerically compared to identify numerical items in the sequence that are less than a preset overlap rate threshold; the overlap rate threshold is determined based on the statistical results of historical intervention effectiveness data. If there are numerical items with a value less than the overlap threshold in the time-series deduplication window, it is determined that the same object has completed the intervention within the corresponding time range, and the current intervention process is terminated. If there are no numerical items with a value less than the overlap threshold in the time-series deduplication window, it is determined that there are no duplicate intervention records in the detection interval. The decision intervention process is initiated based on the intervention node time, and the object identifier and execution time corresponding to this intervention are recorded and updated to the historical intervention record set.

[0010] A decision optimization system based on dynamic labels includes: a dynamic label time series construction module, a time series feature discrimination module, an intervention node localization module, and a time series deduplication intervention module; The dynamic tag time series construction module collects multi-dimensional time series feature data reported by user devices according to a preset period, generates dynamic tags based on the multi-dimensional time series feature data, and constructs a dynamic tag time series set in chronological order; The temporal feature discrimination module compares the current dynamic tag with the historical tags in the dynamic tag temporal set, calculates the temporal deviation and feature fluctuation value of the tag within the continuous window, and distinguishes the stable temporal features and abnormal temporal features of the tag accordingly. When the intervention node localization module identifies abnormal temporal features, it traces back a temporal interval corresponding to the length of the continuous window, with the time of the abnormal occurrence as the endpoint, and determines this temporal interval as the optimal response period. It calculates the relative rate of change of dynamic tags at adjacent acquisition times within the optimal response period and constructs a rate of change sequence. It identifies local maxima in the rate of change sequence and uses the next time point corresponding to the latest local maxima as the intervention node. If there is no local maxima, the next time point corresponding to the maximum relative rate of change is used as the intervention node. When the identified temporal features are stable, the intervention node localization operation is not performed. The time-series deduplication intervention module performs time-series deduplication detection on intervention nodes. When it detects that the same object has completed the intervention within a preset time window, the current intervention process is canceled. When no duplicate execution record is detected, decision intervention is carried out based on the determined intervention nodes.

[0011] Furthermore, the data dynamic label time series construction module includes a feature acquisition unit and a label generation unit; The feature acquisition unit is used to collect terminal interaction event features, session behavior quantification features, system resource access features and device network status features, and construct the original feature vector to complete the feature data normalization process. The label generation unit is used to perform weighted summation of normalized features to obtain comprehensive feature values, obtain dynamic labels through equally divided interval mapping, and construct a dynamic label time series set.

[0012] Furthermore, the time series feature discrimination module includes a separation degree calculation unit and a time series state determination unit; The deviation calculation unit calculates the temporal deviation between historical dynamic tags and the current dynamic tag, the average temporal deviation, and the characteristic fluctuation value; The timing state determination unit is used to compare the calculation results with the corresponding threshold to determine whether the timing feature of the label is a stable timing feature or an abnormal timing feature.

[0013] Furthermore, the intervention node location module includes a response time period construction unit and a rate of change analysis unit; The response time period construction unit traces back to the optimal response time period with the time when the anomaly occurred as the endpoint. The rate of change analysis unit is used to calculate the relative rate of change of adjacent labels, identify local maxima in the sequence, and determine the corresponding intervention nodes.

