A customer portrait matching positioning method and system

Abstract

The application discloses a customer portrait matching positioning method and system, and particularly relates to the technical field of portrait matching positioning. The method performs cleaning and time correction on multi-source behavior data under a unified time base, and constructs behavior time series and label time series. On this basis, a label high-intensity activation interval is identified, and a multi-time scale portrait structure composed of a long-term structural portrait, a life cycle stage portrait and an event portrait is formed. According to the time sensitivity and risk tolerance of a business task, a comprehensive analysis index is calculated to determine whether to introduce time-related dynamic attenuation, and each portrait layer score is obtained and fused according to the weight to generate a comprehensive matching score and output a target customer set. Feedback event sequences are constructed in combination with recommendation, marketing or risk control feedback, and the multi-time scale portrait and the fusion weight are adjusted. The method can distinguish long-term characteristics, stage behavior and short-term events, and improve the accuracy and interpretability of customer portrait matching positioning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image matching positioning, and more particularly to a customer image matching positioning method and system. BACKGROUND

[0002] With the development of online businesses such as e-commerce transactions, content browsing, search services, and third-party advertising, enterprises can continuously collect multi-source behavior data from multiple business systems and external cooperation channels. After time correction and cleaning, these data can be organized into customer behavior time series, and on this basis, a timestamped label time series can be constructed, providing a more detailed input basis for customer image modeling.

[0003] Existing customer image matching solutions are mostly based on statistical results of labels on a single time scale, making it difficult to distinguish between long-term structural characteristics, behaviors with stage concentration and explosive characteristics, and short-term single events with significant business significance. There is also a lack of systematic mechanisms to combine environmental events and customer basic attributes, classify behavior segments semantically, and abstract them into life cycle stages. Therefore, it is easy to have problems such as short-term promotions masking long-term preferences or stage status being difficult to express in the image.

[0004] On the other hand, existing technologies usually use fixed time windows or simple time decay rules set by experience when dealing with time factors. Quantitative indicators such as label time instability index and stage behavior proportion index have not been introduced, making it difficult to decide whether to enable time-dependent dynamic decay according to the time sensitivity and risk tolerance of different tasks. At the same time, for the result feedback generated by recommendation, marketing or risk control, there is a lack of ability to align feedback event sequences in a unified time base and adjust multi-time scale images and the fusion weight of each image layer accordingly, making the image matching result still insufficient in accuracy, time stability and explainability. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a customer image matching positioning method and system to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A customer image matching positioning method, comprising the following steps:

[0008] Cleaning and time correcting multi-source behavior data under a unified time base to construct behavior time series and label time series;

[0009] identify a high-intensity activation interval of the label based on a label time sequence, form a multi-time scale portrait structure, and the multi-time scale portrait structure includes a long-term structural portrait, a life cycle stage portrait, and an event portrait;

[0010] analyze the time sensitivity and risk tolerance of the business task, calculate a comprehensive analysis index based on a label time instability index and a stage behavior proportion index, determine whether to introduce a time-related dynamic attenuation based on the comprehensive analysis index, calculate a long-term structural portrait score, a life cycle stage portrait score, and an event portrait score, and generate a comprehensive matching score by fusing the portrait layer scores according to weighted summation, and output a target customer set based on the comprehensive matching score;

[0011] Construct a feedback event sequence in combination with feedback from recommendation, marketing, or risk control, and adjust the multi-time scale portrait and fusion weight according to the feedback event sequence.

[0012] In a preferred embodiment, the process of identifying a high-intensity activation interval of the label based on a label time sequence includes: using a sliding time window of a preset length to scan along the time axis of the label time sequence; comparing the label activation intensity in the window with the historical baseline intensity of the corresponding customer; when the label activation intensity in a certain continuous time interval is continuously and significantly raised relative to the historical baseline, marking the time interval as a high-intensity activation interval of the label, and recording its start and end time and peak intensity.

[0013] In a preferred embodiment, forming a multi-time scale portrait structure includes: aligning the high-intensity activation intervals of the customer on each label dimension along the time axis, identifying stage event candidate segments; combining environmental event sequences and basic attributes of the customer to semantically classify the stage event candidate segments, and construct a life cycle stage portrait; excluding high-intensity segments within the life cycle event influence window from the label time axis, and constructing a long-term structural portrait based on the label activation records of the remaining time period.

[0014] In a preferred embodiment, constructing a life cycle stage portrait includes: for the identified life cycle event types, recording the type, start time, end time, stage intensity, and related label set of the corresponding life cycle event for the customer as the granularity; when the same type of life cycle event occurs multiple times on the behavior time axis of the customer, connecting these life cycle events in chronological order to form a stage trajectory.

[0015] In a preferred embodiment, the comprehensive analysis index is calculated based on the label time instability index and the periodic behavior proportion index, including: weighting and fusing the label time instability index and the periodic behavior proportion index to obtain a comprehensive analysis index value; comparing the comprehensive analysis index value with a preset dynamic attenuation enabling threshold, and determining whether to introduce time-related dynamic attenuation when calculating the portrait layer score based on the comparison result.

[0016] In a preferred embodiment, the label time instability index is used to reflect the average level of the difference in activation intensity between the recent observation window and the long-term historical window for the label concerned in the current task; and the periodic behavior proportion index is used to measure the proportion of the behavior amount falling within the life cycle stage influence window or the short-term event influence window in the overall behavior within the current task and the candidate customer range.

[0017] In a preferred embodiment, the portrait layer score is calculated respectively, including: calculating a long-term structural portrait score based on the similarity between the label typical value of the customer in a long-term range and the target portrait template;

[0018] calculating a life cycle stage portrait score based on the affinity between the current life cycle stage of the customer and the business task, and introducing a timeliness weight when time-related dynamic attenuation is enabled;

[0019] calculating an event portrait score based on the relevance between the single event recently occurred by the customer and the business task, and introducing a time attenuation weight when time-related dynamic attenuation is enabled.

