Milk powder new customer identification method based on time window and register fusion

Abstract

The application belongs to the technical field of data processing and customer identification, and more particularly relates to a milk powder new customer identification method based on a time window and storage fusion. Through standardized processing of multi-source data collection and business time extraction, system record time interference is excluded, and the time sequence dislocation problem of multi-source heterogeneous data is solved from the source. A unified consumer behavior timeline is constructed through event causal relationship inference, further calibrating the time sequence of behavior events and ensuring the authenticity of customer consumption tracks. Through equivalent modeling of storage and extraction behaviors, the purchase and consumption logic under the special consumption mode of the milk powder industry is accurately restored. Finally, through dynamic time window backtracking and weighted judgment, accurate identification of new customers is realized. The accuracy of milk powder new customer identification is significantly improved, avoiding misjudgment and missed judgment, ensuring the rationality of new customer identification results, and helping enterprises optimize customer hierarchical management, improve marketing resource utilization rate and customer satisfaction.

Description

Technical Field

[0001] This application belongs to the field of data processing and customer identification technology, and more specifically, relates to a method for identifying new milk powder customers based on time window and registration fusion. Background Technology

[0002] In the marketing and customer management process of the infant formula industry, new customer identification is a core prerequisite for formulating precise marketing strategies and optimizing customer service, and is of great significance for enhancing brand competitiveness. Currently, infant formula companies generally use multi-source heterogeneous systems to conduct business operations, including POS sales systems, storage systems, pickup systems, and online ordering systems, etc., with customer consumption behavior data scattered across these heterogeneous systems. However, existing new customer identification methods mostly analyze raw data from each system directly, without fully considering the characteristics of multi-source heterogeneous data, resulting in significant technical shortcomings. On the one hand, there is confusion in the time dimension of data records in various systems, often mistakenly taking the system record time as the business occurrence time, resulting in a time sequence misalignment of customer behavior events between different systems, making it impossible to accurately reconstruct the customer's true consumption trajectory; On the other hand, for the common deposit-pickup consumption model in the milk powder industry, existing methods do not perform equivalent modeling for deposit and pickup behaviors, only treating a single pickup as an independent purchase event, or omitting the purchase attributes corresponding to deposit, leading to biases in the judgment of new customers; for example, misjudging old customers who deposit and pick up multiple times as new customers, or failing to accurately identify first-time purchase customers corresponding to deposit behavior as new customers.

[0003] In summary, the two issues mentioned above directly result in the low accuracy of existing new customer identification methods, making it impossible to provide reliable data support for the precision marketing of infant formula companies. Summary of the Invention

[0004] This invention provides a new customer identification method for milk powder based on the fusion of time window and storage, which aims to solve the technical problem of low accuracy in existing new customer identification methods.

[0005] This invention provides a method for identifying new milk powder customers based on a fusion of time windows and registration, comprising the following steps: Step 1: Collect raw data from heterogeneous systems, eliminate interference from system recording times, extract the business occurrence time as the core event benchmark, standardize the heterogeneous data, and form a unified format dataset containing consumer identifiers, behavior types, and business occurrence times; Step 2: Infer the causal relationship between behavioral events based on business rules and lightweight statistical models. At the same time, combine the reliability of data sources and the validity of event types to assign a confidence weight to each event, generating an enhanced event dataset containing causal relationship labels and confidence weights. Step 3: Using the time of business occurrence as the initial sorting basis, and combining causal relationship markers and confidence weights, dynamically adjust the order of events through a priority queue mechanism to construct a unified behavior timeline for each consumer; Step 4: Equip the deposit behavior on the unified behavior timeline into a single purchase event, and the subsequent pickup behavior into multiple consumption events, and establish the association between each pickup event and the corresponding deposit event. Step 5: For suspected new purchase events on the timeline, backtrack a preset time window based on the time the suspected new purchase event occurred, filter related behaviors of the same consumer within the window, and make a weighted judgment based on the confidence level. If there are no records of the same brand that meet the judgment criteria within the window, the customer is marked as a new customer; otherwise, the customer is marked as an old customer.

