Intelligent customer information matching method based on big data

Abstract

The invention belongs to the technical field of big data customer information matching, and discloses an intelligent customer information matching method based on big data, which comprises the following steps of: firstly, establishing a scene customer behavior semantic dictionary according to three groups of new customers, high-value customers and loss early-warning customers, and determining behavior label mapping rules of different customer types; for example, the browsing duration of a new customer exceeds 3 minutes, the consultation frequency exceeds 2, the behavior is defined as a high-intention behavior, the semantic dictionary is dynamically updated every 24 hours based on cross-platform semantic conflict feedback, and the cross-platform data semantic standard is unified; secondly, screening weak association features by adopting a mutual information and customer value double-weight algorithm, setting weights according to customer value differences, and preferentially retaining features with higher association degree with customer demands to generate structured feature vectors; and finally, deploying a preprocessing effect verification module at an edge computing node, performing field secondary labeling, null interpolation complementation and time format unification operation on the data, and improving the data standardization level.

Description

Technical Field

[0001] This invention belongs to the field of big data customer information matching technology, specifically a smart customer information matching method based on big data. Background Technology

[0002] In fields such as finance, e-commerce, and customer service, customer information matching is a crucial link supporting precise service and business decision-making. Customer data is generally aggregated from multiple channels, including online platforms and offline stores. However, customer information matching faces complex challenges due to data diversification, dynamic demands, and cross-domain scenarios. Current technical solutions still have the following technical problems: Different platforms often have semantic discrepancies in the definition and coding of the same customer behavior. For example, the "add to cart" operation on e-commerce platforms and the "favorite" function on social platforms both reflect potential customer needs, but their field identifiers and semantic tags are completely different. Current technology can only complete basic data format conversion and has not established a unified semantic mapping system, resulting in a large number of cross-platform behavioral features being unable to be effectively integrated due to semantic conflicts. The value mining of weakly correlated data such as customer browsing time and consultation tone is almost non-existent, relying only on strongly correlated fields such as order information and user ID, resulting in a waste of a large amount of data value.

[0003] Customer needs are inherently sudden and phased. For example, ordinary customers may suddenly browse baby and maternity products frequently, or during holidays, customer needs may shift from essential purchases to gift selection. Traditional matching models rely on historical static data for training and lack a mechanism to quickly capture short-term dynamic behaviors. Once customer needs change abruptly, the matching accuracy will drop significantly. Furthermore, model iteration takes a long time and cannot adapt to real-time changing demand scenarios.

[0004] Offline store membership data and online platform data belong to different data domains. Current technology has not established an effective cross-domain feature mapping relationship. For example, there is a lack of unified feature dimension definition between offline membership level and online account activity, which makes it impossible for cross-domain data to participate in customer information matching. Cross-domain data also suffers from feature distribution offset problems. For example, the average order value offline is generally higher than online, which further reduces the matching accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent customer information matching method based on big data to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent customer information matching method based on big data, the specific steps of which are as follows: The specific steps of S1 are as follows: First, a scenario-based customer behavior semantic dictionary is established based on new customers, high-value customers, and customers at risk of churn. Mapping rules for different behavioral tags are defined according to customer type. For new customers, browsing time of more than 3 minutes and more than 2 inquiries are mapped as high-intent behavior, and page scrolling depth of more than 80% is mapped as in-depth understanding behavior. For high-value customers, repeat purchase interval of less than 30 days, average order value more than twice the category average, and dedicated customer service consultation are mapped as core value behavior. For customers at risk of churn, a 50% decrease in login frequency in nearly 30 days and items left in shopping cart for more than 7 days are mapped as reactivation demand behavior. The scenario-based customer behavior semantic dictionary is dynamically updated every 24 hours based on cross-platform semantic conflict feedback. Simultaneously, a dual-weight algorithm of mutual information and customer value is used to screen weakly correlated features, including consultation tone and page dwell time. Different weights are set according to customer value, and features with a correlation greater than 0.6 with customer needs are selected to generate structured feature vectors. Finally, a preprocessing effect verification module is deployed on the edge computing node. If the basic accuracy of matching after normalization of data on a certain platform is less than 85%, secondary field annotation is automatically triggered, and interpolation is performed to complete fields with a null value ratio greater than 10%, and the time format is unified.

[0007] Furthermore, step S2 is specifically as follows: Based on the structured feature vectors and cleaned data from step S1, the granularity of sensitive field desensitization is first designed according to customer value stratification. The desensitization process adopts... A differential privacy mechanism with a value of 0.3 ensures privacy and security; Subsequently, a federated adversarial learning framework was established, deploying feature alignment sub-models at each data source node. The generator simulates online features based on offline store visit frequency and paper membership levels, while the discriminator optimizes parameters through an adversarial loss function to reduce cross-domain feature differences. The federated center sets up a cross-domain alignment effect feedback module. If the KL divergence between offline store consumption amount and online APP consumption amount is greater than 0.15, the generator loss function is adjusted and customer activity weights are added. If the mismatch rate after alignment at a certain node is greater than 20%, parameter transmission is paused and feature dimensions are verified. At the same time, SHA-256 hash mapping is performed on offline unstructured sensitive data to convert it into irreversible feature codes containing only category and weight.

[0008] Furthermore, step S3 is specifically as follows: Based on the cross-domain feature data aligned in step S2, a multimodal triggering mechanism is first established by integrating three types of demand signals: text, behavior, and device. Text signals are triggered by the popularity of keywords in customer service conversations, behavior signals are triggered by the density of behavior within one hour, and device signals are triggered by the frequency of APP opening within one hour. When two or more types of signals are satisfied, a demand mutation feature vector is generated. Subsequently, a demand intensity calculation model is established, and priority is quantified by demand intensity. Demand intensity ≥ 0.8 is pushed to the inventory for priority matching, 0.5 ≤ demand intensity < 0.8 is used to execute personalized recommendations, and less than 0.5 is included in the long-term cultivation pool. Simultaneously, customer behavior dictionary filtering is performed on unstructured data, including customer service dialogues and product reviews. A dedicated dictionary is built based on customers' historical behavior, and after filtering out redundant keywords, it is input into the BERT-base model to generate a 768-dimensional semantic vector.