[0014] Furthermore, the temporal deduplication intervention module includes an overlap calculation unit and a duplicate detection execution unit; The overlap calculation unit calculates the intervention overlap at each time point within the temporal deduplication window and constructs the overlap sequence; The duplicate detection execution unit is used to compare the degree of overlap with the corresponding threshold, determine the duplicate intervention status, and execute the process termination or decision intervention operation, while updating the historical intervention records synchronously.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention generates dynamic labels through multi-dimensional temporal feature normalization and weighted fusion, which can uniformly quantify the value differences of different types of data, improving the accuracy and stability of label representation. By determining the temporal state through dual indicators of temporal deviation and feature fluctuation value, it can more accurately distinguish between stable and abnormal scenarios, reducing misjudgment problems caused by a single threshold. Based on backtracking from abnormal moments to construct the optimal response period, and using the extreme points of relative change rate to determine intervention nodes, it can accurately locate the location of changes and abrupt changes, achieving more timely and effective intervention triggering. The use of a local maximum filtering rule for relative change rate can eliminate sequence endpoint interference, improving the reliability of intervention node selection. Through temporal deduplication windows and intervention overlap detection mechanisms, it can effectively avoid repeated interventions on the same object within a short period, reducing resource waste and redundant operations. The overall solution realizes a closed-loop process from data collection, label generation, anomaly detection to intervention execution, with clear operational logic and strong adaptability, significantly improving the accuracy, timeliness, and resource utilization of the dynamic label decision-making system. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a decision optimization method based on dynamic labels according to the present invention. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1 The present invention provides the following technical solution: A decision optimization method based on dynamic labels includes the following steps: S1. Collect multi-dimensional time-series feature data reported by user devices according to a preset cycle, generate dynamic tags based on the multi-dimensional time-series feature data, and construct a dynamic tag time-series set in chronological order; S2. Compare the current dynamic tag with the historical tags in the dynamic tag time series set to calculate the time series deviation and feature fluctuation value of the tag within the continuous window, thereby distinguishing the stable time series features and abnormal time series features of the tag. S3. When an abnormal time series feature is identified, starting from the time when the anomaly occurs, trace back to the time series interval corresponding to the length of the continuous window in S2, and determine this time series interval as the optimal response period; calculate the relative rate of change of dynamic labels of adjacent collection times within the optimal response period, and construct a rate of change sequence; identify local maxima in the rate of change sequence, and take the next time point corresponding to the latest local maxima as the intervention node; if there is no local maxima, take the next time point corresponding to the maximum relative rate of change as the intervention node; when a stable time series feature is identified, no additional intervention operation is performed, and data is collected and anomaly detection is performed according to a fixed period. S4. Perform time-series deduplication detection on intervention nodes. If the same object has been detected to have completed the intervention within the preset time window, cancel the current intervention process. If no duplicate execution record is detected, carry out decision intervention based on the determined intervention nodes.

[0019] S1 includes the following: Multi-dimensional time-series feature data of user terminals and business service systems are collected according to a preset collection cycle. The multi-dimensional time-series feature data includes terminal interaction event features, session behavior quantification features, system resource access features, and device network status features. The multi-dimensional time-series feature data collected at the same time are used to form the original feature vector of the corresponding time. Each type of feature data in the original feature vector is normalized and the values ​​of each type of feature are uniformly mapped to the range of 0 to 1. Each type of feature data that has undergone normalization is assigned a preset weight coefficient. Each feature data is multiplied by its corresponding weight coefficient, and all the product results are summed to obtain the comprehensive feature value at the current time. The sum of all weight coefficients is one, and the weight coefficients are determined by statistical methods based on historical data. The overall range of values ​​of the comprehensive feature value is divided into equal intervals based on the preset total number of dynamic tags, forming a continuous numerical interval with the same number of dynamic tags. The continuous numerical intervals are assigned positive integer numbers starting from 1 in ascending order, and different intervals correspond to different dynamic tags. The dynamic label corresponding to the numerical range into which the comprehensive feature value falls is used as the dynamic label of the current moment. The dynamic labels generated at each moment are associated with and stored with the corresponding timestamps, and arranged in chronological order of collection time to form a dynamic label time series set.