[0020] In a preferred embodiment, the feedback event sequence includes customer feedback data from recommendation, marketing or risk control operations, which is aligned with the behavior time sequence after desensitization and formatting processing.

[0021] In a preferred embodiment, the feedback event sequence adjusts the fusion weight, including: performing summary analysis on the comprehensive performance of several similar business tasks; and slowly adjusting the fusion weight ratio of the long-term structural portrait, the life cycle stage portrait and the event portrait through iterative learning based on the consistency degree between the comprehensive matching score and the actual result.

[0022] In a preferred embodiment, a customer portrait matching positioning system includes the following modules:

[0023] A data acquisition and processing module is used to clean and time correct multi-source behavior data under a unified time base, and to construct a behavior time sequence and a label time sequence;

[0024] An image modeling module is configured to identify a label high-intensity activation interval based on the label time sequence, form a multi-time scale image structure, and the multi-time scale image structure includes a long-term structural image, a life cycle stage image, and an event image;

[0025] A matching calculation module is configured to analyze time sensitivity and risk tolerance for a business task, calculate a comprehensive analysis index based on a label time instability index and a stage behavior proportion index, determine whether to introduce time-related dynamic attenuation based on the comprehensive analysis index, calculate a long-term structural image score, a life cycle stage image score, and an event image score, fuse image layer scores according to weighted summation, and generate a comprehensive matching score; and output a target customer set based on the comprehensive matching score.

[0026] A feedback adjustment module is configured to construct a feedback event sequence in combination with recommendation, marketing, or risk control feedback, and adjust the multi-time scale image and the fusion weight according to the feedback event sequence.

[0027] Technical effects and advantages of the present application:

[0028] The present application performs time calibration and cleaning on multi-source behavior data from e-commerce transaction systems, content browsing systems, search log systems, third-party advertisement delivery systems, and external cooperation channel interfaces under a unified time base, constructs behavior time sequences and label time sequences with timestamps, and forms multi-time scale image layers such as long-term structural images, life cycle stage images, and event images based on label high-intensity activation intervals, thereby distinguishing long-term stable features, stage concentrated burst behaviors, and single events with significant business significance within the same image framework, avoiding misjudgment of short-term activities or accidental events as long-term preferences, and significantly improving the accuracy and robustness of customer image matching positioning.

[0029] The present application introduces a label time instability index and a stage behavior proportion index to quantitatively describe the time stability and stage behavior proportion of the label within an observation window, and determines whether to enable a time-related dynamic attenuation strategy in combination with time sensitivity, observation window length, and risk tolerance of the business task. When needed, the contribution of long-term structural image scores, life cycle stage image scores, and event image scores from expired behaviors or behaviors at the boundary of the influence window is decayed and corrected, and when not needed, the baseline scores in the layer are maintained or only weak decay is implemented, thereby achieving differentiated control of time factors, making the matching result sensitive to recent changes in the time dimension, and avoiding excessive response to short-term noise, thereby improving the time stability and adaptability of the system.

[0030] The application calculates long-term structural portrait scores, life cycle stage portrait scores and event portrait scores respectively, and fuses the three types of scores by weighting according to the weight set by the task portrait template to obtain a comprehensive matching score. When making recommendation, marketing or risk control decisions, the upper business system can directly use the comprehensive matching score to sort and select customers, or can analyze the reasons why customers are selected or excluded based on the scores of each portrait layer, which significantly improves the explainability of the matching result and is beneficial to the strategy verification and decision explanation of operation and risk control personnel. Meanwhile, the application constructs the recommendation, marketing or risk control execution result as a feedback event sequence, aligns the behavior time sequence and the label time sequence under a unified time base, and iteratively adjusts the fusion weight ratio of the multi-time scale portrait structure and the three types of portrait layers based on the consistency between the comprehensive matching score and the actual feedback, so that the system can automatically correct the influence of each portrait layer in the process of continuously introducing real business feedback, form a closed-loop optimization mechanism, and reduce the dependence on manual parameter adjustment and experience rules. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to facilitate the understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings;

[0032] Figure 1 FIG. 1 is a flowchart of a customer portrait matching positioning method according to the present application;

[0033] Figure 2 FIG. 2 is a structural diagram of a customer portrait matching positioning system according to the present application. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0035] Embodiment 1, a customer portrait matching positioning method according to the present application, as shown in FIG. 1, includes the following steps: Figure 1

[0036] Step one: label time sequence construction

[0037] In an embodiment of the customer portrait matching positioning method according to the present application, the purpose of step one is to arrange the customer behaviors scattered in different business systems into a behavior time sequence on a unified time axis, and on this basis, to construct a label time sequence, providing a structured input for subsequent multi-time scale portrait modeling.

[0038] ​Specifically, within a pre-set data collection period, without touching the customer's privacy, the behavior data and transaction data related to the customer are continuously collected from the e-commerce transaction system, content browsing system, search log system, third-party advertisement delivery system, and external cooperation channel interface; the above-mentioned behavior data can include the search keywords and their triggering time of the customer on the platform, the exposure record and the stay time of the browsing page or content, the operation of adding goods to the shopping cart, the ordering and payment behavior, the rating and evaluation of goods or services, the business operation record of after-sales and refund application, and the basic attribute information and terminal device environment information associated with the customer account.

[0039] Since the above-mentioned behavior data often comes from different systems, covers different regions and may be in different time zones, this step first corrects the time stamp of all behavior records. The system converts the local time recorded in each business system into a unified time base through time zone conversion and format standardization processing according to the pre-set standard time reference; on this basis, taking the customer main identifier as the aggregation key, the multi-source behavior records of the same customer in the e-commerce transaction system, content browsing system, search log system and third-party advertisement system are merged, and sorted from early to late according to the corrected time stamp. The behavior time sequence of a certain customer is denoted as ;

[0040] wherein u represents the customer main identifier, represents the time stamp of the kth behavior record under the unified time base, represents the business content or operation type related to the behavior, is the number of behaviors of the customer within the collection period. In the process of constructing the sequence , the obviously abnormal time stamp, missing time field and repeated record are cleaned and corrected to ensure that the position of each behavior on the unified time base is clear and traceable, thereby forming a standardized behavior data set with customer ID and behavior time sequence as the basic unit. The customer ID can be set as a virtual number, etc., without touching the customer's privacy.

