A customer portrait construction method based on big data collection

Abstract

The application relates to the technical field of customer portrait construction, and discloses a customer portrait construction method based on big data collection, which comprises the following steps: collecting two types of core data sources of website click logs and transaction records, constructing a customer feature matrix, and performing missing value processing and standardization; automatically evaluating feature significance in combination with the processed customer feature matrix, and generating a feature weight vector; calculating a customer portrait scalar by using weighted summation, and realizing unsupervised clustering by using adaptive threshold density clustering based on the portrait and a near neighbor distance set; finally, combining the portrait score and the clustering label to perform differentiated marketing management and resource allocation, and real-time self-correcting clustering parameters according to the marketing conversion rate change, so that closed-loop optimization from data collection to marketing execution is realized, and high real-time performance and adaptability are achieved.

Description

Technical Field

[0001] This invention relates to the field of customer profile building technology, specifically a method for building customer profiles based on big data collection. Background Technology

[0002] Customer profiling is a key supporting technology for precision marketing and customer relationship management in modern enterprises. Its construction process relies on the in-depth collection, analysis and modeling of customer behavior data. With the development of digital platforms and online businesses, customers frequently leave behavioral traces such as visits, transactions and interactions in multiple channels, making the sources of profile data increasingly rich and the dimensional structure increasingly complex. However, despite the sufficient amount of data, existing customer profiling construction technologies still have shortcomings in several aspects, especially in terms of the linkage of actual marketing strategies and the ability of the system to adaptively optimize.

[0003] Most current customer profiling technologies rely on offline data collection based on platform logs and construct user profiles through manually set rules or predefined weight combinations. Feature selection is still largely expert-driven, using fixed weights or single-indicator ranking for subjective screening. While some technologies incorporate statistical methods such as information gain and mutual information to assess feature importance, they lack a comprehensive ability to measure the stability of the global feature distribution. In clustering modeling, most solutions depend on classic algorithms like K-Means or DBSCAN, but often use fixed parameter settings, lacking a dynamic adaptation mechanism to data distribution fluctuations. For cold-start customers or short-term fluctuation scenarios, model construction struggles to provide stable support. Furthermore, in the profile output application stage, most systems only generate static labels or classification results for subsequent use, lacking an effective feedback mechanism between the profile and actual marketing campaign results.

[0004] To this end, this invention proposes a customer profile construction method that is fully adaptive, has no experience-based weighting, and allows for online feedback, in order to support the continuous optimization of enterprise digital marketing and precision marketing solutions. Summary of the Invention

[0005] This invention provides a customer profile construction method based on big data collection, which helps to solve the problems mentioned in the background art, such as data collection relying on offline logs, excessive reliance on expert preset rules or fixed weights, lack of comprehensive evaluation of global distribution stability, fixed algorithm parameters, lack of dynamic adaptation to data fluctuations, weak cold start and short-term fluctuation handling, and only outputting static labels, lacking a closed-loop optimization mechanism with actual marketing effects.

[0006] This invention provides the following technical solution: a method for constructing customer profiles based on big data collection, comprising:

[0007] Collect multi-source behavioral and transaction data containing unique customer identifiers to generate the original dataset D. clicks and D tx ;

[0008] Based on the original dataset, numerical features related to customer behavior are extracted, and a customer feature matrix X is constructed.

[0009] The customer feature matrix is ​​processed for missing values ​​and normalized to output a standardized matrix X' = [x' ij ];

[0010] The significance score s for each feature is calculated based on the normalized matrix. j Generate feature saliency vector

[0011] Based on the standardized features and the aforementioned saliency vector, calculate the scalar value p for each customer's profile. i ;

[0012] Based on the customer profile scalar and the nearest neighbor distance set calculated from it, a density-based clustering algorithm is used to segment customers, generating a customer segmentation label set {C}. i};

[0013] Based on the aforementioned customer profile scalar and group tag set, personalized marketing tiered management is implemented, including resource allocation and content delivery strategy formulation, and the grouping parameters are adaptively adjusted based on changes in customer conversion rates after grouping.

[0014] Optionally, the collection includes multi-source behavioral data and transaction data containing unique customer identifiers, generating an original dataset D. clicks and D tx ,include:

[0015] Request permission to collect customer website click logs and CRM transaction records;

[0016] Set the time interval for collecting click logs from the customer's website, and denot it as the log collection interval;

[0017] Set the daily collection time for customer CRM transaction records, and record it as the record collection node;

[0018] For the customer's website click logs, export the fields {customer_id, page_url, event_time} from the web server at every log collection interval;

[0019] For customer CRM transaction records, at each record collection node, export a standard table containing the fields {customer_id, transaction_amount, transaction_time, channel} from the CRM system;

[0020] Create the original record set D of the click log clicks and the original record set D of transaction records tx ;

[0021] Write the exported website click logs to D clicks In the format, each record is formatted as follows:

[0022] d i =(customer_id,page_url,event_time)

[0023] Write the exported CRM transaction records to D tx In the format, each record is formatted as follows:

[0024] d i =(customer_id,transaction_amount,transaction_time,channel)

[0025] Where d i Indicates that they are respectively in D clicks and D tx The website click logs and CRM transaction records of customer i.

