Novel customer relationship management method

Abstract

The invention provides a novel customer relationship management method, and belongs to the technical field of artificial intelligence and application. Establishing a data base which is used for gathering and integrating multi-source data and consists of client omni-channel interaction data and enterprise business data; based on feature engineering and RFM, the FCI model defines customer feature portraits including behavior preferences, management cycle stages and marketing response laws based on a data base. And taking the customer feature portrait as an input parameter of the random forest model, training the random forest model in combination with historical data, and outputting a marketing strategy including a recommendation category, a communication channel, an interaction frequency, a discount threshold value and a docking person level. According to the invention, based on the multi-dimensional marketing strategy output by the random forest model, the execution link can be directly connected, and manual secondary adjustment is not needed; and meanwhile, the model can quickly adapt to different customer group characteristics, so that'homogenization 'and'hysteresis' of traditional manual strategy formulation are avoided, and manual dependence and management cost are reduced.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and application technology, and in particular relates to a novel customer relationship management method. Background Technology

[0002] Customer relationship refers to the dynamic connection between a company and its customers formed through long-term interaction based on value exchange, emotional connection, and demand fulfillment. The core is that the company builds a relationship with its customers from "strangeness" to "trust" and then to "loyalty" through continuous service, communication, and value delivery, ultimately achieving a win-win situation of "long-term customer retention" and "corporate value growth".

[0003] Marketing strategy refers to the systematic action plan and execution logic formulated by enterprises based on market analysis and customer needs insights to achieve business goals such as "customer acquisition, conversion, retention and value enhancement". The core is to accurately locate target customers, design appropriate value delivery methods, efficiently reach customers and stimulate their consumption or interaction behavior, and continuously optimize the plan based on the effect feedback.

[0004] In existing technologies, customer interaction data such as social communication and service feedback, which are relied upon by customer relationship management, and enterprise business data such as product inventory and marketing campaign history, which are required for marketing strategy formulation, are usually stored in different systems, forming data silos. This makes it difficult to effectively explore the correlation between customer behavior characteristics and marketing resources and business rules, and fails to provide comprehensive data support for strategy formulation.

[0005] In the traditional model, customer relationship management, such as customer segmentation and needs identification, and marketing strategy development, such as marketing plan generation and execution optimization, are mostly independent processes that rely on manual information transmission, resulting in process breaks. For example, when the customer lifecycle stage changes, the marketing strategy cannot be adjusted in a timely manner, leading to a lag in matching customer needs with marketing actions.

[0006] In existing technologies, the formulation of marketing strategies relies heavily on the historical experience of operations personnel, lacking automated and in-depth analysis of customer behavior preferences and marketing response patterns. This leads to strategy generalization and difficulty in achieving personalized adaptation. For example, it is difficult to cater to different needs by using homogeneous marketing for high-value customers and ordinary customers. Under the manual decision-making model, the strategy iteration cycle is long and cannot quickly respond to dynamic changes in the market environment, fluctuations in customer demand, and other business scenarios.

[0007] In key scenarios such as personalized operations for high-value customers, targeted recovery of lost customers, and cross-selling across product categories, existing technologies often suffer from problems such as misallocation of marketing resources, delayed recovery opportunities, and blind cross-selling recommendations due to the lack of linkage analysis between customer value evolution patterns and marketing response characteristics. This results in low utilization of marketing resources and poor achievement of core business goals. Summary of the Invention

[0008] To address the above problems, this invention proposes a novel customer relationship management method, namely, an integrated management method for customer relationships and marketing strategies based on artificial intelligence, comprising the following processes:

[0009] S1 establishes a data foundation composed of customer omnichannel interaction data and enterprise business data. The customer omnichannel interaction data includes communication content, consumption records, and service feedback; the enterprise business data includes marketing campaign history and transaction data.

[0010] S2, based on a data foundation, constructs customer profiles including behavioral preferences, management cycle stages, and marketing response patterns:

[0011] By combining consumption records, transaction data, marketing campaign history, and communication content, customer behavioral preference characteristics are defined based on feature engineering methods.

[0012] Based on communication content and consumption records, management cycle stages are defined using feature engineering, RFM, and FCI methods.

[0013] By combining communication content, marketing campaign history, and transaction data, marketing response patterns are defined based on feature engineering methods.

[0014] S3 uses customer profiles as input parameters for the random forest model, combines historical data to train the random forest model, and outputs marketing strategies including recommended product categories, communication channels, interaction frequency, discount thresholds, and contact person job levels.