[0041] The environment-related event information is collected from the operation management system and external public data sources; the environmental events mainly include large-scale promotion activity period within the platform, cross-brand joint activity, statutory holidays, major social events, etc. For each type of environmental event, the system records its effective start time, effective end time and event type, and converts it to the unified time base to form an environmental event sequence parallel to the behavior time sequence.

[0042] After obtaining the standardized behavior time series and the environment event series, this step maps the original behavior records to the activation records at the label level according to a pre-designed label system. The label system can be divided into different dimensions such as interest preference labels, consumption category labels, risk characteristic labels, and environment-related labels, and each dimension contains a plurality of specific labels. For each record in the behavior time series , the system looks up and matches the corresponding label code according to the characteristic fields such as the product category, brand ownership, price interval, payment method, transaction channel, and content theme involved, marks the behavior on one or more labels related to it as activated, and retains the original behavior timestamp , forming a label activation record with time information. For the case where one behavior triggers multiple labels, this step does not compress and merge, but retains the correspondence between multiple label activations, so as to analyze the label co-occurrence pattern and label combination characteristics in the subsequent analysis.

[0043] After completing the label mapping, this step constructs a label time series with a customer as the granularity and a single label as the observation object under a unified time basis. For any customer u and label , the system accumulates all records of the activation of the label in chronological order on the behavior time axis of the customer, forming a sequence , where represents the jth activation time of the label in the customer u dimension, is the number of activations of the label in the collection period. Around the above sequence, this step calculates and saves the first activation time, the last activation time, the cumulative activation number, the statistical characteristics of the adjacent activation time interval, and the activation intensity in the preset time slice of the label in the customer dimension. Activation intensity can be obtained by comprehensively considering the number of activations of the label in the time slice, the amount of related behavior, and the duration, etc., and is used to describe the activity level of the label at different time periods. After obtaining the label time series, the system scans along the time axis using a sliding time window of a pre-set length, compares the activation intensity in the window with the historical baseline intensity of the customer, and when the activation intensity in a certain continuous time interval is continuously and significantly higher than the baseline, marks the time interval as a high-intensity activation interval of the label, and records its start and end time and peak intensity in the label time series.

[0044] Step 2: Stage behavior portrait modeling

[0045] Step two is to further identify the behavior segments with stage characteristics based on the label time series view obtained in step one, and build a multi-time scale hierarchical portrait structure on this basis, so that long-term preferences, stage status and short-term events can be expressed in the portrait respectively, rather than simply superimposed together. For ease of understanding, step two can be understood as first finding out which behaviors on the timeline are concentrated in a certain stage, and then modeling these stages and daily, relatively stable behaviors separately.

[0046] Specifically, this step first aligns the high-intensity activation interval of each customer in each label dimension on the time axis. For the same customer, if multiple semantically related labels show significant intensity uplift in the label time series at the same time within a certain continuous time window, or closely connected in time, the behavior segment corresponding to this time window is regarded as a stage event candidate segment. In order to quantify the degree of this concentrated outbreak, the stage intensity of a certain customer u, a certain label set L in the time interval can be calculated under a unified time basis, for example, a simple intensity index is introduced.

[0047] Among them, and respectively represent the start and end time of the segment, represents the basic intensity corresponding to a label activation in this time interval; for example, for browsing behavior, the value of the normalized duration of this browsing can be used; is the weight coefficient of different labels or different behavior types, for example, the weight of payment success behavior can be set to 1.0, the weight of adding to cart behavior can be set to 0.7, and the weight of product detail page browsing behavior can be set to 0.3; is the duration of the segment. By comparing with the baseline intensity of the customer in a longer historical window, it can be directly identified which segments belong to stage abnormal activity, so as to mark them as stage event candidates.

[0048] Combine the sequence of environmental events and the basic attributes of the customer to semantically classify the candidate segments; here, semantic classification means not only to see which category or interest the label itself belongs to, but also to see whether it is superimposed with certain environmental factors when it occurs; this step classifies the segments that meet the characteristics into different life cycle event types according to the pre-agreed pattern library, such as parenting stage, decoration stage, preparation stage, treatment stage, etc., and configures reasonable influence window and decay period for each type of life cycle event.

[0049] After the lifecycle event identification is completed, the lifecycle stage portrait is constructed at the granularity of a customer. For the same customer, if the same type of lifecycle event appears multiple times in its behavior timeline, these events are connected in chronological order to form a stage trajectory. Each stage trajectory usually contains the type of lifecycle event, the start and end time of the stage, the intensity of the stage calculated based on the aforementioned intensity indicators, the impact radius inferred from business experience or models, and a set of labels that are significantly related to the stage, etc. By maintaining such stage trajectories, the present application can explicitly record the life scenes and stage states of customers at different periods in the portrait layer, so that subsequent matching and positioning no longer rely solely on what this person likes, but can also answer what stage this person is in now. The lifecycle stage portrait, as a relatively independent layer, is related to the label time series and the environmental event sequence, providing a structure that reflects the medium and long-term state for multi-time scale portraits.

[0050] From the label timeline, the high-intensity segments that have been labeled as the lifecycle event impact window are removed, and only the label activation records in the remaining time period are retained. For these remaining records, this step no longer emphasizes the burst in a short period of time, but focuses on the average level and relative ranking over a longer time span. Through time-weighted average, quantile statistics or other robust statistical methods, the typical value and activity level of each label in the long-term range are calculated.