[0026] Optionally, the step of extracting numerical features related to customer behavior and constructing a customer feature matrix X based on the original dataset includes:

[0027] From D clicks and D tx In this context, the following basic characteristics are constructed for each customer i:

[0028] f i1 Website visits in the past 7 days;

[0029] f i2 Total transaction amount over the past 30 days;

[0030] f i3 : Number of days since the last visit;

[0031] f i4 : The number of days since the last transaction;

[0032] f i5 : Main transaction channel codes, including online=1, offline=2;

[0033] Arrange all customers into a matrix:

[0034] X = [f ij ] n×5

[0035] Each row of the matrix represents a customer, and the i-th row is the original feature vector of customer i, which is f in sequence. i1 f i2 f i3 f i4 f i5 .

[0036] Optionally, the step of performing missing value processing and normalization on the customer feature matrix to output a standardized matrix X' includes:

[0037] Obtain the number of missing values ​​n in the j-th column of the customer feature matrix X. miss,j The absence indicates the presence of a null value, null, or a special placeholder in the j-th column;

[0038] Calculate the missing percentage for the j-th column of the customer feature matrix X:

[0039]

[0040] Where n is the number of rows in the customer feature matrix X;

[0041] Set a threshold τ to determine whether there are too many missing values;

[0042] If r j <τ, then the median m of the non-empty values ​​in the j-th column. j Fill missing information;

[0043] Among them, the median m j This refers to the (n / 2)th data point after sorting the non-empty values ​​in column j in ascending order;

[0044] Otherwise, if the j-th column is found to be missing too much data, the data in the j-th column will be deleted.

[0045] Optionally, the step of performing missing value processing and normalization on the customer feature matrix to output a standardized matrix X' further includes:

[0046] Calculate for each column after missing data filtering:

[0047]

[0048] Where, μ j Let σ represent the mean of the j-th feature. j This represents the standard deviation of the j-th feature;

[0049] Perform normalization transformation:

[0050]

[0051] Where f ij Derived from the customer feature matrix X, it represents the j-th feature of the i-th record;

[0052] Output the normalized matrix X' = [x' ij ].

[0053] Optionally, the step of calculating the saliency score of each feature based on the normalized matrix to generate a feature saliency vector is performed. include:

[0054] For each column of the normalized matrix X':

[0055] Calculate the probability assignment matrix p ij :

[0056]

[0057] Calculate information entropy e j :

[0058]

[0059] Among them, e j ∈[0.1], representing the information entropy of the j-th feature;

[0060] Assign significance scores:

[0061]

[0062] Where m represents the number of features in the customer feature matrix, and its value is 5; e k s represents the information entropy of the k-th feature among all features; j ∈(0,1);

[0063] The significance score of each column of the standardized matrix X' is calculated, and the scores are summed to obtain the significance vector.

[0064] Optionally, calculating the profile scalar value for each customer based on the standardized features and the saliency vector includes:

[0065] By weighting the normalized features using significance scores, a scalar profile of each customer is obtained;

[0066]

[0067] Where m represents the number of features in the customer feature matrix, with a value of 5; s jx' is the significance score of the j-th feature; ij This is the standardized value of the i-th customer after normalization on the j-th feature.

[0068] Optionally, based on the customer profile scalar and combined with the nearest neighbor distance set calculated therefrom, a density-based clustering algorithm is used to segment customers and generate a customer segmentation label set, including:

[0069] Let k be used for neighbor distance statistics, and for each customer profile point p i Calculate the distance d between its k-th nearest neighbor. i,(k) , forming a set n is the total number of customers;

[0070] Construct an adaptive threshold ε:

[0071]

[0072] in, This is the distance from the median. α is the interquartile range; α is the initial empirical coefficient used to adjust the distance threshold in density clustering.

[0073] The distances to the median and interquartile range are specifically as follows:

[0074] Will Sort the elements in ascending order, and take the n / 2th element after sorting. Use this element as the median distance.

[0075] Will After sorting in ascending order, select the 25%n element as Q1 and the 75%n element as Q3.

[0076] Density clustering technology is used, and DBSCAN({p i}, ε, minPts=k), generate customer segmentation label set {C i}

[0077] Optionally, the personalized marketing tiered management based on the customer profile scalar and the grouping tag set includes resource allocation and content delivery strategy formulation, and adaptive adjustment of grouping parameters based on changes in customer conversion rates after grouping, including:

[0078] All customers are categorized by their label C i Perform clustering, with each customer segmentation tag set forming a cluster;

[0079] For each cluster, calculate the average image value:

[0080]

[0081] Where |c| represents the number of customers within cluster c; ∑ i∈c p i This represents the sum of profiles for all customers within cluster c;

[0082] Allocate the promotion budget proportionally:

[0083]

[0084] Among them, B c B represents the marketing budget allocated to customers within cluster c; total This indicates the total marketing budget available during the current marketing cycle; The sum of the average values ​​of the images for all clusters.