[0015] Preferably, the customer omnichannel interaction data specifically includes: communication content between customers and enterprises, which refers to the original communication content data group composed of communication time, communication channel, communication topic, and interactive text or voice recordings; consumption records, which refers to the original consumption record data group composed of consumption time, consumed product ID, consumption amount, and payment method; and service feedback data, which specifically consists of the original service feedback data group composed of feedback time, feedback type, feedback content, and satisfaction rating.

[0016] The enterprise business data includes marketing campaign historical data, which refers to the original marketing campaign data set consisting of campaign ID, campaign theme, campaign time, campaign rules, and participating customer scope; and transaction data, which refers to the original transaction data set consisting of transaction order number, transaction amount, transaction status, transaction completion time, and refund record.

[0017] Preferably, the data in S1 is preprocessed, including performing duplicate record removal operations, defining and performing operations on rules for missing or incomplete values ​​of key fields, and defining and performing filtering operations on rules for identifying abnormal data.

[0018] Simultaneously, the preprocessed data undergoes data format conversion; including:

[0019] Convert all amounts to yuan;

[0020] Align the time in the format "year-month-day-hour-minute-second";

[0021] The customer's digital ID is a unique 12-digit code consisting of the "year-month-day" plus the communication sequence number of the day from the first communication record.

[0022] Preferably, in S2, customer behavior preference features are defined based on feature engineering methods by combining consumption records, transaction data, marketing activity history, and communication content. Specifically, these features refer to the 1st, 2nd, and 3rd categories of total purchase value and their corresponding discount thresholds, as well as the 1st communication channel in terms of usage time.

[0023] Combining communication content and consumption records, management cycle stages are defined based on feature engineering, RFM, and FCI methods, including: defining customer stage tags using feature engineering methods based on communication content and consumption records, including potential customer stage, new customer stage, and old customer stage; defining high-value customers using RFM methods based on consumption records as input, i.e., customers whose total transaction amount ranks in the top 30% of all customers; and defining customers at risk of churn using FCI methods based on communication content as input, i.e., customers whose inactivity period exceeds the average communication interval of the previous year.

[0024] By combining communication content, marketing campaign history, and transaction data, marketing response patterns are defined based on feature engineering methods. These patterns include the job level of the communication contact person and the response cycle, as well as the time interval between the transaction product and its initial release.

[0025] Preferably, the process of defining customer behavior preference characteristics is as follows:

[0026] 1) Using "Customer Digital ID" as the grouping dimension, further subdivide each customer's consumption records by "Product Category ID" to calculate the total purchase value of each customer in each category, i.e., the total transaction amount; sort the total purchase value in descending order from high to low, and sum them as the total purchase value; extract the categories ranked 1st, 2nd, and 3rd, and label them as category 1, category 2, and category 3 of the customer's total purchase value, respectively; if the customer consumes less than 3 categories, mark the remaining unmet ranking positions as none; for categories 1, 2, and 3 of the customer's total purchase value, backtrack and match the discount data that the customer has enjoyed in all historical transactions under this category, and calculate the average value as the discount threshold corresponding to category 1, 2, and 3;

[0027] 2) Extract records of interactions between target customers via telephone, online customer service, email, and WeChat from the customer communication logs of the data base, calculate the cumulative usage time, and mark the communication channel ranked first as the customer's preferred communication channel;

[0028] 3) Establish a customer behavior preference feature data group based on customer digital ID, including the total value of purchased goods; the first, second, and third ranked categories and their corresponding discount thresholds; and preferred communication channels.

[0029] Preferably, the process of defining the management cycle stage is as follows:

[0030] 1) Extract communication records from the data base using customer digital IDs. Collect all communication records within the past year. Customers with zero purchases in their transaction records are labeled as "potential customers". Customers with exactly one purchase are labeled as "new customers". Customers with more than one purchase are labeled as "existing customers".

[0031] 2) Calculate the total amount of goods purchased by all customers in the past year using the RFM model and sort them. Mark the top 30% of customers in the sorting results as "high-value customers".

[0032] 3) Using the interval analysis logic of the FCI model, statistically analyze all communication records of all customers in the past year and calculate the time interval between two consecutive communications to obtain the average communication interval of the previous year; calculate the time from the date of the customer's most recent communication to the current date based on the current date to define the customer's silence period; compare the customer's silence period with the average communication interval of the previous year. If the customer's silence period is greater than the average communication interval of the previous year, mark it as a "customer at risk of churn".

[0033] 4) Establish a customer management cycle stage data group based on the customer's digital ID, which includes customer stage tags and customer status tags. If the customer has no status tag, mark it as none.