[0051] On the other hand, for those single events that do not form a complete lifecycle stage but have significant business meaning in a short period of time, this step extracts them to construct a separate event portrait layer. For such events, this step records the event occurrence time, event type, event intensity, and a set of labels related to the event, and configures appropriate decay strategies for related labels to have an impact on matching calculations within a reasonable time window after the event occurs, and gradually fade out after exceeding the window.

[0052] Step three: task matching score calculation;

[0053] Step three is to generate a target customer set that can be directly used for execution around specific business demands on the premise that multi-time scale portraits have been constructed. To this end, when a matching positioning request is initiated by the upper-level business system, the system does not simply make a static screening of the existing portraits, but first analyzes the business information carried by the request. The request usually contains business target types, requirements for time sensitivity, acceptable risk levels, and portrait templates describing ideal target customer characteristics. Business target types can be short-term promotional activities, phased benefit operations, annual value stratification assessment, or risk screening, etc. Time sensitivity reflects whether the current task places more emphasis on behavior performance in the recent period or relies more on long-term stable characteristics. Risk tolerance is used to constrain the fluctuation range of the matching results in terms of conversion rate, bad debt rate, or complaint rate. The portrait template gives an approximate outline of what kind of customers are expected to be found in the form of tags and feature intervals. The system derives the basic weight ratio of the three portrait layers, i.e. long-term structural portrait, life cycle stage portrait, and event portrait, in subsequent calculations according to these information, for example, appropriately increasing the weights of event portrait and current life cycle stage in short-term activities before holidays, and increasing the weight of long-term structural portrait in annual customer stratification scenarios.

[0054] After completing the above task analysis, the system selects a candidate customer set from the portrait storage. The scope of the candidate set can be determined according to the constraints set by the business party in advance. After the candidate set is determined, the matching degree between each candidate customer and the target portrait template needs to be calculated on the three portrait levels. For each portrait layer, calculate the long-term structural portrait score, life cycle stage portrait score, and event portrait score.

[0055] In an optional implementation of the present application, whether to introduce a time-related dynamic attenuation mechanism in the matching score calculation process of each portrait layer is based on the calculation of the label time instability index and the stage behavior proportion index , by fusing the two indexes in a weighted manner to construct a comprehensive analysis index J for decision-making, which can be defined as: ; wherein the weight coefficients and , + =1, for example =0.4, =0.6;

[0056] The larger the value, the higher the short-term hotspot warming degree and the stage behavior coverage ratio, indicating that short-term behavior and stage events have more significant disturbance on the overall portrait in the current task.

[0057] After obtaining the comprehensive analysis index J, a dynamic decay enabling threshold associated with the business scenario is introduced in this step , which also takes values in [0, 1] and can be set through offline simulation, historical task backtracking analysis and operation experience, and can be moderately adjusted according to feedback during long-term system operation. For task types with high time sensitivity, , a relatively low value can be set so that time-related dynamic decay is enabled in time once a certain degree of time instability and periodic behavior concentration is detected; for tasks mainly focusing on annual value assessment and long-term portrait stratification and not sensitive to short-term fluctuations, , a relatively high value can be set so that the static portrait is still the main one under most circumstances, and decay is only enabled when short-term disturbance is extremely significant. In actual operation, the system compares the J calculated for the current task with the threshold ; when J≥ , it is determined that the influence of short-term hotspots and periodic behaviors on label distribution in the current task has reached a level that needs to be focused on, and there is a high risk of misfixing short-term behaviors as long-term preferences, so in the layer score calculation, , the label contribution in the expired or edge of the influence window is decayed according to the preset time decay curve; when J< , it means that the label is overall stable in the time dimension, and the coverage of periodic behaviors and short-term events in the observation window is limited, and the current task can be safely considered as dominated by long-term structural features, so the system can directly use the layer baseline score without time decay, or only implement weak decay in the event portrait layer, thereby avoiding unnecessary time modeling overhead while ensuring matching accuracy.

[0058] , which is used to reflect the average level of the difference in activation intensity between the recent observation window and the long-term historical window for the label concerned in this task. By comparing the recent and historical label activation frequency, associated amount or dwell time, etc. , when most key labels are significantly warmer in the recent period and in the long-term low baseline state, the value of tends to increase; when the recent behavior is basically consistent with the long-term distribution or slightly decreases, the value of the index is smaller. Normalize to the interval [0, 1], the closer the value is to 1, the more significant the disturbance of short-term hotspots to label distribution, and the higher the risk of long-term structural features being obscured by short-term behaviors.

[0059] Periodic behavior proportion index The proportion of the behavior quantity falling into the life cycle stage influence window or the short-term event influence window in the overall behavior quantity in the current task and the candidate customer range. The system divides the behavior in the observation window into stage behavior and non-stage behavior based on the identified life cycle stage portrait and event portrait, and calculates the proportion of the stage behavior in the total behavior quantity according to this. The index can also be normalized to the interval [0, 1], and the larger the value is, the more the current business occurs in the time segment dominated by the specific stage or stage event, and the stronger the stage behavior dominates the current task; when the value is low, it indicates that most of the behavior occurs in the stable period, and the influence of the life cycle stage and the short-term event is relatively limited.

[0060] Long-term structural portrait score The acquisition aims to measure the consistency of the customer with the target portrait template in the steady-state characteristics, and the score is obtained by calculating the similarity between the typical value of the customer in the long-term range and the template requirements. Specifically, the long-term structural portrait layer contains a set of labels , each of which has a long-term typical value in the customer u dimension obtained by time-weighted average or quantile statistics , which reflects the steady-state level after excluding short-term fluctuations; the target portrait template defines an expected value or reasonable interval for each label . The calculation formula of the score is: ;

[0061] wherein, is the weight coefficient of the label , and can be preset by the system administrator according to the importance of the label in the business; is a similarity function for quantifying the closeness of the customer value and the target value; as a preferred implementation, the similarity function is based on normalized absolute difference calculation, and the specific mathematical form is as follows: ; wherein, represents the long-term typical value of the customer u in the label ; represents the expected value or reference value defined by the target portrait template for the label ; is a very small positive number, for example =0.00000001, to avoid the case that the denominator is zero.