[0085] Optionally, the step of performing personalized marketing tiered management based on the customer profile scalar and the grouping tag set, including resource allocation and content delivery strategy formulation, and adaptively adjusting the grouping parameters based on changes in customer conversion rates after grouping, further includes:

[0086] Set behavioral goals for successful conversion;

[0087] After each round of marketing campaigns, the total number of customers in cluster c who exhibited the desired behavior is recorded as the number of successfully converted customers, V.

[0088] Calculate the conversion rate A of cluster c. (t) :

[0089]

[0090] Where |c| represents the number of customers within cluster c;

[0091] Calculate the difference in conversion rate ΔA between two adjacent rounds of marketing campaigns. c :

[0092]

[0093] in, This indicates the conversion rate after this round of marketing campaign. This indicates the conversion rate after the previous round of marketing campaigns;

[0094] Set a threshold δ to determine whether to update the clustering threshold coefficient α;

[0095] If |ΔA|≤δ, then keep the original clustering threshold coefficient α unchanged;

[0096] When |ΔA|>δ, then update the clustering threshold coefficient α:

[0097] α (t+1) =α t ×(1+βsign(ΔA))

[0098] Where, α t This represents the clustering threshold coefficient used in this round of marketing; β represents the step size factor, which controls the magnitude of each α update; sign(ΔA) is the sign function, which increases the threshold with a positive sign and decreases the threshold with a negative sign.

[0099] With α (t+1) Reconstruct the adaptive threshold ε:

[0100]

[0101] ε t+1 Used for customer profiling and clustering in the next round of marketing.

[0102] The present invention has the following beneficial effects:

[0103] 1. This solution only collects data from two core data sources: website click logs and CRM transaction records, significantly reducing system integration complexity. Compared to traditional approaches that require access to multiple platforms and data formats, this solution only requires maintaining two stable data retrieval scripts and setting clear field formats, reducing engineering integration and maintenance costs. Simultaneously, through a fixed-frequency batch retrieval mechanism, data synchronization can be ensured within a controllable time window, thus meeting near real-time data requirements while avoiding the high concurrency and system pressure associated with streaming processing. This design balances system stability and flexibility, providing a solid data foundation for subsequent feature construction and model applications.

[0104] 2. Based on the original data, this solution extracts only five key features directly related to marketing management, such as recent visit frequency, total transaction amount, and behavior time interval, avoiding computational redundancy caused by "feature dimension inflation." Through clear feature definitions and a fixed dimension order, features can be reused across different projects or business scenarios, and new marketing campaigns can be launched quickly. This method can also flexibly add or remove features according to actual needs, effectively balancing feature richness and computational efficiency. More importantly, outputting multi-source heterogeneous data through the same matrix greatly simplifies the subsequent algorithm integration interface, making engineering implementation more standardized and easier to test. This step ensures that key business features are fully preserved while avoiding interference from irrelevant features, thereby improving the training speed and performance stability of subsequent models.

[0105] 3. This solution employs a "median-filled or column-wide removal" strategy, achieving precise completion and cleaning through dynamic assessment of the missing percentage in each column. When the missing percentage is below a threshold, the column is filled with the true median, preserving most of the true data distribution while preventing bias introduced by mean-filled data. When the missing percentage is too high, the entire column is deleted, avoiding excessive noise for subsequent analysis. Compared to traditional methods that use deletion or mean-filled data, this method ensures the integrity of the model input while effectively resisting the influence of outlier data. Furthermore, the median calculation rule is independent of distribution assumptions, making it applicable to any non-normal distribution scenario, further enhancing the versatility and stability of the solution in multi-industry applications.

[0106] 4. By performing standard deviation normalization on the cleaned matrix, all features can be unified to the range of zero mean and unit variance, eliminating the weight shift caused by differences in feature dimensions and value ranges. Compared with methods such as min-max scaling that are sensitive to outliers, this normalization method can handle extreme data points more robustly. This step ensures that when calculating the significance of features or distance metrics in the subsequent calculation, it will not cause dominance or distortion due to a single feature value being too large or too small. At the same time, the standardized data is more likely to meet the assumptions of statistics and algorithms regarding the input distribution, improving the convergence speed and numerical stability of the algorithm.