[0006] This invention addresses the issue of temporal misalignment in heterogeneous multi-source data by standardizing the processing of multi-source data acquisition and business time extraction, eliminating interference from system recording times. It constructs a unified consumer behavior timeline through causal relationship inference, further calibrating the temporal sequence of behavioral events and ensuring the authenticity of customer consumption trajectories. Through equivalent modeling of deposit and pickup behaviors, it accurately reconstructs the purchasing and consumption logic under the unique consumption model of the infant formula industry. Finally, through dynamic time window backtracking and weighted judgment, it achieves accurate identification of new customers. This significantly improves the accuracy of new customer identification in the infant formula industry, avoids misjudgments and omissions, ensures the rationality of new customer identification results, and helps enterprises optimize customer segmentation management, improve marketing resource utilization, and increase customer satisfaction.

[0007] Preferably, when extracting the time of a business transaction, a time verification mechanism is introduced to identify abnormal records, including the following steps: Based on the characteristics of the business processing flow of each heterogeneous system, a reasonable time difference range between the business occurrence time and the system record timestamp is set in advance; the difference between the extracted business occurrence time and the corresponding data system record timestamp is calculated. If the difference of the timestamp exceeds the preset reasonable time difference range, the data is determined to be an abnormal record. For the identified abnormal records, the completeness of the preset key fields is further verified. If the key fields are complete, they are temporarily stored in the abnormal data pool. After subsequent manual review and confirmation, it will be determined whether to include them in the dataset.

[0008] Preferably, the heterogeneous system includes a POS sales system, a storage system, a pickup system, and an online ordering system; The fields of the original data are standardized through a field mapping table to form standard fields, which include system record timestamps.

[0009] Preferably, the inference of causal relationships between behavioral events includes the following steps: First, mandatory causal relationships between behavioral events are identified based on preset business rules. These mandatory causal relationships include those between orders and pickups, storage and pickups, payments and sales, and returns and sales. For behavioral events that are not covered by business rules and whose causal relationship cannot be determined by business rules, a lightweight statistical model is used to analyze the temporal correlation, data correlation and historical occurrence patterns of the events, thereby inferring potential causal relationships. After manual verification, the corresponding causal relationships are marked.

[0010] Preferably, the confidence weights include confidence based on the data source and confidence based on the event type; The confidence level based on the data source is determined by the standardization of data collection, data transmission failure rate and historical data accuracy of each heterogeneous system, and confidence levels are pre-classified and assigned corresponding values. The confidence level based on event type is pre-classified and assigned a corresponding value based on the traceability of business processes, the integrity of record fields, and the pass rate of historical authenticity verification for various behavioral events. The overall confidence level is calculated by multiplying the confidence levels of the data source and the event type for the same event.

[0011] Preferably, the priority queue mechanism is as follows: A confidence threshold is preset, and events with a comprehensive confidence level higher than the threshold are judged as high-confidence events. High-confidence events are given priority in the initial sorting based on the time of business occurrence. Events with a comprehensive confidence level lower than or equal to the confidence threshold are identified as low-confidence events. The causal relationship markers of low-confidence events are traversed, and the positions of the events in the timeline are adjusted according to the causal chain to ensure that all causal events are located before their corresponding result events.

[0012] Preferably, when the storage behavior is equivalent to a one-time purchase event, the business occurrence time of the storage event is used as the purchase date, the product information in the storage event is inherited, and the corresponding purchase amount is determined.

[0013] Preferably, when the pickup behavior is equivalent to a series of consumption events, the time when the pickup event occurs is taken as the consumption date. The corresponding deposited event is associated with the consumer identifier, product information and time range, and the consumption amount for this pickup is calculated.

[0014] Preferably, the preset time window is determined by combining the common consumption cycle of milk powder at different stages, the average repurchase cycle of milk powder consumers in the industry, and the statistical results of the brand's own customer consumption data.

[0015] Preferably, the weighted determination includes the following steps: First, filter all brand-related behavioral events of the same consumer within a preset time window, extract the comprehensive confidence weight of each event, and calculate the cumulative value. Based on the sample data for determining new and old customers in the brand's history, statistical analysis was conducted in combination with the consumption characteristics of different customer groups to calibrate and obtain the cumulative threshold that can distinguish between new and old customers. The calculated cumulative confidence weight is compared with the cumulative threshold. If the cumulative value does not exceed the cumulative threshold, the customer is identified as a new customer; if the cumulative value exceeds the preset cumulative threshold, the customer is identified as a returning customer.