[0020] In this embodiment, a unified collection period T is set, and multi-dimensional time-series feature data is collected from user terminals and business service systems according to the collection period T. The multi-dimensional time-series feature data includes terminal interaction event features, session behavior quantification features, system resource access features, and device network status features. For example, terminal interaction event features include click frequency, swiping speed, etc.; session behavior quantification features include page dwell time, interaction interval, etc.; system resource access features include CPU utilization, memory usage, etc.; and device network status features include network latency, packet loss rate, etc. Multi-dimensional temporal feature data collected at the same time are used to construct the original feature vector Xt for the corresponding time, and Xt is composed of x t1 x t2 x t3 x t4 Composed in sequence, x t1 x t2 x t3 x t4 These correspond to the quantitative values ​​of terminal interaction event characteristics, session behavior quantitative characteristics, system resource access characteristics, and device network status characteristics, respectively. Normalization is performed on each component of the original feature vector Xt, mapping the values ​​to the interval between 0 and 1 to eliminate dimensional differences; and the normalization calculation formula is: x' ti =(x ti -min(x i )) / (max(x i )-min(x i ), where min(x i ) and max(x i ) represent the minimum and maximum values ​​of the i-th feature within the historical period, respectively, x' ti The result after component normalization, i takes values ​​from 1 to 4; after normalization, a standardized feature vector X't is formed, X't is derived from x' t1 、x' t2 、x' t3 、x' t4 Composed in sequence; A weighted fusion calculation is performed on each component of the standardized feature vector to obtain the comprehensive feature value Ct at the current time. The weighted fusion calculation formula is: Ct = w1 × x' t1 +w2×x' t2 +w3×x' t3 +w4×x' t4 Among them, w1, w2, w3 and w4 are preset weight coefficients, the sum of the weight coefficients is 1, and the weight coefficients are determined by historical data statistics. Based on the preset total number of dynamic tags, the total range of values ​​for the comprehensive feature value is divided into several consecutive numerical intervals, which are the same as the number of tags. Each consecutive numerical interval uniquely corresponds to a dynamic tag. The consecutive numerical intervals are assigned positive integer numbers starting from 1 in ascending order, that is, the first interval corresponds to tag 1, the second interval corresponds to tag 2, and so on. When the comprehensive feature value Ct falls into the corresponding consecutive numerical interval, the dynamic tag corresponding to the consecutive numerical interval is taken as the dynamic tag Lt at the current time. The dynamic tags Lt generated at each time point are associated with and stored with the corresponding timestamps t, and arranged in chronological order of collection time to form a dynamic tag time series set L.

[0021] S2 includes the following: Calculate the temporal deviation between each historical dynamic tag and the current dynamic tag within the continuous window. The temporal deviation is equal to the absolute value of the difference between the value of the historical dynamic tag and the value of the current dynamic tag. The length of the continuous window is greater than 1, and the continuous window contains the dynamic tag at the current acquisition time and several historical dynamic tags before the current acquisition time. The number of historical dynamic tags is equal to the length of the continuous window minus one. The temporal deviations corresponding to all historical dynamic tags within the continuous window are summed to obtain the total deviation. The total deviation is then divided by the number of historical dynamic tags within the continuous window to obtain the average temporal deviation, where the number of historical dynamic tags is equal to the length of the continuous window minus one. Calculate the difference between each time series deviation and the average time series deviation within the continuous window, sum the squares of each difference to obtain the sum of squares; divide the sum of squares by the number of historical dynamic tags, and then take the square root of the division result to obtain the characteristic fluctuation value. A preset deviation threshold and a fluctuation threshold are set, both of which are determined by statistical analysis of historical dynamic tag time-series data; When the average time series deviation is less than or equal to the deviation threshold and the feature fluctuation value is less than or equal to the fluctuation value threshold, the tag time series feature within the continuous window is determined to be a stable time series feature; when the average time series deviation is greater than the deviation threshold or the feature fluctuation value is greater than the fluctuation value threshold, the tag time series feature within the continuous window is determined to be an abnormal time series feature.

[0022] In this embodiment, the length of the continuous window is set to k, where k is a positive integer greater than 1. The continuous window consists of dynamic tags from the current time and the previous k-1 acquisition times, and the tag sequence within the continuous window is L. t-k+1 L t-k+2 ... L t L t For the current dynamic tag, L t-k+1To L t-1 For historical tags in the dynamic tag time series set L; Calculate the temporal deviation d between each historical label and the current dynamic label within a continuous window. tj The formula for calculating the timing deviation is: d tj =|L t -L j | where Lj is the historical label at the j-th time within the continuous window, j takes values ​​from t-k+1 to t-1, d tj This represents the deviation value between the j-th historical tag and the current dynamic tag; The average time series deviation D within a continuous window is calculated using the formula: D = ∑d tj / (k-1), where ∑d tj The time series deviation d within the continuous window tj The sum, where k-1 is the number of historical tags within a consecutive window; The characteristic fluctuation value V within a continuous window is calculated using the formula: V = [∑(d tj -D) 2 / (k-1)]^(1 / 2), where, ∑(d tj -D) 2 The time series deviation d within the continuous window tj The sum of squares of the differences from the average time series deviation D; A deviation threshold D0 and a fluctuation threshold V0 are set. Both the deviation threshold D0 and the fluctuation threshold V0 are determined by statistical analysis of historical tag time series data and are used to distinguish between stable time series characteristics and abnormal time series characteristics of tags. When the average time series deviation D is less than or equal to the deviation threshold D0, and the characteristic fluctuation value V is less than or equal to the fluctuation value threshold V0, the label time series feature within the continuous window is determined to be a stable time series feature; when the average time series deviation D is greater than the deviation threshold D0, or the characteristic fluctuation value V is greater than the fluctuation value threshold V0, the label time series feature within the continuous window is determined to be an abnormal time series feature.