[0107] 5. This solution automatically assigns a significance score to each feature based on information entropy, without relying on empirical coefficients or external training data, and measures its information value from the perspective of the feature's own distribution. Unlike traditional expert scoring or PCA dimensionality reduction, the entropy weight method quantifies the dispersion and information contribution of features on the global sample, thereby dynamically adjusting feature weights. This mechanism not only reduces human subjective bias but also improves model interpretability, clearly indicating which features have the greatest impact on customer profiles. At the same time, this method can adapt to fluctuations in sample size and the dynamic addition of new features, and has excellent online adaptive capabilities.

[0108] 6. By weighted summing of standardized features and saliency vectors, this solution generates a single comparable profile scalar value, significantly simplifying the cost of using multi-dimensional profile information. This scalar has both behavioral activity and consumer value meanings, and can be directly used for customer ranking and priority determination. Compared to the complex method of requiring multiple queries and combining multiple tags, the profile scalar is clear at a glance and easy to use directly in marketing management systems, BI reports, and decision-making processes, improving business execution efficiency. In addition, customer groups can be segmented by threshold or level in different business scenarios or advertising channels, enabling highly flexible strategy formulation.

[0109] 7. By dynamically calculating the cluster radius based on the median and interquartile range of the k-th nearest neighbor distance set in the sample, combined with empirical coefficients, no manual parameter tuning is required. Compared with the traditional DBSCAN mode that requires a fixed threshold, the dynamic threshold makes the clustering process more in line with the actual data distribution and can automatically adapt to changes in sample density, avoiding over-clustering or splitting caused by improper parameter settings in a single instance. This method can also resist clustering failure caused by local density imbalance, making the clustering results more robust. At the same time, the threshold is calculated based on statistical distribution and is not sensitive to extreme values, making it suitable for large-scale online business environments.

[0110] 8. By using the average profile value of customer groups as the basis for budget allocation, marketing funds are invested more accurately and with higher returns. Unlike traditional strategies based on historical experience or uniform discounts, this method ensures that high-value groups receive more resource support and improve their conversion efficiency. At the same time, low-value groups can receive targeted advertising, saving costs. The budget allocation formula is simple to implement and can be seamlessly integrated with the company's existing financial system and advertising platform.

[0111] 9. This solution monitors the conversion rate changes of each group after each round of marketing, automatically adjusts the density threshold coefficient, and recalculates the cluster radius to achieve closed-loop self-adaptation of profiles and clustering parameters. This mechanism enables customer segmentation and profiles to be continuously optimized with market feedback, avoiding the decrease in accuracy caused by model obsolescence. Compared with manual periodic review and parameter tuning, closed-loop self-calibration significantly reduces labor costs, improves model response speed, and provides key support for continuous iteration and large-scale application. Attached Figure Description

[0112] Figure 1 This is a schematic diagram of the basic process of the present invention. Detailed Implementation

[0113] 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.

[0114] An example, a method for constructing customer profiles based on big data collection, refers to... Figure 1 ,include:

[0115] Collect multi-source behavioral and transaction data containing unique customer identifiers to generate the original dataset D. clicks and D tx ;

[0116] Based on the original dataset, numerical features related to customer behavior are extracted, and a customer feature matrix X is constructed.

[0117] The customer feature matrix is ​​processed for missing values ​​and normalized to output a standardized matrix X' = [x' ij ];

[0118] The significance score s for each feature is calculated based on the normalized matrix. j Generate feature saliency vector

[0119] Based on the standardized features and the aforementioned saliency vector, calculate the scalar value p for each customer's profile. i ;

[0120] Based on the customer profile scalar and the nearest neighbor distance set calculated from it, a density-based clustering algorithm is used to segment customers, generating a customer segmentation label set {C}. i};

[0121] Based on the aforementioned customer profile scalar and group tag set, personalized marketing tiered management is implemented, including resource allocation and content delivery strategy formulation, and the grouping parameters are adaptively adjusted based on changes in customer conversion rates after grouping.

[0122] The data collection includes multi-source behavioral and transaction data containing unique customer identifiers, generating an original dataset D. clicks and D tx ,include:

[0123] Request permission to collect customer website click logs and CRM transaction records;

[0124] Set the time interval for collecting click logs from the customer's website, and denot it as the log collection interval;

[0125] Set the daily collection time for customer CRM transaction records, and record it as the record collection node;

[0126] For the customer's website click logs, export the fields {customer_id, page_url, event_time} from the web server at every log collection interval;

[0127] For customer CRM transaction records, at each record collection node, export a standard table containing the fields {customer_id, transaction_amount, transaction_time, channel} from the CRM system;

[0128] Create the original record set D of the click log clicks and the original record set D of transaction records tx ;

[0129] Write the exported website click logs to D clicks In the format, each record is formatted as follows:

[0130] d i =(customer_id,page_url,event_time)

[0131] Write the exported CRM transaction records to D tx In the format, each record is formatted as follows:

[0132] d i =(customer_id,transaction_amount,transaction_time,channel)

[0133] Where d i Indicates that they are respectively in D clicks and D tx This solution collects website click logs and CRM transaction records from customer i. It focuses solely on these two core data sources, significantly reducing system integration complexity. Compared to traditional approaches that require access to multiple platforms and data formats, this solution only requires maintaining two stable data retrieval scripts with clearly defined field formats, reducing engineering integration and maintenance costs. Furthermore, the fixed-frequency batch retrieval mechanism ensures data synchronization within a controllable time window, meeting near real-time data requirements while avoiding the high concurrency and system pressure associated with streaming processing. This design balances system stability and flexibility, providing a solid data foundation for subsequent feature construction and model applications.