[0009] Furthermore, step S4 is specifically as follows: First, a dual time-series matrix is ​​established with a short-term series of nearly 72 hours and a long-term series of 90 days as the unit of time. The short-term series is weighted according to the behavioral value, and the long-term series is statistically analyzed to form a baseline based on the daily behavior frequency and category preference ratio. The influence of short-term high-value behaviors is strengthened by weighted fusion. Then, a long-term behavioral trend database is established to store the stability of category preferences and the consistency of behavioral patterns over 90 days. If short-term behavior conflicts with long-term trends, the related data is checked. When the correlation is less than 0.3, the short-term weight is reduced to 0.5 times the original weight; when the correlation is 0.3≤R≤0.5, the short-term weight is retained at 0.8 times the original weight; when the correlation is greater than 0.5, the short-term weight is retained. At the same time, time-series feature attention pooling is used to compress high-frequency repetitive behaviors, retaining the first or last time and frequency, and merging consecutive identical behaviors.

[0010] Furthermore, step S5 is specifically as follows: Based on the time-series feature vector fused in step S4, a dynamic weight matrix with two dimensions of customer value and business scenario is first established based on the DQN algorithm. The core dimensions include five categories: identity, needs, behavior, preferences, and scenario, and different weights are assigned according to scenario and customer value. Deploy a real-time performance monitoring unit to calculate the matching conversion rate and customer satisfaction every hour. If the conversion rate is 10% lower than the target, the weight of the dimension that did not meet expectations will be increased. If the satisfaction rate is less than 80%, the weight of the dimension that affects the experience will be decreased. The adjustment range is ≤0.2 to avoid fluctuations. For scenarios including holidays, member days, and inventory clearance, add peripheral dimensions to participate in the matching calculation together with the core dimensions to form a scenario-based matching matrix.

[0011] Furthermore, step S6 is specifically as follows: Based on the scenario-based matching matrix in step S5, an abnormal high-value behavior identification and verification process is first established. For behaviors including single consumption exceeding 5 times the historical average and high-frequency inquiries about high-value products within 1 hour, historical association verification and scenario association verification are performed. If both verifications are passed, a fast channel is triggered, skipping the regular feature verification and entering the high-priority queue. Based on a short-term behavior index pool of nearly 72 hours, a scenario hotspot preloading module is deployed. 24 hours before the promotion, historical promotional categories and coupon records are preloaded to generate pre-computed vectors. 12 hours before the new product launch, historical new product trajectories and purchase records of the same category are preloaded to establish a pre-matching candidate set. During the event, the preloaded features are directly called to shorten the response time.

[0012] Incremental computing is scheduled using a dual-dimensional approach that considers both customer value and behavioral urgency. 40% of computing resources are allocated to high-value customers who make purchases or inquire about after-sales service, 35% to ordinary customers who add items to their cart or inquire about new products, 15% to customers with churn warnings who log in or click on promotions, and 10% to low-value customers who simply browse. Feature extraction and matching updates are performed only on newly added data, and old results are replaced by incremental indexes to optimize the allocation of computing resources.

[0013] Furthermore, step S7 is specifically as follows: First, cross-platform semantic conflict feedback is divided into three levels: Level 1 is a fatal conflict, in which the customer behavior semantic dictionary is updated and edge nodes are synchronized within 2 hours; Level 2 is a significant conflict, in which the tag mapping rules are corrected within 12 hours; and Level 3 is a minor conflict, in which the descriptions are updated in batches within 24 hours. Then, a dual-indicator evaluation system of matching accuracy and response time is established. Matching accuracy is monitored by accuracy, recall, and mismatch rate, while response time is monitored by regular matching, fast channel, and peak period. If the accuracy improves but the response time increases, the number of BERT layers is reduced and the number of attention heads of Attention-LSTM is simplified to achieve lightweighting. Meanwhile, feedback is collected from unsatisfied customers through lightweight questionnaires. Key needs are extracted from the text feedback using NLP, and the feedback data is integrated into the model iteration. If the product category does not match, the BERT semantic weights are adjusted; if the price does not match, the price sensitivity parameter in the dynamic matrix is ​​optimized.

[0014] The beneficial effects of this invention are as follows: 1. This invention first establishes a scenario-based customer behavior semantic dictionary for three customer groups: new customers, high-value customers, and customers at risk of churn. It clarifies the behavioral tag mapping rules for different customer types. For example, new customers are defined as having a browsing time exceeding 3 minutes or more than 2 inquiries as exhibiting high-intent behavior. The semantic dictionary is dynamically updated every 24 hours based on cross-platform semantic conflict feedback, unifying cross-platform data semantic standards. Second, it employs a dual-weight algorithm of mutual information and customer value to filter weakly correlated features, setting weights according to differences in customer value and prioritizing the retention of features more relevant to customer needs to generate structured feature vectors. Finally, a preprocessing effect verification module is deployed on edge computing nodes to perform secondary field annotation, null value interpolation completion, and time format standardization operations on the data, improving data standardization. This effectively eliminates semantic barriers between multi-source data, fully releasing the potential value of weakly correlated data and providing high-quality data support for accurate matching of subsequent customer information.

[0015] 2. This invention employs a tiered customer value design to desensitize sensitive fields. For example, high-value customers retain their phone numbers with tiered desensitization labels, while ordinary customers only undergo basic desensitization. A differential privacy mechanism is used to ensure data privacy and security. A federated adversarial learning framework is established, deploying feature alignment sub-models at each data source node. A generator simulates online features based on offline store visit frequency and paper membership levels, while a discriminator optimizes parameters using an adversarial loss function to reduce cross-domain feature differences. A cross-domain alignment effect feedback module is set up in the federated center, dynamically adjusting model parameters based on the KL divergence of offline and online consumption amounts and node alignment mismatch rate to ensure cross-domain feature synergy. SHA-256 hash mapping is performed on unstructured sensitive offline data, converting it into irreversible feature codes containing only category and weight. This solves the problem of cross-domain data not being able to participate in matching collaboratively, avoids the risk of sensitive data leakage, and significantly improves the impact of cross-domain feature distribution offset on matching accuracy.