[0016] The beneficial effects of the invention include: This invention addresses the issue of temporal misalignment in heterogeneous multi-source data by standardizing the processing of multi-source data acquisition and business time extraction, eliminating interference from system recording times. It constructs a unified consumer behavior timeline through causal relationship inference, further calibrating the temporal sequence of behavioral events and ensuring the authenticity of customer consumption trajectories. Through equivalent modeling of deposit and pickup behaviors, it accurately reconstructs the purchasing and consumption logic under the unique consumption model of the infant formula industry. Finally, through dynamic time window backtracking and weighted judgment, it achieves accurate identification of new customers. This significantly improves the accuracy of new customer identification in the infant formula industry, avoids misjudgments and omissions, ensures the rationality of new customer identification results, and helps enterprises optimize customer segmentation management, improve marketing resource utilization, and increase customer satisfaction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The overall process flowchart provided for embodiments of the present invention.

[0019] Figure 2 The flowchart for step 2 provided in this embodiment of the invention is shown. Detailed Implementation

[0020] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0021] See Figure 1 As shown, this embodiment provides a method for identifying new milk powder customers based on a fusion of time windows and registration, including the following steps: Step 1: Collect raw data from heterogeneous systems, eliminate interference from system recording times, extract the business occurrence time as the core event benchmark, standardize the heterogeneous data, and form a unified format dataset containing consumer identifiers, behavior types, and business occurrence times; The heterogeneous systems include a POS sales system, a storage system, a pickup system, and an online ordering system. In this embodiment, raw data is collected from the heterogeneous systems through interface calls or direct database connections. The raw data fields of each system are as follows: The original fields of the POS sales system include member card number, transaction time, product code, transaction amount, and system entry time; The original fields of the online order system include user ID, order time, payment status, product ID, and data reporting time; The original fields of the storage system include customer number, processing time, stored goods information, deposit amount, and record generation time; The original fields of the pickup system include the pickup person's ID, pickup time, pickup product details, corresponding storage slip number, and system recording time.

[0022] As mentioned above, the heterogeneous data of each system is not consistent. Therefore, in this embodiment, a field mapping table is pre-constructed to uniformly map the original fields of each system to standard fields. The mapping table is shown in Table 1 below: Table 1

[0023] In Table 1, the membership card number in the POS sales system serves as the consumer identifier, and the transaction time is the checkout time, i.e., the time when the business occurs; the user ID in the online order system serves as the consumer identifier, and the order placement time is the time when the business occurs; the customer number in the storage system serves as the consumer identifier, and the processing time is the time when the business occurs; and the pickup person number in the pickup system serves as the consumer identifier, and the pickup time is the time when the business occurs.

[0024] Based on the above field mapping table, the collected raw data is uniformly converted to ensure the consistency of data structure across different systems. Among them, the behavior type standard field is directly defined according to the data source system. Specifically: the behavior type corresponding to POS sales system data is sales, the behavior type corresponding to deposit system data is deposit, the behavior type corresponding to pickup system data is pickup, and the behavior type corresponding to online order system data is order.

[0025] In this embodiment, since the system recording time is susceptible to factors such as data acquisition delay and transmission failure, while the business occurrence time is the time when the consumer actually completes the behavior, which can truly reflect the temporal relationship of the behavior, this embodiment discards the system recording timestamp with low reliability and extracts the business occurrence time as the core benchmark for subsequent event sorting. Furthermore, in this embodiment, after extracting the time of the business transaction, a time verification mechanism is introduced to identify abnormal records to ensure the accuracy of the time data. The specific steps are as follows: Based on the characteristics of the business processing flow of each heterogeneous system, a reasonable range of difference between the business occurrence time and the system record timestamp is set through historical data statistical analysis. For example, the POS sales system has high data real-time requirements, so a reasonable time difference range is set between -10 minutes and 30 minutes. A negative value indicates that the business occurrence time is earlier than the system record time, which is a normal input delay. Similarly, a corresponding reasonable time difference range is set for the characteristics of other systems.