[0023] S3 includes the following: When the timing characteristics of the tags within a continuous window are determined to be abnormal timing characteristics, the acquisition time when the abnormality occurs is taken as the endpoint, and a timing interval is formed by tracing back the window length by acquisition time. This timing interval is determined as the optimal response period. The window length is the same as the length of the continuous window defined in S2. Traverse the dynamic tag sequence within the optimal response period and calculate the relative change rate of dynamic tags at adjacent acquisition times. The relative change rate is obtained by taking the absolute value of the difference between adjacent tags and dividing it by the larger of the two adjacent tags. Arrange the calculated relative change rates in chronological order to form a relative change rate sequence. The relative rate of change sequence is compared point by point to identify the numerical points that meet the local maximum condition. The criterion for determining the local maximum is that the current relative rate of change value is greater than the relative rate of change values ​​of the adjacent previous time and the adjacent next time. The beginning and end points of the relative rate of change sequence are not involved in the local maximum determination. Only the points in the middle of the sequence are judged for maximum value. Record the time information corresponding to all numerical points that satisfy the local maximum condition, select the one with the largest time value from the recorded results, and use the next data collection time point calculated according to the fixed collection cycle as the intervention node. When there are no numerical points in the relative rate of change sequence that satisfy the local maximum condition, all values ​​in the sequence are compared to determine the time information corresponding to the largest value, and the next data collection time point calculated according to the fixed collection cycle is used as the intervention node. When stable temporal features are identified, the original collection cycle remains unchanged, and the multi-dimensional temporal feature collection and temporal feature comparison operations are continued. No intervention node determination or decision-making intervention related operations are performed.

[0024] In this embodiment, when the tag temporal characteristics within a continuous window are determined to be abnormal, the time t at which the abnormality occurs is taken as the endpoint, and k collection times are traced back to form a temporal interval. This temporal interval is determined as the optimal response period, and the dynamic tag sequence contained within the optimal response period is L. t-k+1 L t-k+2 ... L t ; Traverse the dynamic tag sequence within the optimal response period, calculate the relative rate of change Rm of the dynamic tags at adjacent acquisition times, and Rm = |L m -L m-1 | / max(L m ,L m-1 ), where L m For the dynamic label at time m within the optimal response period, L m-1 The dynamic label at time m-1 within the optimal response period, where m ranges from t-k+2 to t; max(L m ,L m-1 The larger of the dynamic labels between two adjacent time points is used; all relative rates of change are arranged in chronological order to form a relative rate of change sequence R. The relative rate of change sequence is compared point by point to identify the numerical points that meet the local maximum condition. The criterion for determining the local maximum is that the current relative rate of change value is greater than the relative rate of change values ​​of the previous time step and the next time step. The beginning and end points of the relative rate of change sequence are not included in the local maximum determination. Only the points in the middle of the sequence are judged for local maximum. Record the time information corresponding to all local maxima that meet the conditions, select the one with the largest time value from the recorded results, and take the next acquisition time of the next time m after the rate of change corresponding to the local maximum as the intervention node; the next acquisition time is the next data acquisition time point calculated with a fixed acquisition period T. When there are no numerical points in the relative rate of change sequence that satisfy the local maximum condition, all values ​​in the relative rate of change sequence are compared to determine the time information corresponding to the largest value, and the next acquisition time of the next time m after the rate of change corresponding to the maximum value is taken as the intervention node. When the label time series features within the continuous window are determined to be stable time series features, the collection period T remains unchanged, and multi-dimensional time series feature data is collected in the manner described in S1. At the same time, time series feature comparison and anomaly detection operations are performed in the manner described in S2. No intervention node determination or decision intervention related operations are carried out.