[0016] 3. This invention establishes a multimodal triggering mechanism by integrating three types of demand signals: text, behavior, and device. Text signals are triggered by the popularity of keywords in customer service dialogues, behavior signals by the density of behavior within one hour, and device signals by the frequency of APP opening within one hour. When two or more types of signals are satisfied, a demand mutation feature vector is generated. Then, the priority is quantified by a demand intensity calculation model to achieve dynamic demand response. A dual time-series matrix of 72 hours (short-term) and 90 days (long-term) is established. The short-term series is weighted according to the value of behavior, and the long-term series is statistically analyzed to form a baseline based on the daily frequency of behavior and the proportion of category preference. The conflict between short-term behavior and long-term trend is corrected using a long-term behavior trend library to strengthen the impact of short-term high-value behavior and avoid the model relying on a single static data. A differentiated resource scheduling mechanism is constructed to perform historical correlation verification and scenario correlation verification on abnormal high-value behavior. After passing the verification, a fast track is opened. At the same time, computing resources are allocated according to customer value and behavior urgency, and historical data is preloaded to generate pre-calculated vectors before scenarios such as promotions and new product launches. This can quickly respond to dynamic customer needs, avoid waste of computing resources, and improve matching efficiency and high-value customer service experience. Attached Figure Description

[0017] Figure 1 This is a flowchart of the intelligent customer information matching method based on big data according to the present invention; Figure 2 This is a flowchart illustrating the data preprocessing and standardization process of this invention. Figure 3 This is a flowchart of the cross-domain data security processing of the present invention; Figure 4 This is a flowchart of the requirements capture and timing fusion process of this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figures 1 to 4 As shown in the figure, this embodiment of the invention provides an intelligent customer information matching method based on big data. The specific steps of the method are as follows: The specific steps of S1 are as follows: To address the issues of semantic heterogeneity and insufficient value mining of weak associations in multi-source data, a scenario-based customer behavior semantic dictionary is first established based on new customers, high-value customers, and customers at risk of churn. Mapping rules for different behavioral tags are defined according to customer type. For new customers, browsing time greater than 3 minutes and more than 2 inquiries are mapped as high-intent behavior, and page scrolling depth greater than 80% is mapped as in-depth understanding behavior. For high-value customers, repeat purchase interval less than 30 days, average order value more than twice the category average, and dedicated customer service consultation are mapped as core value behaviors. For customers at risk of churn, a 50% decrease in login frequency in nearly 30 days and items remaining in the shopping cart for more than 7 days are mapped as reactivation demand behavior. Furthermore, the scenario-based customer behavior semantic dictionary is dynamically updated every 24 hours based on semantic conflict feedback across platforms, such as the difference in conversion efficiency between adding items to cart on e-commerce platforms and social collections, and the semantic deviation between canceling payments on payment platforms and canceling orders through customer service. Simultaneously, a dual-weight algorithm combining mutual information and customer value is used to filter weakly correlated features such as consultation tone and page dwell time. Different weights are set according to customer value; for example, the urgency weight of consultation tone is increased by 50% for high-value customers, and the completeness weight of browsing path is increased by 30% for new customers. Features with a correlation greater than 0.6 with customer needs are selected to generate structured feature vectors. Finally, a preprocessing effect verification module is deployed on the edge computing node. If the basic accuracy of matching after data normalization on a certain platform is less than 85%, secondary field annotation is automatically triggered, and interpolation is performed on fields with a null value ratio greater than 10%. The time format is uniformly converted to YYYY-MM-DD HH:MM.

[0020] Weakly related features are uniformly classified into four categories: inquiry tone, page dwell time, page scrolling depth, and device operation frequency, such as the frequency of APP opening within 1 hour. The non-strongly related features mentioned in subsequent chapters are all included in this category. The tone of inquiry is extracted using NLP to determine two dimensions: urgency and clarity of need. Urgency is calculated based on the frequency of keywords such as "immediately" and "urgently needed," while clarity of need is scored based on the completeness of information such as the specific product model and specifications. Page dwell time is represented by the page jump sequence and the percentage of time spent on each page, such as homepage → product details page (40% dwell time) → review page (30% dwell time). The baseline accuracy is based on the manually labeled matching results of similar customer data over the past 30 days, i.e., accuracy = (number of customers correctly matched by the model / total number of manually labeled customers) × 100%.

[0021] Formula for screening weakly correlated features based on mutual information:

[0022] In the formula: Indicating weak customer association characteristics Customer needs tags The mutual information between them quantifies the strength of their association. By calculating this value, features with an association degree greater than 0.6 are selected. The larger the value, the stronger the explanatory power of the feature for the demand. This represents weakly correlated characteristics of the customers to be screened, including non-strongly correlated data such as inquiry tone, page dwell time, device operation frequency, and browsing path completeness; this variable can be discrete or continuous and needs to be preprocessed through edge nodes to be converted into a computable format; Customer demand tags are defined in the contextualized customer behavior semantic dictionary and set differently according to customer segmentation: high intention demand of new customers, core value demand of high-value customers, and reactivation demand of customers with churn warning, etc. This refers to customers who simultaneously possess weakly correlated feature values. and the value of the demand tag The probability is calculated by statistically analyzing the co-occurrence frequency of features and demands from multiple sources. This indicates that the customer has a weak correlation characteristic value. The probability, without considering the influence of demand tags, is obtained by statistically analyzing the characteristics of all customers. The frequency ratio of occurrence is calculated to eliminate the interference of the frequency of occurrence of the feature itself on the correlation judgment; This refers to the customer's requirement tag value. The probability of demand among all customers is determined by statistical analysis. The frequency of occurrence is calculated to standardize the joint probability and ensure that the mutual information value reflects only the specific association between the feature and the demand, rather than the prevalence of the demand itself. Used to measure weak association characteristics The uncertainty of the value means that the larger the entropy value, the more dispersed the feature values ​​are, reflecting the information richness of the feature itself. Features with too low marginal entropy will be initially excluded to avoid meaningless calculations. Labels indicating known customer needs Afterwards, characteristics The remaining uncertainty of the value; if the value is small, it indicates demand. It can significantly reduce The uncertainty at this time Large differences indicate high mutual information values ​​and strong feature correlations.