[0062] Life cycle stage portrait score The acquisition of the score is used to evaluate the dynamic fit degree between the current life cycle stage of the customer and the business task, and the calculation significantly reflects the timeliness of the stage. Suppose that the customer u is currently in a set of life cycle stages , each stage records its type, start time , end time , i.e. the impact window, stage intensity and a set of related labels . The calculation formula of the stage intensity H is: ;

[0063] wherein, represents the basic intensity of a label activation in the time interval , and is the weight coefficient of different behavior types. In order to reflect the dynamic influence of the stage, the calculation of the score introduces a timeliness weight. When J≥ , the formula is: ; when J< , the formula is: ;

[0064] wherein, is the affinity of the stage p to the current business task, which can be pre-configured in a mapping table; for example, for the parenting stage and the infant milk promotion task, the affinity can be set to 0.9; for the decoration stage and the same task, the affinity can be set to 0.1; is the timeliness weight of the stage p at the current calculation time , for example, a linear decay function is used for calculation: ; wherein is a preset decay buffer length, for example, 30 days.

[0065] The acquisition of the event portrait score aims to measure the relevance of a single significant event recently occurred by the customer to the business task, and strictly control its short-term influence through a decay mechanism. Suppose that the customer u has recently occurred a set of events , each event contains event type, occurrence time , event intensity and a set of related labels . When J≥ , the calculation formula of the score is: ; when J< , the calculation formula of the score is: ;

[0066] wherein, This is the weighting coefficient for event e, which is usually related to the event intensity. Positive correlation; It is the relevance of event e to the current business task. Similar to stage affinity, it can be obtained through a predefined mapping table; It is a time decay weighting function, which can be set as follows: ;

[0067] in, Indicates the current time and the time the event occurred. The interval between This is the attenuation coefficient, used to control the rate at which the weights decay, for example... This formula is typically used to represent the decay constant in an exponential decay process. With half-life Relationship, If half-life The decay constant is calculated for each unit of time (e.g., day, hour). This mechanism ensures that recent events contribute significantly to the score, while their impact diminishes rapidly over time, thus effectively preventing short-term behavior from being misjudged as long-term preferences.

[0068] After time decay correction within each profile layer, the matching scores from the three layers can be fused to obtain a comprehensive matching score tailored to the current business scenario. For ease of description, an optional implementation method defines the following fusion relationship for each candidate customer u: ;

[0069] in, This indicates the overall matching score of customer u under the current task. The score is calculated based on a long-term structural profile. The score is obtained based on the life cycle stage profile. The score is obtained based on the event profile; , , The fusion weights for long-term profiles, phase profiles, and event profiles in this task should meet the following requirements. + + =1, for example , =0.4、 =0.4.

[0070] The candidate set is sorted from highest to lowest score. Combining the pre-set score thresholds by the business stakeholders, the tiered strategies corresponding to different score ranges, and resource constraints, the system sequentially selects customers from the front of the sorted results to form the final target customer set.

[0071] Step four: feedback-driven weight adjustment;

[0072] The role of step four is to realign the calculated results with the actual situation on the business, so that the multi-time scale portrait established before is no longer a one-time static snapshot, but a dynamic structure that can be slowly corrected according to the subsequent behavior of the customer. To this end, after the target customer set is output in step three and the recommendation, marketing, benefit distribution or risk control treatment is completed by the upper system, this step first collects and organizes the execution results of these actions under a unified time base, extracts the information related to customer feedback, and constitutes an independent feedback event sequence. The so-called feedback event can be whether the customer has seen, clicked or ignored a certain recommendation content, whether the customer has generated an order and payment behavior within a limited time, whether the customer has unsubscribed or complained about marketing touch, whether the customer has used the distributed benefits, or whether the credit approval has passed, whether the customer has repaid on time after credit, whether the customer has occurred overdue, and whether the customer has triggered manual review, etc. After these data are desensitized and format standardized, they are reordered according to the customer's main identification and standard time base, so that they can be one-to-one corresponding with the aforementioned behavior time sequence and label time sequence on the time axis.

[0073] After obtaining the feedback event sequence, this step does not directly recalculate the portrait as a whole, but first extracts several quantifiable effect indicators from the feedback to measure whether the matching decisions driven by the portrait are reliable. For example, for a certain matching task, the click rate, conversion rate, unsubscribe rate and complaint rate of customers can be counted, and then these results can be compared by score interval, label combination or life cycle stage. In order to facilitate automatic use in algorithms, a simplified feedback score can be constructed, for example, the comprehensive feedback of a certain customer u in a certain task is defined as: ;

[0074] Among them, can be understood as a conversion-related indicator, such as whether the expected behavior is completed or its probability value, represents the strength of negative feedback such as unsubscribe and complaint, , is a positive weight to reflect the relative importance of positive and negative results in comprehensive evaluation, for example , .

[0075] In the long-term structural portrait layer, this step pays more attention to whether the performance of certain long-term labels in multiple tasks is stable and reliable. For those labels that are used to filter customers in different scenarios and have feedback scores Tags or tag combinations that consistently exceed the preset baseline can be considered to truly reflect the long-term characteristics of the customer, at which point this step will appropriately extend the remaining validity period of these tags and slow down the natural decay rate of their weights over time; on the contrary, for some tags, although they are assigned higher weights during the modeling phase, the corresponding customer group feedback is weak or even negative in multiple rounds of tasks, which triggers the tag aging strategy, gradually weakening its influence in the long-term structural portrait by reducing the weight, shortening the validity period, and if necessary, downgrading it from a strong preference tag to a weak preference tag, or even marking it as temporarily not participating in matching calculations. By strengthening the long-term characteristics that perform well and weakening the long-term characteristics that repeatedly fail, the long-term structural portrait can gradually converge to a more accurate portrayal of the customer's intrinsic attributes in continuous operation, rather than remaining static at the initial definition.