[0026] Then, for each record that has undergone field standardization, the difference between the business occurrence time and the system record timestamp is calculated. If the difference exceeds the reasonable time difference range preset by the corresponding system, the data is determined to be an abnormal record. For the identified abnormal records, the completeness of preset key fields is further verified. In this embodiment, the preset key fields include consumer identifier, behavior type, product information, and business occurrence time. During verification, if the key fields are complete, the abnormal record is temporarily stored in the abnormal data pool, and it will be determined whether to include it in the dataset after subsequent manual review and confirmation. If the key fields are incomplete, the record is directly removed to avoid interfering with the subsequent analysis results.

[0027] Step 2: Infer the causal relationship between behavioral events based on business rules and lightweight statistical models. At the same time, combine the reliability of data sources and the validity of event types to assign a confidence weight to each event, generating an enhanced event dataset containing causal relationship labels and confidence weights. In this embodiment, see Figure 2 As shown, the business rules include order-pickup rules, storage-pickup rules, payment-sales rules, and return-sales rules; Order-Pickup Rules: If the same consumer has both an order and a pickup, and the product information (product ID, brand, model) of the two items is completely matched, and the pickup occurs within a reasonable period after the order (in this example, it is set to 90 days, corresponding to the regular shelf life and consumption cycle of milk powder), then the order event is inferred to be the cause and the pickup event to be the effect.

[0028] Storage-Pickup Rules: If the same consumer has both storage and pickup behaviors, and the product information of the two behaviors matches, and the pickup behavior occurs within a reasonable period after the storage behavior (365 days in this embodiment, corresponding to the regular validity period of the storage service), then it is inferred that the storage event is the cause and the pickup event is the effect.

[0029] Payment-Sales Rule: If the same consumer has both a payment activity (payment record in the online order system, corresponding to the behavior type of payment) and a sales activity, and the payment amount is consistent with the sales amount and the product information matches, and the payment time is within 30 minutes before the sales time, then it is inferred that the payment event is the cause and the sales event is the effect.

[0030] Returns-Sales Rules: If the same consumer returns an item (corresponding to the behavior type of "return"), and the returned item information matches the item information of a certain sales behavior, and the return time is within 30 days after the sales time, then the causal relationship of the sales event is weakened (the confidence level needs to be adjusted accordingly in subsequent calculations); if the quantity of returned items is the same as the quantity sold, then the sales event is considered invalid.

[0031] In this embodiment, the standardized dataset generated in step 1 is grouped by consumer identifier, the events in each group are traversed, causal relationship pairs are matched according to the above rules, and the preceding time ID (i.e., the unique identifier of the cause event) is marked for the result event, thus completing the forced causal relationship marking.

[0032] For behavioral events not covered by the aforementioned business rules (such as emerging online pre-order pickup behavior, and split pickup behavior after purchasing bundled packages), the Apriori algorithm is used to analyze the temporal correlation, data correlation, and historical occurrence patterns of the events to infer potential causal relationships, as detailed below: We selected a standardized dataset from the past year, sorted it by consumer identifier and the time of the business transaction, and constructed a behavior sequence for each consumer.

[0033] The minimum support is set to 0.05, which means that the occurrence rate of a certain event combination in all behavioral sequences is not less than 5%, and the minimum confidence is set to 0.8, which means that if the causal event occurs, the probability of the result event occurring is not less than 80%.

[0034] The Apriori algorithm is then used to mine frequently occurring event combinations in the behavioral sequence. For example, the combination of online reservation and in-store pickup occurs frequently, and frequent itemsets are formed accordingly.

[0035] Calculate the confidence scores of event A (preceding event) and event B (following event) in the frequent itemset. The confidence score calculation formula is Confidence(A→B) = Support(A∪B) / Support(A), where the support score is the proportion of the event combination in all sequences. If the confidence score is higher than the set minimum confidence score, and the business occurrence time of event A is within a reasonable window before event B (set to 7 days in this embodiment), then it is preliminarily inferred that event A is a potential cause of event B.

[0036] The preliminary inference of potential causal relationships is submitted to domain experts for manual evaluation to verify whether it conforms to the actual business logic. For example: if the experts verify that online reservation is indeed a prerequisite for in-store pickup, then the causal relationship is confirmed and the preceding event ID is marked for the result event; if it is verified to be an accidental co-occurrence (such as the co-occurrence of browsing products and picking up goods), then the potential causal relationship is eliminated.