[0034] Preferably, the process of defining the marketing response pattern is as follows:

[0035] 1) Extract communication records from the data base using the customer's digital ID, and statistically analyze all communication records within the past year. Extract the initiation time, contact person's job level, and response time for each communication record. Calculate the time difference between the communication initiation time and the final response time, and define it as a single response cycle. Calculate the average of all single response cycles by classifying the contact person as junior customer service, senior consultant, or sales director, and use this as the response cycle indicator for the contact person's job level corresponding to the user's digital ID.

[0036] 2) Extract transaction records from the data base using the customer's digital ID, and statistically analyze all transaction records within the past year. Extract the first release time of the corresponding product ID from each transaction record. For each transaction, calculate the interval between the transaction completion time and the first release time of the product, which is defined as the single new product response cycle. Calculate the average of all single new product response cycles based on the total value of the product categories purchased by the customer's digital ID in categories 1, 2, and 3, and use this as the new product response cycle indicator corresponding to the user's digital ID.

[0037] 3) Establish a marketing response pattern data group based on customer digital IDs, including the response cycle of the contact person's job level and the new product response cycle.

[0038] Preferably, the step S3, which combines historical data to train a random forest model, specifically includes:

[0039] Random forest model parameter configuration, including: assigning a value of 100 to the number of decision trees; assigning a value of 10 to the maximum tree depth; assigning a value of 5 to the minimum number of split samples; assigning a value of 5 to the minimum number of leaf node samples; and selecting the Gini coefficient as the feature selection method.

[0040] Input the input features and output labels of the training set into the random forest model to generate the model's prediction results;

[0041] Input the features from the test set into the trained model to obtain the predicted marketing strategy labels, compare them with the true labels from the test set, and calculate the core evaluation metrics:

[0042] Customer conversion rate metric indicates the percentage of customers who make their first purchase as a result of the prediction strategy;

[0043] The strategy response rate metric represents the customer's click / engagement rate with the predictive strategy;

[0044] The customer repurchase rate metric represents the percentage of customers who repurchase within 30 days after the prediction strategy is implemented.

[0045] If the average of the three indicators is ≥80%, the model is considered to have met the performance standards; if it does not meet the standards, the model parameters are adjusted and the model is retrained until it meets the standards.

[0046] Compared with the prior art, the present invention has the following innovative features:

[0047] (1) Constructing a data foundation for the integration of customer omnichannel interaction and enterprise business data: Customer interaction data and enterprise business data are stored separately in different systems, resulting in a lack of complete information support for customer analysis. By aggregating and integrating multi-source data, the two types of data are uniformly incorporated into the data foundation to achieve a deep correlation between customer "behavior-business" data, providing a full-dimensional data foundation for subsequent customer feature mining and strategy formulation.

[0048] (2) Integrating feature engineering with RFM and FCI models to construct multi-dimensional customer feature profiles: Combining feature engineering technology, based on RFM and FCI models, define customer behavior preferences such as category preferences and communication habits, management cycle stages such as potential or new customers or loyal customers, and marketing response patterns such as discount threshold sensitivity, to construct multi-dimensional and three-dimensional customer feature profiles.

[0049] (3) Automated mapping of customer characteristics to marketing strategies based on random forest model: Multidimensional customer characteristic profiles are used as input to random forest model. The model is trained by combining historical data and automatically outputs full-dimensional marketing strategies including recommended categories, communication channels, interaction frequency, discount thresholds and contact person job level, so as to realize the integration and automated association between customer characteristics and marketing strategies.

[0050] The beneficial effects brought about by the innovation of this invention include:

[0051] (1) Improve the accuracy of customer demand identification: By integrating customer behavior and enterprise business data through a unified data base, and combining multi-dimensional feature profiles to mine potential customer needs, the one-sided judgment of "only looking at behavior and not business" or "only looking at business and not behavior" under the traditional data silos is avoided, and the error rate of customer demand identification is significantly reduced.

[0052] (2) Improve the personalization and implementation efficiency of marketing strategies: The multi-dimensional marketing strategies output by the random forest model can be directly connected to the execution stage without the need for manual secondary adjustments; at the same time, the model can quickly adapt to the characteristics of different customer groups, avoiding the "homogenization" and "lag" of traditional manual strategy formulation, thus shortening the strategy implementation cycle and improving customer response speed.

[0053] (3) Reduce reliance on manual labor and management costs: By automatically integrating data, automatically mining features, and automatically generating strategies, the cumbersome process of "manually summarizing data, manually classifying customer levels, and manually formulating strategies" in traditional customer management is replaced, reducing human resource input; at the same time, subjective biases in manual operation are avoided, reducing the waste of marketing resources caused by strategy errors, and indirectly reducing the cost of enterprise customer relationship management and marketing operations.