[0023] Specifically, step S2 is as follows: Based on the structured feature vector and cleaned data from step S1, to achieve secure collaborative matching of cross-domain data, the granularity of sensitive field desensitization is first designed according to customer value stratification. For example, high-value customer mobile phone numbers are desensitized to "138***1234 (VVIP)" and ID card numbers are desensitized to "3101011234 (High Value)" to retain the level label; ordinary customer mobile phone numbers are desensitized to "138***1234 (Ordinary)" and ID card numbers are desensitized to "3101011234" to reduce redundancy; new customer mobile phone numbers are desensitized to "138***1234 (New Customer)" to retain only the "New Customer" label. The desensitization process adopts... A differential privacy mechanism with a value of 0.3 ensures privacy and security; The value of 0.3 is set based on the following criteria: After multiple rounds of testing, this value meets two objectives: ① Privacy and security: The risk of leakage of sensitive customer information (phone number, ID number) is <0.01%; ② Data availability: Cross-domain feature matching accuracy is ≥88%, which is significantly higher than... =0.2 (accuracy 82%) =0.4 (0.05% risk of leakage) is better, achieving a balance between privacy and accuracy; Subsequently, a federated adversarial learning framework was established, deploying feature alignment sub-models at each data source node (offline stores, online APP, customer service system). The generator simulates online features based on offline store visit frequency and paper membership levels, while the discriminator optimizes parameters through an adversarial loss function to reduce cross-domain feature differences. The federated center sets up a cross-domain alignment effect feedback module. If the KL divergence between offline store consumption amount and online APP consumption amount is greater than 0.15, the generator loss function is adjusted and customer activity weights are added, with active customer weight = 1.2 and inactive customer weight = 0.8. If the mismatch rate of a node after alignment is greater than 20%, parameter transmission is paused and feature dimensions are verified. At the same time, SHA-256 hash mapping is performed on offline unstructured sensitive data (such as paper consumption preferences and customer service recording transcriptions) to convert them into irreversible feature codes containing only category and weight, such as maternal and infant preferences → #A73F2D: 0.75.

[0024] The structure and parameters of the feature alignment sub-model are based on: ① The generator uses a 3-layer fully connected network (input layer dimension = number of offline features, hidden layer dimensions 256 and 128, output layer dimension = number of online features, activation function ReLU); ② The discriminator uses a 2-layer fully connected network (input layer dimension = number of online features, hidden layer dimension 128, output layer dimension 1, activation function Sigmoid); ③ =0.6、 The setting of 0.4 is based on the fact that in historical cross-domain data alignment experiments, this ratio resulted in the smallest KL divergence between offline and online consumption amounts (average ≤ 0.12). Customer activity level classification criteria: Customer activity is determined by a combination of login frequency and purchase frequency in the past 30 days: ① Active customers: ≥10 logins and ≥3 purchases in the past 30 days; ② Low-activity customers: <5 logins or <1 purchase in the past 30 days; ③ Customers in between are calculated as active customer weight × 0.9. Category weighting is calculated based on the percentage of spending on that category over the past 90 days. Adversarial loss function for federated adversarial learning:

[0025] In the formula: This represents the total adversarial loss value, which measures the overall error in federated adversarial learning between the generator simulating online features and the discriminator distinguishing between real and simulated features. The smaller the loss value, the better the cross-domain feature alignment effect. This represents the classification loss weight coefficient, set to 0.6, used to adjust the importance of the discriminator to the accuracy of feature classification. The classification loss reflects whether the discriminator can correctly distinguish whether the feature source is real online data or generator-simulated data. The distribution loss weight coefficient is set to 0.4 and is used to adjust the importance of the discriminator to the consistency of cross-domain feature distribution. The distribution loss reflects the difference in probability distribution between online and offline features. The smaller the difference, the better the cross-domain alignment effect. This represents the classification loss value, which is the loss of the discriminator in determining whether a feature is a true online feature. For example, the cross-entropy loss. The smaller the value, the higher the classification accuracy. This represents the distribution loss value. The discriminator measures the loss of the difference between the probability distribution of offline features simulated by the generator and the real online features, such as the loss corresponding to KL divergence and JS divergence. The smaller the value, the closer the cross-domain feature distribution is.

[0026] Specifically, step S3 is as follows: Based on the cross-domain feature data aligned in step S2, in order to accurately capture dynamic customer needs, a multimodal triggering mechanism is first established by integrating three types of demand signals: text, behavior, and device. Text signals are triggered by the popularity of keywords in customer service dialogues (frequency of occurrence / total dialogue length > 0.3), behavior signals are triggered by the density of behavior within 1 hour (≥3 views of the same category, ≥2 items added to cart, density > 0.5), and device signals are triggered by the frequency of APP opening within 1 hour (≥4 times, frequency > 0.4). When two or more types of signals are satisfied, a demand mutation feature vector is generated. Subsequently, a demand intensity calculation model is established, and priority is quantified by demand intensity. Demand intensity ≥ 0.8 is pushed to prioritize inventory matching, 0.5 ≤ demand intensity < 0.8 is executed with personalized recommendations, and less than 0.5 is included in the long-term cultivation pool. Behavioral density and device signal frequency calculation method: ① Behavioral density = (Number of views + Number of items added to cart in the same category within 1 hour) / (Total number of products in this category on the platform × 0.5), for example: if a category has 100 products, and 3 products are viewed and 2 items are added to cart within 1 hour, the density = (3 + 2) / (100 × 0.5) = 0.1; ② Device signal frequency = Number of times the app is opened within 1 hour / Average number of times the app is opened by customers on the platform per day, for example: if the platform is opened an average of 10 times per day, and a customer opens the app 4 times within 1 hour, the frequency = 4 / 10 = 0.4; Keywords with strong demand relevance are divided into 5 categories: ① Product attribute (size, color, material); ② Service (delivery time, after-sales repair, installation); ③ Price (discounts, coupons, full reduction); ④ Inventory (in stock, out of stock, restock); ⑤ Decision-making (recommendation, comparison, place order). Only keywords within this category are counted to calculate popularity. Simultaneously, customer behavior dictionary filtering is performed on unstructured data including customer service dialogues and product reviews. A dedicated dictionary is built based on the customer's historical behavior, such as a high-value customer dictionary containing repeat purchase policies and a new customer dictionary containing new customer discounts. After filtering out redundant keywords, the data is input into the BERT-base model to generate a 768-dimensional semantic vector.