[0025] S4 includes the following: Traverse all historical acquisition times within the time-series deduplication window, and calculate the intervention overlap degree corresponding to each time. The intervention overlap degree is equal to the absolute value of the time difference between the intervention node time and the corresponding historical time, divided by the length of the time-series deduplication window. The length of the time-series deduplication window is greater than 1, and the time-series deduplication window contains several historical acquisition times before the intervention node time. The number of historical acquisition times is equal to the length of the time-series deduplication window. All intervention overlap rates are arranged in chronological order of data collection time to form an overlap rate sequence; the overlap rate sequence is numerically compared to identify numerical items in the sequence that are less than a preset overlap rate threshold; the overlap rate threshold is determined based on the statistical results of historical intervention effectiveness data. If there are numerical items with a value less than the overlap threshold in the time-series deduplication window, it is determined that the same object has completed the intervention within the corresponding time range, and the current intervention process is terminated. If there are no numerical items with a value less than the overlap threshold in the time-series deduplication window, it is determined that there are no duplicate intervention records in the detection interval. The decision intervention process is initiated based on the intervention node time, and the object identifier and execution time corresponding to this intervention are recorded and updated to the historical intervention record set.

[0026] In this embodiment, the length of the time series deduplication window is set to p, where p is a positive integer greater than 1. The length of the time series deduplication window is determined statistically from historical intervention records. Taking the time t0 corresponding to the intervention node as the benchmark, p collection times are extracted backward to form a time series detection interval that does not include t0. The time series contained in the time series detection interval is t. 0-p+1 t 0-p+2 ... t 0-1 ; Iterate through all time points within the time-series detection interval, calculate the intervention overlap degree Hn corresponding to each time point, and Hn = |t0 - t n | / p, where t n Let n be the timestamp of the nth time point within the time-series detection interval, where n takes the value t. 0-p+1 To t 0-1 ,|t0-t n | represents the absolute time difference between the intervention node moment and the moment within the time-series detection interval; All intervention overlaps are arranged in chronological order to form an overlap sequence H; the overlap sequence is compared numerically to identify numerical items in the sequence that are less than the preset overlap threshold H0, and the overlap threshold H0 is determined by statistical analysis of historical intervention effectiveness data; if there are numerical items less than the overlap threshold H0, it is determined that the same object has completed the intervention within the preset time window, and the current intervention process is canceled. If there are no numerical items with a value less than the overlap threshold H0, it is determined that there are no duplicate intervention records in the time series detection interval; the decision intervention process is initiated according to the intervention node time, the object identifier and time information corresponding to this intervention are recorded, and the recorded results are synchronously updated to the historical intervention record set to provide a data foundation for subsequent time series deduplication detection; After completing the single intervention process judgment, keep the collection period T unchanged, return to S1 to continue to perform multi-dimensional temporal feature collection operations, and carry out the dynamic label generation, temporal feature judgment, intervention node determination and temporal deduplication detection process in a loop to achieve continuous dynamic label decision optimization.

[0027] A decision optimization system based on dynamic labels includes: a dynamic label time series construction module, a time series feature discrimination module, an intervention node localization module, and a time series deduplication intervention module; The dynamic tag time series construction module collects multi-dimensional time series feature data reported by user devices according to a preset period, generates dynamic tags based on the multi-dimensional time series feature data, and constructs a dynamic tag time series set in chronological order; The temporal feature discrimination module compares the current dynamic tag with the historical tags in the dynamic tag temporal set, calculates the temporal deviation and feature fluctuation value of the tag within the continuous window, and distinguishes the stable temporal features and abnormal temporal features of the tag accordingly. When the intervention node localization module identifies abnormal temporal features, it traces back a temporal interval corresponding to the length of the continuous window, with the time of the abnormal occurrence as the endpoint, and determines this temporal interval as the optimal response period. It calculates the relative rate of change of dynamic tags at adjacent acquisition times within the optimal response period and constructs a rate of change sequence. It identifies local maxima in the rate of change sequence and uses the next time point corresponding to the latest local maxima as the intervention node. If there is no local maxima, the next time point corresponding to the maximum relative rate of change is used as the intervention node. When the identified temporal features are stable, the intervention node localization operation is not performed. The time-series deduplication intervention module performs time-series deduplication detection on intervention nodes. When it detects that the same object has completed the intervention within a preset time window, the current intervention process is canceled. When no duplicate execution record is detected, decision intervention is carried out based on the determined intervention nodes.