[0134] The step of extracting numerical features related to customer behavior based on the original dataset and constructing a customer feature matrix X includes:

[0135] From D clicks and D tx In this context, the following basic characteristics are constructed for each customer i:

[0136] f i1 Website visits in the past 7 days;

[0137] f i2 Total transaction amount over the past 30 days;

[0138] f i3 : Number of days since the last visit;

[0139] f i4 : The number of days since the last transaction;

[0140] f i5: Main transaction channel codes, including online=1, offline=2;

[0141] Arrange all customers into a matrix:

[0142] X = [f ij ] n×5

[0143] Each row of the matrix represents a customer, and the i-th row is the original feature vector of customer i, which is f in sequence. i1 f i2 f i3 f i4 f i5 Based on the original data, this solution extracts only five key features directly related to marketing management, such as recent visit frequency, total transaction amount, and behavior time interval, avoiding computational redundancy caused by "feature dimension inflation." Through clear feature definitions and a fixed dimension order, features can be reused across different projects or business scenarios, supporting the rapid deployment of new marketing campaigns. This method can also flexibly add or remove features according to actual needs, effectively balancing feature richness and computational efficiency. More importantly, outputting multi-source heterogeneous data through the same matrix greatly simplifies subsequent algorithm integration interfaces, making engineering implementation more standardized and easier to test. This step ensures that key business features are fully preserved while avoiding interference from irrelevant features, thereby improving the training speed and performance stability of subsequent models.

[0144] The step of performing missing value processing and normalization on the customer feature matrix to output a standardized matrix X' includes:

[0145] Obtain the number of missing values ​​n in the j-th column of the customer feature matrix X. miss,j The absence indicates the presence of a null value, null, or a special placeholder in the j-th column;

[0146] Calculate the missing percentage for the j-th column of the customer feature matrix X:

[0147]

[0148] Where n is the number of rows in the customer feature matrix X;

[0149] Set a threshold τ to determine whether there are too many missing values;

[0150] If r j <τ, then the median m of the non-empty values ​​in the j-th column. j Fill missing information;

[0151] Among them, the median m j This refers to the (n / 2)th data point after sorting the non-empty values ​​in column j in ascending order;

[0152] Otherwise, if column j is deemed to have too many missing values, the data in column j is deleted. This solution employs a "median-filled or entire column removed" strategy, achieving accurate completion and cleaning through dynamic judgment of the missing value ratio of each column. When the missing value ratio is below a threshold, the column is filled with the true median, which preserves most of the true data distribution and prevents bias introduced by mean filling. When the missing value ratio is too high, the entire column is directly deleted, avoiding excessive noise for subsequent analysis. Compared with the traditional method of deleting or filling with the mean, this method can effectively resist the influence of abnormal data while ensuring the integrity of the model input. In addition, the median calculation rule does not depend on the distribution assumption, making it applicable to any non-normal distribution scenario, further enhancing the versatility and stability of the solution in multi-industry applications.

[0153] The step of performing missing value processing and normalization operations on the customer feature matrix to output a standardized matrix X' further includes:

[0154] Calculate for each column after missing data filtering:

[0155]

[0156] Where, μ j Let σ represent the mean of the j-th feature. j This represents the standard deviation of the j-th feature;

[0157] Perform normalization transformation:

[0158]

[0159] Where f ij Derived from the customer feature matrix X, it represents the j-th feature of the i-th record;

[0160] Output the normalized matrix X' = [x' ij By performing standard deviation normalization on the cleaned matrix, all features can be unified to a range of zero mean and unit variance, eliminating weight shifts caused by differences in feature dimensions and value ranges. Compared to methods sensitive to outliers, such as min-max scaling, this normalization method can handle extreme data points more robustly. This step ensures that subsequent calculations of feature significance or distance metrics will not be dominated or distorted by a single feature value that is too large or too small. At the same time, standardized data is more likely to meet the statistical and algorithmic assumptions about the input distribution, improving the algorithm's convergence speed and numerical stability.