[0076] In the life cycle stage portrait layer, this step focuses on when the stage should end and whether the stage impact is overestimated or underestimated. For a customer currently in a certain life cycle stage, the system will align the stage trajectory corresponding to the stage with the feedback events, and observe whether the matching strategy designed around the stage continues to obtain positive feedback within the stage impact window. If the relevant behavior and feedback have weakened significantly before the end of the theoretical impact window, it can be considered that the customer has crossed the stage prematurely, and this step will shorten the termination time of the stage trajectory and reduce the stage weight in subsequent matching; on the contrary, if the behavior around the stage still occurs frequently near the originally set stage end, and the recommendations or marketing triggered based on the stage portrait still have good performance, the impact window of the life cycle event can be extended accordingly, or the stage strength can be increased, so that it continues to play a role in subsequent matching positioning. In short, the division of the life cycle stage is not fixed once and for all, but needs to be dynamically adjusted in combination with the subsequent real behavior of the customer to adjust the stage boundary and stage strength.

[0077] For the event portrait layer, this step deals with single major events that are very noticeable in the short term but may not have a long-term impact on customer behavior. On the one hand, if the subsequent behavior frequency of related goods or services falls rapidly within the system's preset observation window after a large consumption or major medical expense, and targeted outreach based on the event portrait does not bring sustained significant benefits, it indicates that the event is more of a one-time impulse or accident. This step will accelerate the reduction of the weight of the corresponding tag according to a faster decay curve, so that it exits the decision-making as soon as possible after playing a role in the short term, avoiding being misinterpreted as a long-term preference; on the other hand, if the behavior around the event theme remains at a high level for a period of time after the event, and the feedback score If the event continues to be above the baseline, then this step can moderately slow down its decay rate, or even aggregate it into a new life cycle stage trajectory, and transfer it to the life cycle stage portrait layer for management, so as to reflect that the short-term event has evolved into a relatively stable life stage.

[0078] After completing the local adjustment based on feedback for the three portrait layers respectively, this step also needs to check whether the current multi-time scale weight configuration is still suitable for subsequent similar tasks as a whole. To this end, the comprehensive performance of several similar business tasks can be summarized and analyzed within a certain observation period, and the consistency between the comprehensive matching score and the actual result under different weight combinations is compared. For example, the ideal weight vector for a certain type of task can be set as , and the weight vector in actual operation is When it is found that the event portrait layer contributes too much in a certain type of task and brings more complaints, or the long-term structural portrait layer weight is too small, resulting in long-term value prediction deviation, slow adjustment can be made in the following form: ; wherein, is the learning step, and the value is between 0 and 1, The smaller the value is, the smoother the adjustment is, avoiding the weight from fluctuating sharply in the short term. Through repeated iteration, the weight vector will gradually approach a more suitable configuration under the guidance of historical effects, so that the same type of matching task in the future can perform more stably in the long term.

[0079] Embodiment 2, the design of a customer portrait matching positioning system based on the method in embodiment 1, as shown in Figure 2 includes the following modules:

[0080] The data acquisition and processing module is used to clean and time correct multi-source behavior data under a unified time base, construct behavior time series and label time series.

[0081] Specifically, within a pre-set data collection period, the data collection and processing module continuously collects behavior data and transaction data related to customers from e-commerce transaction systems, content browsing systems, search log systems, third-party advertisement placement systems, and external cooperation channel interfaces, including search keywords and their triggering times, exposure records and dwell times, adding-to-cart operations, ordering and payment behaviors, ratings and reviews, after-sales and refund applications, and account basic attribute information and terminal device environment information; for behavior records from different systems, different regions, and different time zones, the data collection and processing module converts the local time recorded by each business system into a unified time base through time zone conversion and format standardization, converts it into a time representation under the unified time base, and uses the customer master identifier as the aggregation key to merge and sort the multi-source behavior records of the same customer, forming a standardized behavior time sequence; on this basis, according to the label system designed in embodiment 1, the original behavior records in the behavior time sequence are mapped to the activation records at the label level, and the label time sequence is constructed under the unified time base, and the first activation time, the last activation time, the cumulative activation times, the adjacent activation time interval statistical features, and the activation intensity in the pre-set time slice are calculated, providing input for the portrait modeling module to identify the label high-intensity activation interval.

[0082] The portrait modeling module is configured to identify the label high-intensity activation interval based on the label time sequence, and form a multi-time scale portrait structure, which includes a long-term structural portrait, a life cycle stage portrait, and an event portrait.

[0083] Specifically, the portrait modeling module scans the label time series with a preset length of sliding time window along the time axis under a unified time base, compares the label activation intensity with the customer historical baseline intensity, and when the activation intensity relative to the baseline appears sustained and significant rise in a certain continuous time interval, marks the time interval as a high-intensity activation interval of the label; on this basis, the portrait modeling module aligns the high-intensity activation intervals of each label dimension on the time axis for each customer, identifies the candidate fragments of the phased events formed by the concentrated outbreak of multiple semantically related labels within a specific time window, and combines the environmental event sequence and the customer basic attributes to classify the candidate fragments into different life cycle event types according to the mode library pre-agreed in embodiment 1, and construct the life cycle phase portrait and phase track recording the life cycle event type, start time, end time, phase intensity, impact window and related label set; at the same time, the high-intensity fragments within the impact window of the life cycle event that have been labeled are excluded from the label time axis, and the long-term typical value and activity level of the remaining label activation records are calculated by time-weighted average, quantile statistics or other robust statistical methods to form a long-term structural portrait; for single events that do not form a complete life cycle phase but have significant business significance in a short period of time, the portrait modeling module extracts the event occurrence time, event type, event intensity and related label set to construct the event portrait layer, thereby completing the modeling of the multi-time scale portrait structure composed of the long-term structural portrait, the life cycle phase portrait and the event portrait.