[0037] Furthermore, in this embodiment, a comprehensive confidence metric is constructed to quantify the reliability of each event, wherein the comprehensive confidence metric consists of a confidence metric based on the data source and a confidence metric based on the event type, as detailed below: Based on the data acquisition standardization, data transmission failure rate, and historical accuracy of each heterogeneous system, confidence levels based on data sources are pre-defined and assigned corresponding values. For example, the confidence level classification based on data sources is shown in Table 2 below: Table 2

[0038] When assigning confidence scores based on data sources, the confidence score value in Table 2 is matched with the data source of each event and used as the data source confidence score for the corresponding time.

[0039] Based on the traceability of business processes, the completeness of record fields, and the pass rate of historical authenticity verification for various behavioral events, confidence levels based on event types are pre-defined and assigned corresponding values; for example, the confidence level classification based on event type is shown in Table 3 below: Table 3

[0040] When assigning confidence scores based on event type, the corresponding confidence score value is matched according to the behavior type of each event, as shown in Table 3, and used as the event type confidence score for that event. For sales events with return associations, the event type confidence score is lowered. For example, when the return quantity is the same as the sales quantity, the confidence score is adjusted to 0, and when the return quantity is 50% of the sales quantity, the confidence score is adjusted to 50% of the original value.

[0041] The overall confidence metric is the product of confidence based on the data source and confidence based on the event type. In this embodiment, the overall confidence metric reflects the reliability of both the data source and the event type. Low reliability in any dimension will lead to a decrease in overall reliability.

[0042] Based on this, the obtained causal relationship labels and overall confidence scores are added to the standardized dataset to form an enhanced event dataset. Each record in the enhanced event dataset contains the following fields: consumer identifier, behavior type, business occurrence time, product information, amount information, causal relationship label, and overall confidence score.

[0043] Step 3: Using the time of business occurrence as the initial sorting basis, and combining causal relationship markers and confidence weights, dynamically adjust the order of events through a priority queue mechanism to construct a unified behavior timeline for each consumer; In this embodiment, the enhanced event dataset generated in step 2 is used as input. The data is grouped by consumer identifier, with each group corresponding to all behavioral events of a single consumer. For each consumer's events, they are sorted in ascending order based on the time of the business transaction to construct the consumer's initial behavioral timeline. For example, the initial event sequence of a certain consumer is: 2025-10-01 09:00:00 (order, business transaction time), 2025-10-01 09:30:00 (deposit), 2025-10-02 10:00:00 (pickup), which are sorted by the time of the business transaction to form the initial timeline.

[0044] In this embodiment, the initial timeline is sorted only based on the time dimension, without considering the causal relationships and confidence differences between events, which may lead to time misalignment, such as the business occurrence time of an online order being mistakenly recorded as later than the pickup time. For this initial timeline that iterates through each consumer, all causal event pairs are identified based on the causal relationship labels in the enhanced dataset, ensuring that the cause event precedes the result event, as detailed below: For each consumer's initial timeline, scan the causal relationship marker (event ID) of each event, locate the corresponding cause event, and construct a causal chain. For example, if event A (order) is the cause of event B (pickup), and event C (payment) is the cause of event A (order), then the causal chain C→A→B is formed.

[0045] Check the order of events in the causal chain. If the result event is found to be before the cause event (e.g., B is before A in the initial timeline), then the cause event is forcibly moved to the position before the result event, and the minimum time interval between the cause event and the result event is retained (i.e., the business occurrence time of the cause event is still the original extraction time, only the sorting order is adjusted, and the time value is not modified).

[0046] After adjusting a single event, since causal relationships may form multi-level chains, it is necessary to recursively check whether its adjacent events are affected. For example, after adjusting the position of the causal event A, it is necessary to check whether the causal event C of A is located before A; if C is moved after A, then the position of C is adjusted to ensure that the entire causal chain is correctly ordered. In this embodiment, the maximum recursion depth is set to 10 levels to avoid infinite loops caused by abnormal causal relationships. Based on this, the timeline is ensured to conform to business logic, but order conflicts may be introduced (such as high-confidence events being moved). Therefore, fine-tuning is required in combination with the overall confidence level, as follows: A priority queue, implemented using a min-heap, is created for each consumer. Each element in the queue represents an event, and its priority is determined by a weighted sorting score. Events from each consumer are inserted into the priority queue sequentially according to the causal-adjusted timeline. During insertion, if a new event conflicts with an existing event in the queue (e.g., the new event occurred earlier than an existing event but has a higher weighted score), their weighted sorting scores are compared, and the event with the lower score is prioritized. The formula for calculating the weighted sorting score is as follows: ; ; In the formula: Indicates the weighted sorted score; Indicates the time when the standardized business occurred; Indicates the overall confidence level of the event; Indicates the weighting factor; Indicates the business timestamp of the event; This indicates the earliest timestamp of any business transaction among all consumer events. This represents the timestamp of the latest business transaction among all consumer events.