[0161] The significance score of each feature is calculated based on the standardized matrix, generating a feature significance vector. include:

[0162] For each column of the normalized matrix X':

[0163] Calculate the probability assignment matrix p ij :

[0164]

[0165] Calculate information entropy e j :

[0166]

[0167] Among them, e j ∈[0.1], representing the information entropy of the j-th feature; measures the degree of dispersion of the feature distribution;

[0168] Assign significance scores:

[0169]

[0170] Where m represents the number of features in the customer feature matrix, and its value is 5; e k s represents the information entropy of the k-th feature among all features; j ∈(0,1); This results in low weights for high-entropy features (uniform distribution) and high weights for low-entropy features (concentrated distribution);

[0171] The significance score of each column of the standardized matrix X' is calculated, and the scores are summed to obtain the significance vector. This solution automatically assigns a significance score to each feature based on information entropy, without relying on empirical coefficients or external training data, and measures its information value from the perspective of the feature's own distribution. Unlike traditional expert scoring or PCA dimensionality reduction, the entropy weight method quantifies the dispersion and information contribution of features on the global sample, thereby dynamically adjusting feature weights. This mechanism not only reduces human subjective bias but also improves model interpretability, clearly indicating which features have the greatest impact on customer profiles. At the same time, this method can adapt to fluctuations in sample size and the dynamic addition of new features, and has excellent online adaptive capabilities.

[0172] The calculation of the scalar value for each customer's profile based on the standardized features and the saliency vector includes:

[0173] By weighting the normalized features using significance scores, a scalar profile of each customer is obtained;

[0174]

[0175] Where m represents the number of features in the customer feature matrix, with a value of 5; s j x' is the significance score of the j-th feature; ijThis is the standardized value of the i-th customer on the j-th feature, normalized. By weighting and summing the standardized features with the saliency vector, this solution generates a single, comparable scalar value for the profile, significantly simplifying the cost of using multi-dimensional profile information. This scalar has both behavioral activity and consumption value meanings, and can be directly used for customer ranking and priority determination. Compared to the complex method of requiring multiple queries and combining multiple tags, the profile scalar is clear at a glance and easy to use directly in marketing management systems, BI reports, and decision-making processes, improving business execution efficiency. In addition, customer groups can be segmented by threshold or level in different business scenarios or advertising channels, enabling highly flexible strategy formulation.

[0176] The customer profile scalar, combined with the nearest neighbor distance set calculated from it, employs a density-based clustering algorithm to group customers and generate a customer grouping label set, including:

[0177] Let k be used for neighbor distance statistics, and for each customer profile point p i Calculate the distance d between its k-th nearest neighbor. i,(k) , forming a set n is the total number of customers; specifically: for each customer profile point p i Calculate its Euclidean distance to all other customers, and select the k-th customer with the closest distance, i.e., d. i,(k) ;

[0178] Construct an adaptive threshold ε:

[0179]

[0180] in, This is the distance from the median. α is the interquartile range; α is the initial empirical coefficient used to adjust the distance threshold in density clustering; α controls the sensitivity of ε, which essentially affects "which customers are considered neighbors". When α is set larger, ε is larger, making the clustering more "broad" and helping to accommodate more customers. When α is set smaller, ε decreases accordingly, making the clustering more "refined" and making it easier to identify outliers or fine-grained groups.

[0181] The distances to the median and interquartile range are specifically as follows:

[0182] Will Sort the elements in ascending order, and take the n / 2th element after sorting. Use this element as the median distance.

[0183] Will After sorting in ascending order, select the 25%n element as Q1 and the 75%n element as Q3. By dynamically calculating the cluster radius based on the median and interquartile range of the k-th nearest neighbor distance set in the sample, combined with empirical coefficients, no manual parameter tuning is required. Compared with the traditional DBSCAN mode that requires a fixed threshold, the dynamic threshold makes the clustering process more in line with the actual data distribution and can automatically adapt to changes in sample density, avoiding over-clustering or splitting caused by improper parameter settings in a single instance. This method can also resist clustering failure caused by local density imbalance, making the clustering results more robust. At the same time, the threshold is calculated based on statistical distribution and is not sensitive to extreme values, making it suitable for large-scale online business environments.

[0184] Density clustering technology is used, and DBSCAN({p i}, ε, minPts=k), generate customer segmentation label set {C i}

[0185] The process involves implementing personalized marketing tiered management based on the customer profile scalar and segmentation tag set, including resource allocation and content delivery strategy formulation, and adaptively adjusting segmentation parameters based on changes in customer conversion rates after segmentation, including:

[0186] All customers are categorized by their label C i Perform clustering, with each customer segmentation tag set forming a cluster;

[0187] For each cluster, calculate the average image value:

[0188]

[0189] Where |c| represents the number of customers within cluster c; ∑ i∈c p i This represents the sum of profiles for all customers within cluster c;

[0190] Allocate the promotion budget proportionally:

[0191]

[0192] Among them, B c B represents the marketing budget allocated to customers within cluster c; total This indicates the total marketing budget available during the current marketing cycle; This is the sum of the average profile values ​​across all clusters. By using the average profile value of customer groups as the basis for budget allocation, marketing investment becomes more precise and ROI is improved. Unlike traditional strategies based on historical experience or uniform discounts, this method ensures that high-value groups receive more resource support, improving their conversion efficiency. At the same time, low-value groups can receive targeted advertising, saving costs. This budget allocation formula is simple to implement and can be seamlessly integrated with the company's existing financial system and advertising platform.