[0027] Formula for calculating customer demand intensity:

[0028] In the formula: This indicates the intensity of demand, used to quantify the urgency and priority of a customer's current needs, and its value ranges from [0, 1]. 0.8 Prioritize matching with pushed inventory; 0.5 0.8 Implement personalized recommendations; 0.5 was included in the long-term cultivation pool; This represents short-term behavior density, indicating the intensity of a customer's browsing and adding-to-cart activities within the same product category within one hour. The value ranges from [0, 1]. For example, adding two items of the same product category to the cart within one hour would indicate [high activity density]. Approaching 1, the weighting is 60%; This indicates the historical category matching degree, referring to the degree of overlap between the categories involved in the customer's current behavior and the categories they have historically purchased / preferred. The value ranges from [0, 1]. For example, a customer who has historically frequently purchased mother and baby products might be browsing mother and baby products if they are currently doing so. Approaching 1, the weighting is 30%; This indicates the urgency of the inquiry tone. Based on NLP analysis, it reflects the urgency level of the customer's tone in the customer service dialogue, taking values ​​[0, 1]. Values ​​containing words like "urgently needed" or "ship immediately" are considered urgency. Approaching 1, the weighting is 10%.

[0029] Formula for calculating the popularity of keywords in text signals:

[0030] Triggering conditions: When the keyword popularity exceeds 0.3, a text demand signal is triggered. In the formula: This indicates keyword popularity and is used to determine the prominence of keywords related to customer needs in customer service conversations. This indicates the frequency of keyword occurrences, referring to the number of times keywords strongly related to customer needs appear in customer service conversations, such as "discounts," "new products," and "after-sales service." For example, if "discounts" appears 3 times, then... ; This indicates the total length of a customer service conversation, referring to the total number of words or tokens in a single conversation between the customer service representative and the customer. For example, if the conversation contains 100 words, then... =100.

[0031] Specifically, step S4 is as follows: To enhance the temporal impact of short-term dynamic behaviors, a dual-time series matrix is ​​first established, consisting of a short-term series of nearly 72 hours and a long-term series of 90 days, with the short-term series weighted according to behavioral value: purchase and repurchase = 3, add to cart and exclusive consultation = 2, and browsing and collection = 1. The long-term series is used to form a baseline by statistically analyzing the daily behavior frequency and category preference ratio. The impact of short-term high-value behaviors is enhanced through weighted fusion. Then, a long-term behavioral trend database is established to store the stability of category preferences (stability is defined as a fluctuation coefficient < 0.3) and the consistency of behavioral patterns (overlap of active periods > 70%) over 90 days. If short-term behavior conflicts with long-term trends, such as home appliance customers frequently browsing maternal and infant products, then related data such as newborn family information are checked. When the correlation is less than 0.3, the short-term weight is reduced to 0.5 times the original weight to filter noise. When the correlation is 0.3 ≤ R ≤ 0.5, the short-term weight is retained at 0.8 times the original weight to balance noise filtering and behavioral effectiveness. When the correlation is greater than 0.5, the short-term weight is retained. At the same time, time-series feature attention pooling is used to compress high-frequency repetitive behaviors, retaining the first or last time and frequency and merging consecutive identical behaviors, such as continuously adding 3 products of the same category to the cart and marking the category concentration as 1.0.

[0032] The relevance R is calculated using a multi-dimensional weighted average: R = Category overlap × 0.5 + Behavioral time relevance × 0.3 + Consumption capacity matching degree × 0.2; Among them, ① Category overlap = number of intersections between short-term behavior categories and long-term preferred categories / total number of long-term preferred categories; ② Behavior time correlation = overlap duration between short-term behavior occurrence period and long-term active period / total long-term active period; ③ Consumption capacity matching degree = average order value of goods involved in short-term behavior / long-term average order value, with a value range of [0,1], and values ​​exceeding 1 are calculated as 1. The volatility coefficient is calculated using the standard deviation / mean of the percentage of weekly spending on this product category over the past 90 days, as shown in the formula:

[0033] in The standard deviation of the percentage for each week. This represents the average percentage for each week. A value <0.3 indicates stable category preference; The short-term behavior value weight is set based on the average conversion rate of each behavior over the past 180 days: purchase conversion rate 35%, repeat purchase conversion rate 30% (weight=3); add-to-cart conversion rate 15%, personalized consultation conversion rate 12% (weight=2); browsing conversion rate 5%, favorite conversion rate 3% (weight=1). The weight value = (conversion rate of this behavior / minimum conversion rate) × 0.1, and is rounded to the nearest integer after business calibration. Short-run weighting formula for time-series conflicting behaviors: When the correlation At 0.3: 0.5 When the correlation 0.5 hours: In the formula: This represents the corrected short-term behavior weight, used to adjust the short-term behavior weight when there is a conflict between short-term behavior and long-term trend, and to avoid noise interfering with the matching results. This indicates the weight of short-term behaviors before the correction, based on the value of short-term behaviors in the document, such as purchase and repeat purchase. Add to cart, personalized consultation Browse and bookmark ; The degree of relevance refers to the matching degree between short-term conflict behavior and customer-related data. For example, if a customer with a preference for home appliances browses mother and baby products in the short term, it is necessary to verify the family information of newborns. This reflects the correlation between the information and short-term behavior, and takes values ​​[0, 1]. 0.5 represents the weighting adjustment coefficient, when When the value is 0.3 (short-term behavior has no reasonable supporting correlation and is likely noise), the weight of short-term behavior is reduced to 50% of the original weight to filter out the influence of invalid behavior.