[0047] In this embodiment, events are extracted sequentially from the priority queue in ascending order of weighted sorting scores to form a unified behavior timeline for each consumer. This ensures the correctness of causal relationships while balancing the impact of different quality data through confidence weights, thus truly reflecting the consumer's behavior sequence. The unified behavior timeline for each consumer is output in list form, and each record contains the following fields: consumer identifier, behavior type, business occurrence time, product information, amount information, causal relationship marker, and overall confidence score.

[0048] Step 4: Equip the deposit behavior on the unified behavior timeline into a single purchase event, and the subsequent pickup behavior into multiple consumption events, and establish the association between each pickup event and the corresponding deposit event. Specifically, in this embodiment, the unified behavior timeline of each consumer is traversed to identify events of the behavior type as "deposit," which are then equated to purchase (deposit) events. The specific steps are as follows: The consignment behavior type is changed to purchase (consignment), clarifying its equivalent purchase attribute and distinguishing it from the purchase attribute of regular sales behavior; then, the business occurrence time of the consignment event is used as the purchase date of the purchase (consignment) event, which is the time when the consumer actually completes the consignment (i.e., advance purchase), accurately reflecting the timing of the first purchase; finally, the product information in the consignment event is completely inherited into the purchase (consignment) event to ensure the consistency of product information, which includes brand, model, quantity, product ID, etc.

[0049] If the storage event includes prepayment or deposit information, that amount will be used directly as the purchase amount for the (storage) event; if the storage event does not have amount information (such as storage services offered by the brand), the purchase amount will be obtained based on the market price of the stored goods.

[0050] Based on this, an equivalent purchase event record is generated, including consumer identifier, behavior type (purchase (deposit)), purchase date (time when the deposit business occurred), product information, purchase amount, original deposit event ID (for tracing), and overall confidence level (inherited from the overall confidence level of the original deposit event); this record replaces the original deposit event and is updated to the unified behavior timeline.

[0051] The act of picking up goods is equivalent to a multi-stage consumption event, which includes the following steps: Traverse the unified behavior timeline of each consumer, identify events with the behavior type of pickup, change the pickup behavior type to consumption (pickup), clarify that it is a phased consumption attribute of previous deposit purchase, and distinguish it from direct purchase sales behavior; The transaction time of the pickup event is used as the consumption date of the consumption (pickup) event to accurately record the consumption sequence. Then, the product information (brand, model, quantity, product ID, etc.) from the pickup event is fully inherited into the consumption (pickup) event to ensure consistency with the corresponding stored product information. Next, the stored events related to this event are found, which can be linked through consumer identifiers and product information. After successful association, the ID of the stored event is recorded as a reference. Based on this, an event record is generated, including consumer identifier, behavior type (consumption (pickup)), consumption date (pickup date), product information, associated event ID (corresponding stored event ID), consumption amount, etc. The consumption amount is calculated in the following way: ; In the formula: This indicates the amount spent on this purchase. This indicates the amount generated by the deposit. This indicates the total number of goods stored. This indicates the quantity of goods picked up in this delivery.

[0052] Based on the above equivalent modeling, the data structure of the unified behavior timeline is adjusted, and new fields are added, including associated event ID (used to record the relationship between consumption (pickup) and purchase (deposit)), event source (records the system to which the original data belongs, such as the deposit system and the pickup system, to facilitate traceability and debugging), and event subtype (subdivided on the basis of behavior type, such as purchase (deposit) and consumption (pickup), to enrich the behavior description). Based on this, the enhanced unified behavior timeline is output.