[0193] The process of implementing personalized marketing tiered management based on the customer profile scalar and segmentation tag set, including resource allocation and content delivery strategy formulation, and adaptively adjusting segmentation parameters based on changes in customer conversion rates after segmentation, also includes:

[0194] Set conversion success behavioral goals; these goals include customer clicks, orders, and shares.

[0195] After each round of marketing campaigns, the total number of customers in cluster c who exhibited the desired behavior is recorded as the number of successfully converted customers, V.

[0196] Calculate the conversion rate A of cluster c. (t) :

[0197]

[0198] Where |c| represents the number of customers within cluster c;

[0199] Calculate the difference in conversion rate ΔA between two adjacent rounds of marketing campaigns. c :

[0200]

[0201] in, This indicates the conversion rate after this round of marketing campaign. This indicates the conversion rate after the previous round of marketing campaigns;

[0202] Set a threshold δ to determine whether to update the clustering threshold coefficient α; the smaller the threshold δ is set, the more sensitive the system is to transformation changes and the more frequently it is updated; the larger the threshold δ is set, the more the system is updated only when there are drastic changes, the more stable the system is but the slower the response.

[0203] If |ΔA|≤δ, then keep the original clustering threshold coefficient α unchanged;

[0204] When |ΔA|>δ, then update the clustering threshold coefficient α:

[0205] α (t+1) =α t ×(1+βsign(ΔA))

[0206] Where, α t α represents the clustering threshold coefficient used in this round of marketing; β represents the step size factor, which controls the magnitude of each update of α; sign(ΔA) is the sign function, which increases the threshold with a positive sign and decreases the threshold with a negative sign; the smaller the step size factor is set, the smaller the magnitude of each update, making the update smoother and the change process more stable. The larger the step size factor is set, the faster each update changes, and the faster it adapts to new situations, but it is more prone to instability.

[0207] With α (t+1) Reconstruct the adaptive threshold ε:

[0208]

[0209] ε t+1 This solution is used for customer profiling and clustering in the next round of marketing. By monitoring the conversion rate changes of each group after each round of marketing, it automatically adjusts the density threshold coefficient and recalculates the cluster radius, achieving closed-loop self-adaptation of profiling and clustering parameters. This mechanism allows customer segmentation and profiling to be continuously optimized based on market feedback, avoiding the decrease in accuracy caused by model obsolescence. Compared with manual periodic review and parameter tuning, closed-loop self-calibration significantly reduces labor costs, improves model response speed, and provides key support for continuous iteration and large-scale application.

[0210] 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.

[0211] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing customer profiles based on big data collection, characterized in that, include: Collect multi-source behavioral and transaction data containing unique customer identifiers to generate a raw dataset. and ; Based on the original dataset, numerical features related to customer behavior are extracted to construct a customer feature matrix. ; The customer feature matrix is ​​processed for missing values ​​and normalized to output a standardized matrix. ; The significance score for each feature is calculated based on the standardized matrix. Generate feature saliency vector ; Based on the standardized features and the aforementioned saliency vector, calculate the scalar value of each customer's profile. ; Based on the customer profile scalar and the nearest neighbor distance set calculated from it, a density-based clustering algorithm is used to segment customers and generate a set of customer segmentation labels. ; Based on the aforementioned customer profile scalar and group tag set, personalized marketing tiered management is implemented, including resource allocation and content delivery strategy formulation, and the grouping parameters are adaptively adjusted based on changes in customer conversion rates after grouping. The collection includes multi-source behavioral data and transaction data containing unique customer identifiers, generating a raw data set. and ,include: Request permission to collect customer website click logs and CRM transaction records; Set the time interval for collecting click logs from the customer's website, and denot it as the log collection interval; Set the daily collection time for customer CRM transaction records, and record it as the record collection node; For the client's website click logs, export the data containing the specified fields from the web server at regular log collection intervals. ; For customer CRM transaction records, export records containing fields from the CRM system at each record collection node. Standard table; Create a raw record set of click logs and the original record set of transaction records ; Write the exported website logs In the format, each record is as follows: Write the exported CRM transaction records to In the format, each record is as follows: in Indicates respectively in and Customers in Website click logs and CRM transaction records; Specifically, the step of calculating the significance score of each feature based on the standardized matrix to generate a feature significance vector is described. ,include: For the standardized matrix Each column: Calculate the probability assignment matrix : Calculate information entropy : in, , indicating the first Information entropy of features; Assign significance scores: in, This indicates the number of features in the customer feature matrix, with a value of 5. Indicates the first of all features Information entropy of each feature; ; The normalized matrix is ​​calculated. The significance scores of each column are summed to obtain a significance vector. ; Specifically, based on the customer profile scalar and combined with the nearest neighbor distance set calculated therefrom, a density-based clustering algorithm is used to group customers and generate a customer grouping label set, including: Configure for neighbor statistics For each customer profile point Calculate its first Nearest neighbor distance , forming a set , The total number of customers; Constructing an adaptive threshold : in, This is the distance from the median. Interquartile range; These are initial empirical coefficients used to adjust the distance threshold in density clustering; The distances to the median and interquartile range are specifically as follows: Will Sort by size in ascending order, and then take the first sorted item. Each element is used as the median distance; Will After sorting in ascending order, select the first in the sorted order. each element as Select the first each element as ,but ; Using density clustering technology, call Generate a set of customer segmentation tags .