[0028] The data dynamic label time series construction module includes a feature acquisition unit and a label generation unit; The feature acquisition unit is used to collect terminal interaction event features, session behavior quantification features, system resource access features and device network status features, and construct the original feature vector to complete the feature data normalization process. The label generation unit is used to perform weighted summation of normalized features to obtain comprehensive feature values, obtain dynamic labels through equally divided interval mapping, and construct a dynamic label time series set.

[0029] The time-series feature discrimination module includes a separation degree calculation unit and a time-series state determination unit; The deviation calculation unit calculates the temporal deviation between historical dynamic tags and the current dynamic tag, the average temporal deviation, and the characteristic fluctuation value; The timing state determination unit is used to compare the calculation results with the corresponding threshold to determine whether the timing feature of the label is a stable timing feature or an abnormal timing feature.

[0030] The intervention node location module includes a response time period construction unit and a change rate analysis unit; The response time period construction unit traces back to the optimal response time period with the time when the anomaly occurred as the endpoint. The rate of change analysis unit is used to calculate the relative rate of change of adjacent labels, identify local maxima in the sequence, and determine the corresponding intervention nodes.

[0031] The temporal deduplication intervention module includes an overlap calculation unit and a duplicate detection execution unit; The overlap calculation unit calculates the intervention overlap at each time point within the temporal deduplication window and constructs the overlap sequence; The duplicate detection execution unit is used to compare the degree of overlap with the corresponding threshold, determine the duplicate intervention status, and execute the process termination or decision intervention operation, while updating the historical intervention records synchronously.

[0032] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0033] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic tag based decision optimization, the method comprising: The method includes the following steps: ​ S1. Collect multi-dimensional time-series feature data reported by user devices according to a preset cycle, generate dynamic tags based on the multi-dimensional time-series feature data, and construct a dynamic tag time-series set in chronological order; S2. Compare the current dynamic tag with the historical tags in the dynamic tag time series set to calculate the time series deviation and feature fluctuation value of the tag within the continuous window, thereby distinguishing the stable time series features and abnormal time series features of the tag. S3. When an abnormal time series feature is identified, starting from the time when the anomaly occurs, trace back to the time series interval corresponding to the length of the continuous window in S2, and determine this time series interval as the optimal response period; calculate the relative rate of change of dynamic labels of adjacent collection times within the optimal response period, and construct a rate of change sequence; identify local maxima in the rate of change sequence, and take the next time point corresponding to the latest local maxima as the intervention node; if there is no local maxima, take the next time point corresponding to the maximum relative rate of change as the intervention node; when a stable time series feature is identified, no additional intervention operation is performed, and data is collected and anomaly detection is performed according to a fixed period. S4. Perform time-series deduplication detection on intervention nodes. If the same object has been detected to have completed the intervention within the preset time window, cancel the current intervention process. If no duplicate execution record is detected, carry out decision intervention based on the determined intervention nodes.

2. The method of claim 1, wherein: S1 includes the following: Multi-dimensional time-series feature data of user terminals and business service systems are collected according to a preset collection cycle. The multi-dimensional time-series feature data includes terminal interaction event features, session behavior quantification features, system resource access features, and device network status features. Multi-dimensional time-series feature data collected at the same time are used to form the original feature vector of the corresponding time. Each type of feature data in the original feature vector is normalized and the values ​​of each type of feature are uniformly mapped to the range of 0 to 1. Each type of feature data that has undergone normalization is assigned a preset weight coefficient. Each feature data is multiplied by its corresponding weight coefficient, and all the product results are summed to obtain the comprehensive feature value at the current time. The sum of all weight coefficients is one, and the weight coefficients are determined by statistical methods based on historical data. The overall range of values ​​of the comprehensive feature value is divided into equal intervals based on the preset total number of dynamic tags, forming a continuous numerical interval with the same number of dynamic tags. The continuous numerical intervals are assigned positive integer numbers starting from 1 in ascending order, and different intervals correspond to different dynamic tags. The dynamic label corresponding to the numerical range into which the comprehensive feature value falls is used as the dynamic label of the current moment. The dynamic labels generated at each moment are associated with and stored with the corresponding timestamps, and arranged in chronological order of collection time to form a dynamic label time series set.