[0053] Step 5: For suspected new purchase events on the timeline, backtrack a preset time window based on the time the suspected new purchase event occurred, filter related behaviors of the same consumer within the window, and make a weighted judgment based on the confidence level. If there are no records of the same brand that meet the judgment criteria within the window, the customer is marked as a new customer; otherwise, the customer is marked as an old customer.

[0054] In this embodiment, the enhanced unified behavior timeline of each consumer is traversed to identify suspected new purchase events, based on the following criteria: The behavior type is sales or purchase (deposit), and the business occurs within a preset valid time range (e.g., 2 years). For each suspected new purchase event identified above, a dynamic time window is set to backtrack historical behavior. Relevant historical behaviors are filtered from the consumer's enhanced unified behavior timeline, with the following filtering criteria: Consumer identification consistency: Consumers with the same unique consumer identifier as the current event; Time range constraint: The business transaction must occur within the backtracking window. Scope of behavioral types: Includes all equivalent purchasing and consumption behaviors such as sales, consumption (pickup), orders, and purchases (deposit); The backtracking window is set to 365 days and can be adjusted according to the common usage cycle of milk powder at different stages. For example, if the consumption cycle of infant formula is 1 to 3 months, the backtracking window can be set to 3 months. It can also be set according to the average repurchase cycle of milk powder consumers in the industry.

[0055] Based on the above filtering, we obtain the historical relevant behavior dataset of the corresponding consumers within the retrospective window, which is used for subsequent determination of records from the same brand. The specific determination is as follows: First, extract the milk powder brand from the product information of the suspected new purchase event. Then, accurately match the brand information with the product information of the historical behavior dataset filtered in the backtracking window to filter out all historical behavior events of the same brand. Then, calculate the weighted brand existence value, which is the sum of the comprehensive confidence of all historical events of the same brand in the backtracking window.

[0056] If the weighted brand value of the suspected purchase event is less than or equal to the set threshold, it is determined to be a new customer; if it is greater than the set threshold, it is determined to be a returning customer. The threshold is obtained based on the following steps: We selected historical customer data from the past two years to determine whether a customer was a new or returning customer. The sample included known behavioral records of both new and returning customers. For each suspected purchase event, we calculated the weighted brand presence value within the backtracking window and statistically analyzed the distribution of the weighted brand presence values ​​for new and returning customers. For example, if the statistics show that the weighted brand presence value for new customers is less than 0.7 and the weighted brand presence value for returning customers is greater than 0.7, then 0.7 is the dividing line for differentiation, and can be used as the threshold.

[0057] This invention solves the problem of missing new customers caused by treating deposit and pickup as independent behaviors in traditional methods by using equivalent modeling of deposit and pickup. Furthermore, based on the timeline adjustment of causal relationships and confidence weights, it corrects the temporal disorder caused by the misalignment of system recording time in traditional methods, making the screening of historical behaviors within the backtracking window more accurate. In addition, the weighted judgment matrix reduces the impact of low-quality data, which can more accurately distinguish between new and old customers and reduce misjudgments.

[0058] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for identifying new milk powder customers based on the fusion of time window and storage, characterized in that, Includes the following steps: Step 1: Collect raw data from heterogeneous systems, eliminate interference from system recording times, extract the business occurrence time as the core event benchmark, standardize the heterogeneous data, and form a unified format dataset containing consumer identifiers, behavior types, and business occurrence times; Step 2: Infer the causal relationship between behavioral events based on business rules and lightweight statistical models. At the same time, combine the reliability of data sources and the validity of event types to assign a confidence weight to each event, generating an enhanced event dataset containing causal relationship labels and confidence weights. The inference of causal relationships between behavioral events includes the following steps: First, mandatory causal relationships between behavioral events are identified based on preset business rules. These mandatory causal relationships include those between orders and pickups, storage and pickups, payments and sales, and returns and sales. For behavioral events that are not covered by business rules and whose causal relationship cannot be determined by business rules, a lightweight statistical model is used to analyze the temporal correlation, data correlation and historical occurrence pattern of the event, thereby inferring potential causal relationships. After manual judgment and verification, the corresponding causal relationship is marked. The confidence weights include confidence based on the data source and confidence based on the event type; The confidence level based on the data source is determined by the standardization of data collection, data transmission failure rate and historical data accuracy of each heterogeneous system, and confidence levels are pre-classified and assigned corresponding values. The confidence level based on event type is pre-classified and assigned a corresponding value based on the traceability of business processes, the integrity of record fields, and the pass rate of historical authenticity verification for various behavioral events. The overall confidence level is calculated by multiplying the confidence levels of the data source and the event type for the same event. Step 3: Using the time of business occurrence as the initial sorting basis, and combining causal relationship markers and confidence weights, dynamically adjust the order of events through a priority queue mechanism to construct a unified behavior timeline for each consumer; Step 4: Equip the deposit behavior on the unified behavior timeline into a single purchase event, and the subsequent pickup behavior into multiple consumption events, and establish the association between each pickup event and the corresponding deposit event. Step 5: For suspected new purchase events on the timeline, backtrack a preset time window based on the time the suspected new purchase event occurred, filter related behaviors of the same consumer within the window, and make a weighted judgment based on the confidence level. If there are no records of the same brand that meet the judgment criteria within the window, the customer is marked as a new customer; otherwise, the customer is marked as an old customer.