[0034] Specifically, step S5 is as follows: Based on the time-series feature vector fused in step S4, in order to adapt to the matching needs of different business scenarios, a dynamic weight matrix with two dimensions of customer value and business scenario is first established based on the DQN algorithm. The core dimensions include five categories: identity, needs, behavior, preferences, and scenario, and different weights are assigned according to scenario and customer value. The DQN algorithm achieves dynamic weight iteration through reinforcement learning: ① Reward function design: A positive reward of +0.2 is given when the matching conversion rate is 10% higher than the target value, a positive reward of +0.15 is given when customer satisfaction is ≥85%, and a negative reward of -0.1 is given when the response time exceeds the threshold (normal matching >100ms); ② Iteration strategy: The weight matrix is ​​updated every 2 hours based on the reward value. - Greedy strategy (ε=0.1, 90% probability of choosing the current optimal adjustment direction, 10% probability of exploring a new direction) to avoid local optima; ③ Constraints: The weight of core dimensions (identity, needs, behavior) shall not be less than 60% of the total weight to prevent non-key dimensions from excessively affecting the matching results; Scenarios and Customer Value Weighting Table: ; Deploy real-time performance monitoring units to calculate matching conversion rates and customer satisfaction every hour. If the conversion rate is 10% lower than the target, the weight of the dimension that did not meet expectations will be increased. For example, if price sensitivity is reduced from 0.3 to 0.5, or satisfaction is less than 80%, the weight of the dimension that affects experience will be reduced, with an adjustment range of ≤0.2, in order to avoid fluctuations. For scenarios such as holidays, member days, and inventory clearance, add peripheral dimensions, such as "holiday category preference" for holidays and "points redemption willingness" for member days, to participate in the matching calculation together with the core dimensions to form a scenario-based matching matrix.

[0035] Dynamic weight matrix adjustment formula: When the conversion rate does not meet the target ( 90%) When satisfaction level is not met ( 80%) In the formula: This indicates the adjusted dimension weights, which refer to the new weights of a core dimension (such as price sensitivity and repurchase intention) in the dual-dimensional matrix of customer value and scenario, used to adapt to the needs of business performance optimization; This indicates the dimension weight before adjustment, the initial weight of that core dimension before adjustment, such as the initial weight of price sensitivity of high-value customers in a promotional scenario. =0.4; Indicates the magnitude of weight adjustment. 0.2, to avoid excessive fluctuations in weights that could cause model oscillations and ensure smooth adjustments; The actual matching conversion rate refers to the proportion of customers who generate target behaviors such as purchase or repeat purchase after matching based on the current weight matrix. This refers to the target conversion rate, the benchmark conversion rate set by the business to match the desired performance, such as the target conversion rate. =30%, then The weight adjustment is triggered at 27%; Customer satisfaction refers to the customer's rating of their satisfaction with the matching results (such as recommended products or service responses), expressed as a percentage. 80% of the time, the weight of dimensions that affect the user experience is reduced.

[0036] Specifically, step S6 is as follows: Based on the scenario-based matching matrix in step S5, in order to improve matching efficiency and response speed to high-value customers, an abnormal high-value behavior identification and verification process is first established. For behaviors such as single consumption exceeding 5 times the historical average and high-frequency inquiries about high-value products within 1 hour (average order value > 3 times the category average), historical association verification (whether there are prior behaviors such as high-value product collection or large-value coupon redemption in the past 90 days, with an association degree > 0.7 for validity) and scenario association verification (whether it is in a scenario such as pre-sale promotion or member day, with a matching degree > 0.8 for reasonableness) are performed. If both verifications are passed, a fast channel is triggered, skipping the regular feature verification and entering the high-priority queue. Dual verification pass criteria: ① Historical correlation verification: The correlation between the previous behavior in the past 90 days (collecting high-value products, receiving large coupons) and the current abnormal behavior is ≥0.7; ② Scenario correlation verification: The matching degree between the current scenario (pre-sale promotion, member day) and the customer's historical high-activity scenarios is ≥0.8 (matching degree = number of times the customer participated in this scenario in history / total number of times the customer participated in active scenarios). If the double verification fails: ① If the historical correlation is 0.5≤R<0.7 and the scenario matching degree is 0.6≤R<0.8, it will be included in the ordinary priority queue and the customer's recent browsing history will be checked; ② If the correlation or matching degree is lower than the above threshold, it will be judged as invalid abnormal behavior, and only the behavior log will be recorded, and it will not participate in the matching calculation. The definition of high frequency for high-value products in one hour is: ≥3 inquiries, with an interval of <10 minutes between each inquiry, and the inquiry content includes decision-making terms such as order placement, inventory, and price confirmation; Based on a short-term behavior index pool of nearly 72 hours, a scenario hotspot preloading module is deployed. 24 hours before the promotion, historical promotional categories and coupon records are preloaded to generate pre-computed vectors. 12 hours before the new product launch, historical new product trajectories and purchase records of the same category are preloaded to establish a pre-matching candidate set. During the event, the preloaded features are directly called to shorten the response time.

[0037] Incremental computing is scheduled using a dual-dimensional approach based on customer value and behavioral urgency. 40% of computing resources are allocated to high-value customers who make purchases or inquire about after-sales service (priority 1), 35% to ordinary customers who add items to their cart or inquire about new products (priority 2), 15% to customers with churn warnings who log in or click on promotions (priority 3), and 10% to low-value customers who simply browse (priority 4). Feature extraction and matching updates are performed only on newly added data, and old results are replaced by incremental indexes to optimize computing resource allocation.