2. The customer profile construction method based on big data collection according to claim 1, characterized in that, Based on the original dataset, numerical features related to customer behavior are extracted to construct a customer feature matrix. ,include: from and In the middle, for each customer The following basic features are constructed: Website visits in the past 7 days; Total transaction amount over the past 30 days; : Number of days since the last visit; : The number of days since the last transaction; : Main transaction channel codes, including online=1, offline=2; Arrange all customers into a matrix: Each row of the matrix represents a customer, and the... Behavioral Customers The original feature vectors are, in order: , , , , .

3. The customer profile construction method based on big data collection according to claim 1, characterized in that, The customer feature matrix is ​​then processed for missing values ​​and normalized to output a standardized matrix. ,include: Obtain customer feature matrix The first in Number of missing items in the column The missing part indicates the first... The column contains null values, null, or special placeholders; Customer feature matrix The first in Calculate the missing percentage for each column: in Customer feature matrix number of rows; Set a threshold to determine if there are too many missing values. ; like Then take the first Median of non-null values ​​in a column Fill missing information; Among them, the median To make the first After sorting the non-null values ​​in ascending order, the first... One data point; Otherwise, determine the first Too many columns are missing, for the first Delete the data in the column.

4. The customer profile construction method based on big data collection according to claim 3, characterized in that, The customer feature matrix is ​​then processed for missing values ​​and normalized to output a standardized matrix. It also includes: Calculate for each column after missing data filtering: in, Indicates the first The mean of each feature, Indicates the first Standard deviation of each feature; Perform normalization transformation: in Derived from customer feature matrix , indicating the first Record No. One feature; Output the normalized matrix .

5. The customer profile construction method based on big data collection according to claim 1, characterized in that, The calculation of the scalar value for each customer's profile based on the standardized features and the saliency vector includes: By weighting the normalized features using significance scores, a scalar profile of each customer is obtained; in, This indicates the number of features in the customer feature matrix, with a value of 5. For the first The significance scores of each feature; For the first The customer in the first The standardized values ​​after normalization on each feature.

6. The customer profile construction method based on big data collection according to claim 5, characterized in that, The process involves implementing personalized marketing tiered management based on the customer profile scalar and segmentation tag set, including resource allocation and content delivery strategy formulation, and adaptively adjusting segmentation parameters based on changes in customer conversion rates after segmentation, including: All customers by their tags Perform clustering, with each customer segmentation tag set forming a cluster; For each cluster, calculate the average image value: in, Cluster Number of customers within; Cluster The sum of profiles of all customers within the company; Allocate the promotion budget proportionally: in, Indicates assignment to cluster Marketing budgets for internal clients; This indicates the total marketing budget available during the current marketing cycle; The sum of the average values ​​of the images for all clusters.

7. The customer profile construction method based on big data collection according to claim 6, characterized in that, The process of implementing personalized marketing tiered management based on the customer profile scalar and segmentation tag set, including resource allocation and content delivery strategy formulation, and adaptively adjusting segmentation parameters based on changes in customer conversion rates after segmentation, also includes: Set behavioral goals for successful conversion; After each round of marketing campaigns, record clusters The total number of customers who engage in the desired behavior is considered the number of successfully converted customers. ; Computational clusters conversion rate : in, Cluster Number of customers within; Calculate the difference in conversion rates between two adjacent rounds of marketing campaigns. : in, This indicates the conversion rate after this round of marketing campaign. This indicates the conversion rate after the previous round of marketing campaigns; Configure the settings for determining whether to update the clustering threshold coefficient. threshold ; like Then the original clustering threshold coefficient is maintained. constant; when Then update the clustering threshold coefficient. : in, This represents the clustering threshold coefficient used in this round of marketing; Represents the step size factor, controlling... The magnitude of each update; The sign function is positive, which increases the threshold, and negative, which decreases the threshold. by Reconstruct adaptive threshold : Will Used for customer profiling and clustering in the next round of marketing.

Images