[0084] The matching calculation module is used to analyze the time sensitivity and risk tolerance for business tasks, and to calculate a comprehensive analysis index based on the label time instability index and the phased behavior proportion index, and to determine whether to introduce time-related dynamic attenuation based on the comprehensive analysis index; long-term structural portrait scores, life cycle phase portrait scores and event portrait scores are calculated respectively, and the scores of each portrait layer are fused according to weighted summation to generate a comprehensive matching score; and the target customer set is output based on the comprehensive matching score.

[0085] Specifically, the matching calculation module parses the business target type, time sensitivity, risk tolerance, and portrait template configuration information from the matching positioning request initiated by the upper-layer business system after receiving the request, derives the basic weight ratio of the three portrait layers of long-term structural portrait, life cycle stage portrait, and event portrait according to the business target type and time sensitivity; within the candidate customer set, the matching calculation module calculates the label time instability index and the stage behavior proportion index based on the label time series and the multi-time scale portrait structure, wherein the label time instability index is used to reflect the average level of the activation intensity difference between the recent observation window and the long-term historical window for the label concerned in this task, and the stage behavior proportion index is used to measure the proportion of the behavior quantity falling into the life cycle stage influence window or the short-term event influence window in the overall behavior, and the two indexes are fused into a comprehensive analysis index for decision-making in the manner described in Embodiment 1; the matching calculation module compares the comprehensive analysis index with the dynamic decay activation threshold value, when the comprehensive analysis index reaches or exceeds the threshold value, introduces time-related dynamic decay in the matching score calculation process of each portrait layer, and corrects the contribution of the label at the edge of the expiration or influence window, the life cycle stage influence, and the short-term event influence according to the preset time decay curve, when the comprehensive analysis index is lower than the threshold value, the layer-based reference score without time decay or only weak decay in the event portrait layer is used; on this basis, the matching calculation module calculates the long-term structural portrait score, the life cycle stage portrait score, and the event portrait score respectively, and weights and sums the three portrait layer scores according to the fusion weight set in Embodiment 1 to obtain the comprehensive matching score of the candidate customer, sorts the candidate set according to the comprehensive matching score from high to low, combines the score threshold, the gradient strategy corresponding to different score intervals, and the resource limit, selects the target customer set from the sorting result, and outputs.

[0086] The feedback adjustment module is used to construct a feedback event sequence in combination with the recommendation, marketing, or risk control feedback, and adjust the multi-time scale portrait and the fusion weight according to the feedback event sequence.

[0087] Specifically, the feedback adjustment module collects information related to customer feedback under a unified time base after the matching calculation module outputs the target customer set and the upper system completes recommendation, marketing, benefit distribution or risk control disposal, including whether the customer sees, clicks or ignores the recommended content, whether the customer completes the order and payment within a limited time, whether the customer unsubscribes or complains about the marketing touch, whether the credit approval passes, whether the customer repays on time after being granted credit, whether the customer has overdue, and whether the customer triggers manual review, etc. The above feedback information is desensitized and standardized, and is sorted according to the customer main identifier and the unified time base to build a feedback event sequence; the feedback adjustment module extracts the click rate, conversion rate, unsubscribe rate, complaint rate and other effect indicators based on the feedback event sequence, builds a feedback score at the customer level and in the score interval, label combination or life cycle stage, and dynamically adjusts the weight and aging strategy of the long-term structural portrait layer, the influence window and stage intensity of the life cycle stage portrait layer, and the decay strategy of the event in the event portrait layer; at the same time, the feedback adjustment module summarizes the comprehensive performance of multiple similar business tasks within a certain observation period, and slowly updates the fusion weight of the long-term structural portrait, the life cycle stage portrait and the event portrait according to the consistency between the comprehensive matching score and the actual result.

[0088] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented in electronic hardware, computer software and a combination of electronic hardware and computer software. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0089] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division, and actual implementation can have another division manner. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0090] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment of the present application according to actual needs.

[0091] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0092] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A customer profile matching positioning method, characterized in that, The method comprises the following steps: cleaning and time correcting multi-source behavior data under a unified time base to construct a behavior time sequence and a label time sequence; identifying a label high-intensity activation interval based on the label time sequence to form a multi-time scale portrait structure, the multi-time scale portrait structure comprising a long-term structural portrait, a life cycle stage portrait and an event portrait; wherein forming the multi-time scale portrait structure comprises: time axis alignment of the high-intensity activation interval of the customer in each label dimension to identify a stage event candidate segment; semantic classification of the stage event candidate segment in combination with an environmental event sequence and basic attributes of the customer to construct the life cycle stage portrait; elimination of the high-intensity segment located within a life cycle event influence window from the label time axis to construct the long-term structural portrait based on the label activation record of the remaining time period; extraction of a single event that does not form a complete life cycle stage but has significant business significance to construct the event portrait; analyzing time sensitivity and risk tolerance for a business task, and calculating a comprehensive analysis index based on a label time instability index and a stage behavior proportion index to determine whether to introduce time-related dynamic attenuation based on the comprehensive analysis index; calculating a long-term structural portrait score, a life cycle stage portrait score and an event portrait score, and fusing the scores of each portrait layer according to weighted summation to generate a comprehensive matching score; and outputting a target customer set based on the comprehensive matching score; combining feedback from recommendation, marketing or risk control to construct a feedback event sequence, and adjusting the multi-time scale portrait and the fusion weight according to the feedback event sequence.