[0038] Customer value is categorized based on average order value and repurchase frequency over the past 90 days: ① High-value customers: Average order value > 2 times the category average and repurchase ≥ 5 times; ② Ordinary customers: Average order value ≤ 1.5 times the category average and repurchase 1-4 times; ③ Customers at risk of churn: Login frequency decreased by > 50% in the past 30 days or no purchase in the past 60 days; ④ Low-value customers: Average order value < 0.5 times the category average and repurchase = 0. Specifically, step S7 is as follows: Based on the business data (conversion rate, mismatch rate) and customer feedback generated during the matching process in step S6, semantic conflicts are detected through two mechanisms: ① Real-time cross-platform comparison: Semantic mapping verification is performed on the behavioral tags of online APP, offline stores, and customer service system every hour. If the tag mapping difference for the same customer and the same behavior is greater than 30% (e.g., e-commerce is labeled with high intent, social media is labeled with low intent), a conflict warning is triggered; ② Customer feedback trigger: If the customer feedback indicates that the recommended category does not match their needs, and verification finds that it is caused by cross-platform semantic mapping deviation, it is classified as a semantic conflict. First, cross-platform semantic conflict feedback is divided into three levels: Level 1 is a fatal conflict, such as an e-commerce order that has been shipped and a logistics order that has not been shipped, or an e-commerce order that has been canceled and a payment platform refund that has been successful. The customer behavior semantic dictionary is updated and the edge nodes are synchronized within 2 hours. Level 2 is a significant conflict, such as a discrepancy between social "interests" and e-commerce "category browsing". The tag mapping rules are corrected within 12 hours. Level 3 is a minor conflict, such as "add to cart" and "wish list". The descriptions are updated in batches within 24 hours. Then, a dual-indicator evaluation system of matching accuracy and response time is established. Matching accuracy is monitored by accuracy (≥90%), recall (≥85%), and mismatch rate (≤5%). Response time is monitored by regular matching (≤100ms), fast channel (≤50ms), and peak period (≤150ms). If the accuracy improves but the response time increases, the number of BERT layers is reduced (from 12 layers to 6 layers) and the number of Attention-LSTM attention heads is simplified (from 8 heads to 4 heads) to achieve lightweighting. Meanwhile, feedback from dissatisfied customers is collected through lightweight questionnaires, including two types of information: first, the reasons for dissatisfaction, including mismatches in category / price / timeliness; second, the desired matching direction. Key needs are extracted from the text feedback using NLP, such as more affordable maternity and baby products, or home appliance accessories of the same brand. At least 1,000 feedback data points are incorporated into the model iteration every month. For mismatches in category, the BERT semantic weights are adjusted, and for mismatches in price, the price sensitivity parameter in the dynamic matrix is ​​optimized to continuously improve the matching fit.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent customer information matching based on big data, characterized in that: The specific steps of this method are as follows: Step S1: First, define mapping rules for different behavioral tags according to customer type. At the same time, use a dual-weight algorithm of mutual information and customer value to filter weakly correlated features, generate structured feature vectors, and finally standardize the data through the preprocessing effect verification module. Step S2: Based on the standardized data in Step S1, design the sensitive field desensitization granularity according to customer value stratification, then establish a federated adversarial learning framework, and perform hash mapping on offline unstructured sensitive data. Step S3: Integrate the three types of demand signals—text, behavior, and device—to establish a multimodal triggering mechanism, generate a demand mutation feature vector, quantify priority through demand intensity, filter and encode unstructured data; Step S4: Establish a dual time series matrix of 72-hour short-term and 90-day long-term, use the long-term behavior trend library to correct conflict behaviors, compress high-frequency repetitive behaviors, and strengthen the impact of short-term high-value behaviors. Step S5: Use the DQN algorithm to establish a dynamic weight matrix with two dimensions: customer value and scenario. Distribute weights according to the differences between scenario and customer value. Adjust the weights according to conversion rate and satisfaction. Set edge dimensions for subdivided scenarios to form a scenario-based matching matrix. Step S6: Based on the matching matrix in step S5, perform double verification on abnormal high-value behaviors. If they pass the verification, they enter the high-priority queue. Then, incremental calculation is scheduled according to customer value and behavior urgency to optimize the allocation of computing resources. Step S7: Handle semantic conflicts in a hierarchical manner, use matching accuracy and response time as dual indicators to evaluate and optimize the model, actively collect feedback and incorporate it into model iteration to improve matching fit.

2. The intelligent customer information matching method based on big data according to claim 1, characterized in that: The specific steps of S1 are as follows: First, a scenario-based customer behavior semantic dictionary is established based on new customers, high-value customers, and customers at risk of churn. Mapping rules for different behavioral tags are defined according to customer type. For new customers, browsing time of more than 3 minutes and more than 2 inquiries are mapped as high-intent behavior, and page scrolling depth of more than 80% is mapped as in-depth understanding behavior. For high-value customers, repeat purchase interval of less than 30 days, average order value more than twice the category average, and dedicated customer service consultation are mapped as core value behavior. For customers at risk of churn, a 50% decrease in login frequency in nearly 30 days and items left in shopping cart for more than 7 days are mapped as reactivation demand behavior. The scenario-based customer behavior semantic dictionary is dynamically updated every 24 hours based on cross-platform semantic conflict feedback. Simultaneously, a dual-weight algorithm of mutual information and customer value is used to screen weakly correlated features, including consultation tone and page dwell time. Different weights are set according to customer value, and features with a correlation greater than 0.6 with customer needs are selected to generate structured feature vectors. Finally, a preprocessing effect verification module is deployed on the edge computing node. If the basic accuracy of matching after normalization of data on a certain platform is less than 85%, secondary field annotation is automatically triggered, and interpolation is performed to complete fields with a null value ratio greater than 10%, and the time format is unified.