3. The decision optimization method based on dynamic labels according to claim 2, characterized in that: S2 includes the following: Calculate the temporal deviation between each historical dynamic tag and the current dynamic tag within the continuous window. The temporal deviation is equal to the absolute value of the difference between the value of the historical dynamic tag and the value of the current dynamic tag. The length of the continuous window is greater than 1, and the continuous window contains the dynamic tag at the current acquisition time and several historical dynamic tags before the current acquisition time. The number of historical dynamic tags is equal to the length of the continuous window minus one. The temporal deviations corresponding to all historical dynamic tags within the continuous window are summed to obtain the total deviation. The total deviation is then divided by the number of historical dynamic tags within the continuous window to obtain the average temporal deviation, where the number of historical dynamic tags is equal to the length of the continuous window minus one. Calculate the difference between each time series deviation and the average time series deviation within the continuous window, sum the squares of each difference to obtain the sum of squares; divide the sum of squares by the number of historical dynamic tags, and then take the square root of the division result to obtain the characteristic fluctuation value. A preset deviation threshold and a fluctuation threshold are set, both of which are determined by statistical analysis of historical dynamic tag time-series data; When the average time series deviation is less than or equal to the deviation threshold, and the feature fluctuation value is less than or equal to the fluctuation value threshold, the label time series feature within the continuous window is determined to be a stable time series feature. When the average time series deviation is greater than the deviation threshold, or the feature fluctuation value is greater than the fluctuation value threshold, the label time series feature within the continuous window is determined to be an abnormal time series feature.

4. The decision optimization method based on dynamic labels according to claim 3, characterized in that: S3 includes the following: When the timing characteristics of the tags within a continuous window are determined to be abnormal timing characteristics, the acquisition time when the abnormality occurs is taken as the endpoint, and a timing interval is formed by tracing back the window length by acquisition time. This timing interval is determined as the optimal response period. The window length is the same as the length of the continuous window defined in S2. Traverse the dynamic tag sequence within the optimal response period and calculate the relative change rate of dynamic tags at adjacent acquisition times. The relative change rate is obtained by taking the absolute value of the difference between adjacent tags and dividing it by the larger of the two adjacent tags. Arrange the calculated relative change rates in chronological order to form a relative change rate sequence. The relative rate of change sequence is compared point by point to identify the numerical points that meet the local maximum condition. The criterion for determining the local maximum is that the current relative rate of change value is greater than the relative rate of change values ​​of the adjacent previous time and the adjacent next time. The beginning and end points of the relative rate of change sequence are not involved in the local maximum determination. Only the points in the middle of the sequence are judged for maximum value. Record the time information corresponding to all numerical points that satisfy the local maximum condition, select the one with the largest time value from the recorded results, and use the next data collection time point calculated according to the fixed collection cycle as the intervention node. When there are no numerical points in the relative rate of change sequence that satisfy the local maximum condition, all values ​​in the sequence are compared to determine the time information corresponding to the largest value, and the next data collection time point calculated according to the fixed collection cycle is used as the intervention node. When stable temporal features are identified, the original collection cycle remains unchanged, and the multi-dimensional temporal feature collection and temporal feature comparison operations are continued. No intervention node determination or decision-making intervention related operations are performed.