2. The method for identifying new milk powder customers based on the fusion of time window and registration as described in claim 1, characterized in that, When extracting the time of a business transaction, a time verification mechanism is introduced to identify abnormal records, including the following steps: Based on the characteristics of the business processing flow of each heterogeneous system, a reasonable time difference range between the business occurrence time and the system record timestamp is set in advance; the difference between the extracted business occurrence time and the corresponding data system record timestamp is calculated. If the difference of the timestamp exceeds the preset reasonable time difference range, the corresponding data is determined to be an abnormal record. For the identified abnormal records, the completeness of the preset key fields is further verified. If the key fields are complete, they are temporarily stored in the abnormal data pool. After subsequent manual review and confirmation, it will be determined whether to include them in the dataset.

3. The method for identifying new milk powder customers based on the fusion of time window and registration as described in claim 1, characterized in that, The heterogeneous system includes a POS sales system, a storage system, a pickup system, and an online ordering system; The fields of the original data are standardized through a field mapping table to form standard fields, which include system record timestamps.

4. The method for identifying new milk powder customers based on the fusion of time window and registration as described in claim 1, characterized in that, The priority queue mechanism is as follows: A confidence threshold is preset, and events with a comprehensive confidence level higher than the threshold are judged as high-confidence events. High-confidence events are given priority in the initial sorting based on the time of business occurrence. Events with a comprehensive confidence level lower than or equal to the confidence threshold are identified as low-confidence events. The causal relationship markers of low-confidence events are traversed, and the positions of the events in the timeline are adjusted according to the causal chain to ensure that all causal events are located before their corresponding result events.

5. The method for identifying new milk powder customers based on time window and registration fusion according to claim 1, characterized in that, When the storage behavior is equated to a one-time purchase event, the business occurrence time of the storage event is used as the purchase date, the product information in the storage event is inherited, and the corresponding purchase amount is determined.

6. The method for identifying new milk powder customers based on the fusion of time window and registration as described in claim 1, characterized in that, When the pickup behavior is equated to a split consumption event, the time when the pickup event occurs is taken as the consumption date. The corresponding stored event is associated with the consumer identifier, product information and time range, and the consumption amount for this pickup is calculated.

7. The method for identifying new milk powder customers based on the fusion of time window and registration as described in claim 1, characterized in that, The preset time window is determined by combining the common consumption cycle of milk powder at different stages, the average repurchase cycle of milk powder consumers in the industry, and the brand's own customer consumption data statistics.

8. The method for identifying new milk powder customers based on the fusion of time window and registration as described in claim 1, characterized in that, The weighted determination includes the following steps: First, filter all brand-related behavioral events of the same consumer within a preset time window, extract the comprehensive confidence weight of each event, and calculate the cumulative value. Based on the sample data for determining new and old customers in the brand's history, statistical analysis was conducted in combination with the consumption characteristics of different customer groups to calibrate and obtain the cumulative threshold that can distinguish between new and old customers. The calculated cumulative confidence weight is compared with the cumulative threshold. If the cumulative value does not exceed the cumulative threshold, the customer is identified as a new customer; if the cumulative value exceeds the preset cumulative threshold, the customer is identified as a returning customer.

Images