3. The intelligent customer information matching method based on big data according to claim 2, characterized in that: Step S2 is as follows: Based on the structured feature vectors and cleaned data from step S1, the granularity of sensitive field desensitization is first designed according to customer value stratification. The desensitization process adopts... A differential privacy mechanism with a value of 0.3 ensures privacy and security; Subsequently, a federated adversarial learning framework was established, deploying feature alignment sub-models at each data source node. The generator simulates online features based on offline store visit frequency and paper membership levels, while the discriminator optimizes parameters through an adversarial loss function to reduce cross-domain feature differences. The federated center sets up a cross-domain alignment effect feedback module. If the KL divergence between offline store consumption amount and online APP consumption amount is greater than 0.15, the generator loss function is adjusted and customer activity weights are added. If the mismatch rate after alignment at a certain node is greater than 20%, parameter transmission is paused and feature dimensions are verified. At the same time, SHA-256 hash mapping is performed on offline unstructured sensitive data to convert it into irreversible feature codes containing only category and weight.

4. The intelligent customer information matching method based on big data according to claim 3, characterized in that: Step S3 is as follows: Based on the cross-domain feature data aligned in step S2, a multimodal triggering mechanism is first established by integrating three types of demand signals: text, behavior, and device. Text signals are triggered by the popularity of keywords in customer service dialogues, behavior signals are triggered by the behavior density within one hour, and device signals are triggered by the frequency of APP opening within one hour. When two or more types of signals are satisfied, a demand mutation feature vector is generated. Subsequently, a demand intensity calculation model was established, and priority was quantified by demand intensity. Demand intensity ≥ 0.8 was pushed to the inventory for priority matching, demand intensity ≤ 0.5 < 0.8 was used for personalized recommendations, and demand intensity less than 0.5 was included in the long-term cultivation pool. Simultaneously, customer behavior dictionary filtering is performed on unstructured data, including customer service dialogues and product reviews. A dedicated dictionary is built based on customers' historical behavior, and after filtering out redundant keywords, it is input into the BERT-base model to generate a 768-dimensional semantic vector.

5. The intelligent customer information matching method based on big data according to claim 4, characterized in that: Step S4 is as follows: First, a dual time-series matrix is ​​established with a short-term series of nearly 72 hours and a long-term series of 90 days as the unit of time. The short-term series is weighted according to the behavioral value, and the long-term series is statistically analyzed to form a baseline based on the daily behavior frequency and category preference ratio. The influence of short-term high-value behaviors is strengthened by weighted fusion. Then, a long-term behavioral trend database is established to store the stability of category preferences and the consistency of behavioral patterns over 90 days. If short-term behavior conflicts with long-term trends, the related data is checked. When the correlation is less than 0.3, the short-term weight is reduced to 0.5 times the original weight; when the correlation is 0.3≤R≤0.5, the short-term weight is retained at 0.8 times the original weight; when the correlation is greater than 0.5, the short-term weight is retained. At the same time, time-series feature attention pooling is used to compress high-frequency repetitive behaviors, retaining the first or last time and frequency, and merging consecutive identical behaviors.

6. The intelligent customer information matching method based on big data according to claim 5, characterized in that: Step S5 is as follows: Based on the time-series feature vector fused in step S4, a dynamic weight matrix with two dimensions of customer value and business scenario is first established based on the DQN algorithm. The core dimensions include five categories: identity, needs, behavior, preferences, and scenario, and different weights are assigned according to scenario and customer value. Deploy a real-time performance monitoring unit to calculate the matching conversion rate and customer satisfaction every hour. If the conversion rate is 10% lower than the target, the weight of the dimension that did not meet expectations will be increased. If the satisfaction rate is less than 80%, the weight of the dimension that affects the experience will be decreased. The adjustment range is ≤0.2 to avoid fluctuations. For scenarios including holidays, member days, and inventory clearance, add peripheral dimensions to participate in the matching calculation together with the core dimensions to form a scenario-based matching matrix.

7. The intelligent customer information matching method based on big data according to claim 6, characterized in that: Step S6 is as follows: Based on the scenario-based matching matrix in step S5, an abnormal high-value behavior identification and verification process is first established. For behaviors including single consumption exceeding 5 times the historical average and high-frequency inquiries about high-value products within 1 hour, historical association verification and scenario association verification are performed. If both verifications are passed, a fast channel is triggered, skipping the regular feature verification and entering the high-priority queue. Based on a short-term behavior index pool of nearly 72 hours, a scenario hotspot preloading module is deployed. 24 hours before the promotion, historical promotional categories and coupon records are preloaded to generate pre-computed vectors. 12 hours before the new product launch, historical new product trajectories and purchase records of the same category are preloaded to establish a pre-matching candidate set. During the event, the preloaded features are directly called to shorten the response time. Incremental computing is scheduled using a dual-dimensional approach that considers both customer value and behavioral urgency. 40% of computing resources are allocated to high-value customers who make purchases or inquire about after-sales service, 35% to ordinary customers who add items to their cart or inquire about new products, 15% to customers with churn warnings who log in or click on promotions, and 10% to low-value customers who simply browse. Feature extraction and matching updates are performed only on newly added data, and old results are replaced by incremental indexes to optimize the allocation of computing resources.

8. The intelligent customer information matching method based on big data according to claim 7, characterized in that: Step S7 is as follows: First, cross-platform semantic conflict feedback is divided into three levels: Level 1 is a fatal conflict, in which the customer behavior semantic dictionary is updated and edge nodes are synchronized within 2 hours; Level 2 is a significant conflict, in which the tag mapping rules are corrected within 12 hours; and Level 3 is a minor conflict, in which the descriptions are updated in batches within 24 hours. Then, a dual-indicator evaluation system of matching accuracy and response time is established. Matching accuracy is monitored by accuracy, recall, and mismatch rate, while response time is monitored by regular matching, fast channel, and peak period. If the accuracy improves but the response time increases, the number of BERT layers is reduced and the number of attention heads of Attention-LSTM is simplified to achieve lightweighting. Meanwhile, feedback is collected from unsatisfied customers through lightweight questionnaires. Key needs are extracted from the text feedback using NLP, and the feedback data is integrated into the model iteration. If the product category does not match, the BERT semantic weights are adjusted; if the price does not match, the price sensitivity parameter in the dynamic matrix is ​​optimized.

Images