5. The decision optimization method based on dynamic labels according to claim 4, characterized in that: S4 includes the following: Traverse all historical acquisition times within the time-series deduplication window, and calculate the intervention overlap degree corresponding to each time. The intervention overlap degree is equal to the absolute value of the time difference between the intervention node time and the corresponding historical time, divided by the length of the time-series deduplication window. The length of the time-series deduplication window is greater than 1, and the time-series deduplication window contains several historical acquisition times before the intervention node time. The number of historical acquisition times is equal to the length of the time-series deduplication window. All intervention overlap rates are arranged in chronological order of data collection time to form an overlap rate sequence; the overlap rate sequence is numerically compared to identify numerical items in the sequence that are less than a preset overlap rate threshold; the overlap rate threshold is determined based on the statistical results of historical intervention effectiveness data. If there are numerical items with a value less than the overlap threshold in the time-series deduplication window, it is determined that the same object has completed the intervention within the corresponding time range, and the current intervention process is terminated. If there are no numerical items with a value less than the overlap threshold in the time-series deduplication window, it is determined that there are no duplicate intervention records in the detection interval. The decision intervention process is initiated based on the intervention node time, and the object identifier and execution time corresponding to this intervention are recorded and updated to the historical intervention record set.

6. A decision optimization system based on dynamic labels, applied to the decision optimization method based on dynamic labels according to any one of claims 1-5, characterized in that: The system includes: a dynamic label time series construction module, a time series feature discrimination module, an intervention node localization module, and a time series deduplication intervention module; The dynamic tag time series construction module collects multi-dimensional time series feature data reported by user devices according to a preset period, generates dynamic tags based on the multi-dimensional time series feature data, and constructs a dynamic tag time series set in chronological order. The temporal feature discrimination module compares the current dynamic tag with the historical tags in the dynamic tag temporal set, calculates the temporal deviation and feature fluctuation value of the tag within the continuous window, and distinguishes the stable temporal features and abnormal temporal features of the tag accordingly. When the intervention node positioning module identifies abnormal temporal features, it traces back a temporal interval corresponding to the length of a continuous window, with the time of the abnormal occurrence as the endpoint, and determines this temporal interval as the optimal response period. It calculates the relative rate of change of dynamic tags at adjacent acquisition times within the optimal response period and constructs a rate of change sequence. It identifies local maxima in the rate of change sequence and takes the next time point corresponding to the latest local maxima as the intervention node. If there is no local maxima, the next time point corresponding to the maximum relative rate of change is taken as the intervention node. When the identified temporal features are stable, the intervention node positioning operation is not performed. The time-series deduplication intervention module performs time-series deduplication detection on intervention nodes. When it detects that the same object has completed the intervention within a preset time window, the current intervention process is canceled. When no duplicate execution record is detected, decision intervention is carried out based on the determined intervention nodes.

7. A decision optimization system based on dynamic labels according to claim 6, characterized in that: The data dynamic label time-series construction module includes a feature acquisition unit and a label generation unit; The feature acquisition unit is used to collect terminal interaction event features, session behavior quantification features, system resource access features and device network status features, and construct the original feature vector to complete the feature data normalization process. The label generation unit is used to perform weighted summation of normalized features to obtain comprehensive feature values, obtain dynamic labels through equally divided interval mapping, and construct a dynamic label time series set.

8. A decision optimization system based on dynamic labels according to claim 6, characterized in that: The time-series feature discrimination module includes a separation degree calculation unit and a time-series state determination unit; The deviation calculation unit calculates the temporal deviation, average temporal deviation, and characteristic fluctuation value between the historical dynamic tag and the current dynamic tag; The timing state determination unit is used to compare the calculation result with the corresponding threshold to determine whether the timing feature of the label is a stable timing feature or an abnormal timing feature.

9. A decision optimization system based on dynamic labels according to claim 6, characterized in that: The intervention node positioning module includes a response time period construction unit and a rate of change analysis unit; The response time period construction unit traces back to form the optimal response time period with the time when the anomaly occurs as the endpoint. The rate of change analysis unit is used to calculate the relative rate of change of adjacent labels, identify local maxima in the sequence, and determine the corresponding intervention nodes.

10. A decision optimization system based on dynamic labels according to claim 6, characterized in that: The time-series deduplication intervention module includes an overlap calculation unit and a duplicate detection execution unit; The overlap calculation unit calculates the intervention overlap at each time point within the temporal deduplication window and constructs an overlap sequence; The duplicate detection execution unit is used to compare the overlap with the corresponding threshold, determine the duplicate intervention status, and execute the process termination or decision intervention operation, while updating the historical intervention records synchronously.

Images