2. The customer profiling and targeting method of claim 1, wherein: The process of identifying a label high-intensity activation interval based on the label time sequence comprises: scanning along the time axis of the label time sequence using a sliding time window of a preset length; comparing the label activation intensity within the window with the historical baseline intensity of the corresponding customer; when the label activation intensity in a certain continuous time interval is continuously and significantly lifted relative to the historical baseline, marking the time interval as a high-intensity activation interval of the label and recording its start and end time and peak intensity. 3.The customer portrait matching positioning method of claim 1, wherein: Constructing the life cycle stage portrait comprises: for the identified life cycle event types, recording the type, start time, end time, stage intensity and related label set of the corresponding life cycle event for the customer as the granularity; when the same type of life cycle event occurs multiple times on the behavior time axis of the customer, connecting these life cycle events in chronological order to form a stage track.

4. The method of claim 1, wherein: Calculating the comprehensive analysis index based on the label time instability index and the stage behavior proportion index comprises: weighted fusion of the label time instability index and the stage behavior proportion index to obtain a comprehensive analysis index value; comparing the comprehensive analysis index value with a preset dynamic attenuation enabling threshold to determine whether to introduce time-related dynamic attenuation when calculating the scores of each portrait layer based on the comparison result. The method comprises the following steps: cleaning and time correcting multi-source behavior data under a unified time base to construct a behavior time sequence and a label time sequence; identifying a label high-intensity activation interval based on the label time sequence to form a multi-time scale portrait structure, the multi-time scale portrait structure comprising a long-term structural portrait, a life cycle stage portrait and an event portrait; wherein forming the multi-time scale portrait structure comprises: time axis alignment of the high-intensity activation interval of the customer in each label dimension to identify a stage event candidate segment; semantic classification of the stage event candidate segment in combination with an environmental event sequence and basic attributes of the customer to construct the life cycle stage portrait; elimination of the high-intensity segment located within a life cycle event influence window from the label time axis to construct the long-term structural portrait based on the label activation record of the remaining time period; extraction of a single event that does not form a complete life cycle stage but has significant business significance to construct the event portrait; analyzing time sensitivity and risk tolerance for a business task, and calculating a comprehensive analysis index based on a label time instability index and a stage behavior proportion index to determine whether to introduce time-related dynamic attenuation based on the comprehensive analysis index; calculating a long-term structural portrait score, a life cycle stage portrait score and an event portrait score, and fusing the scores of each portrait layer according to weighted summation to generate a comprehensive matching score; and outputting a target customer set based on the comprehensive matching score; combining feedback from recommendation, marketing or risk control to construct a feedback event sequence, and adjusting the multi-time scale portrait and the fusion weight according to the feedback event sequence. The process of identifying a label high-intensity activation interval based on the label time sequence comprises: adopting a sliding time window of a preset length to scan along the time axis of the label time sequence; comparing the label activation intensity within the window with the historical baseline intensity of the corresponding customer; when the label activation intensity in a certain continuous time interval is continuously and significantly lifted relative to the historical baseline, marking the time interval as a high-intensity activation interval of the label and recording its start and end time and peak intensity. Constructing the life cycle stage portrait comprises: for the identified life cycle event types, recording the type, start time, end time, stage intensity and related label set of the corresponding life cycle event for the customer as the granularity; when the same type of life cycle event occurs multiple times on the behavior time axis of the customer, connecting these life cycle events in chronological order to form a stage track. Calculating the comprehensive analysis index based on the label time instability index and the stage behavior proportion index comprises: weighted fusion of the label time instability index and the stage behavior proportion index to obtain a comprehensive analysis index value; comparing the comprehensive analysis index value with a preset dynamic attenuation enabling threshold to determine whether to introduce time-related dynamic attenuation when calculating the scores of each portrait layer based on the comparison result.

5. The customer profiling and matching positioning method of claim 4, wherein: The label time instability index is used to reflect the average level of the difference in activation intensity between the recent observation window and the long-term historical window of the label concerned in the task; and the phased behavior proportion index is used to measure the proportion of behaviors falling into the life cycle stage influence window or the short-term event influence window in the overall behaviors within the current task and the candidate customer range.

6. The method of claim 1, wherein: The scores of the respective portrait layers are calculated, including: calculating a long-term structural portrait score based on the similarity between the typical value of the label of the customer in a long-term range and the target portrait template; calculating a life cycle stage portrait score based on the affinity between the current life cycle stage of the customer and the business task, and introducing a timeliness weight when time-related dynamic attenuation is enabled; calculating an event portrait score based on the correlation between the single event recently occurred by the customer and the business task, and introducing a time attenuation weight when time-related dynamic attenuation is enabled.

7. The method of claim 1, wherein: The feedback event sequence includes customer feedback data from recommendation, marketing or risk control operations, which is aligned with the behavior time sequence after desensitization and formatting processing.

8. The method of claim 1, wherein: The feedback event sequence adjusts the fusion weight, including: analyzing the comprehensive performance of several similar business tasks; based on the consistency between the comprehensive matching score and the actual result, the fusion weight ratio of the long-term structural portrait, the life cycle stage portrait and the event portrait is slowly adjusted through iterative learning.

9. A client profiling and matching positioning system, characterized by, The positioning system is used to implement the method of any one of claims 1-8, comprising the following modules: a data acquisition and processing module for cleaning and time correction of multi-source behavior data under a unified time base, constructing a behavior time sequence and a label time sequence; a portrait modeling module for identifying a label high-intensity activation interval based on the label time sequence, forming a multi-time scale portrait structure, and the multi-time scale portrait structure including a long-term structural portrait, a life cycle stage portrait and an event portrait; a matching calculation module for analyzing time sensitivity and risk tolerance for a business task, and calculating a comprehensive analysis index based on a label time instability index and a phased behavior proportion index, and determining whether to introduce time-related dynamic attenuation based on the comprehensive analysis index; long-term structural portrait score, life cycle stage portrait score and event portrait score are calculated respectively, and the scores of the respective portrait layers are fused according to weighted summation to generate a comprehensive matching score; outputting a target customer set based on the comprehensive matching score; a feedback adjustment module for constructing a feedback event sequence in combination with recommendation, marketing or risk control feedback, and adjusting the multi-time scale portrait and the fusion weight according to the feedback event sequence.

Images