[0054] (4) Wide range of applicable scenarios: Based on the integrated logic of "data base + feature model + strategy model", it can quickly adapt to the needs of the scenario, rather than optimizing parameters for a single scenario, and has a wide range of applications. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the overall technical process of the present invention.

[0056] Figure 2 This is a diagram showing the total transaction amount for each product category.

[0057] Figure 3 This is a schematic diagram of the random forest model.

[0058] Figure 4 The diagram shows the experimental process for setting parameters for the random forest model.

[0059] Figure 5 The graph shows the marketing strategy recommendation results of the random forest model. Detailed Implementation

[0060] This invention proposes an integrated management method for customer relationship and marketing strategies based on artificial intelligence. The overall process is as follows: Figure 1 As shown:

[0061] S1 establishes a data foundation that integrates multi-source data, comprising customer omnichannel interaction data and enterprise business data. Through the deployment of a data interface adaptation layer, it connects customer omnichannel interaction touchpoints with internal enterprise business systems, extracting in real-time customer interaction data (comprising communication content, consumption records, and service feedback) and enterprise business data (comprising marketing campaign history and transaction data). Data preprocessing includes removing duplicate records, filling in missing values ​​in key fields, and filtering out abnormal data. The preprocessed data undergoes standardization transformation, including converting amounts to yuan, formatting time as "year-month-day-hour-minute-second," and tagging customers with numeric IDs, achieving unified formatting for heterogeneous data. A structured array is then created based on the customer's numeric ID, where customer omnichannel interaction data includes communication content, consumption records, and service feedback; and enterprise business data includes marketing campaign history and transaction data.

[0062] S2, based on customer interaction and business data in the data foundation, uses feature engineering techniques for targeted feature extraction and calculation: It calculates the total purchase value of each customer across different product categories from consumption records and sorts them in descending order to determine the first, second, and third categories of total purchase value. Simultaneously, it matches the historical transaction discount levels for each category in the marketing campaign's historical data to generate the discount threshold for that category. It also calculates the cumulative time customers spend using each communication channel from communication content, selecting the communication channel with the longest usage time to complete the construction of behavioral preference features. Combining the transaction amount dimension of the RFM model and the interval analysis logic of the FCI model, it divides the management cycle into stages: customers with interaction in communication content but no purchase records are marked as potential customers, those with one purchase record are marked as new customers, and those with more than one purchase record are marked as returning customers. It calculates the total transaction amount of all customers and sorts them in descending order, classifying the top 30% of customers as high-value customers. Finally, it calculates the average communication interval duration of customers in the previous year; if the current customer's inactivity period exceeds this average, they are marked as a customer at risk of churn. Extract the job level of the contact person and the corresponding response cycle for each interaction from the communication content; retrieve the first release time of the transaction product from the historical data of the marketing campaign; calculate the interval between the transaction time and the release time to form the marketing response pattern characteristics; and link and integrate the three types of characteristics—behavioral preferences, management cycle stages, and marketing response patterns—according to the customer's unique digital ID to finally construct a multi-dimensional and three-dimensional customer profile.

[0063] S3 uses the constructed multi-dimensional customer profile as input features for the random forest model. Simultaneously, it extracts "effective marketing strategy tags" for corresponding customer groups from historical business data—the strategy combinations that historically yielded the best customer conversion / retention results. The tags include the best-performing recommended product category, communication channel, interaction frequency, discount threshold, and contact person's job level corresponding to the customer's digital ID. Categorical variables in the input features, such as customer stage and communication channel type, are coded with labels. Discount thresholds and interaction frequencies are normalized to ensure data format compatibility for model training. The dataset is divided into training and testing sets in a 7:3 ratio. The random forest model is trained using the training set, while the model's performance is validated using the testing set. Customer conversion rate, response rate, and repurchase rate after marketing strategy implementation are the core evaluation metrics. When a new customer's multi-dimensional profile is input into the model, the model automatically outputs a suitable full-dimensional marketing strategy based on the "customer feature-marketing strategy" correlation learned during training. Ultimately, this achieves integrated matching of customer features and marketing strategies, automated output, and traceable decision-making.

[0064] The method of the present invention will be further described below with reference to specific embodiments.

[0065] I. Establish a data foundation composed of customer omnichannel interaction data and enterprise business data.

[0066] A multi-source data foundation is constructed, centered on customer omnichannel interaction data (communication content, consumption records, service feedback) and enterprise business data (product inventory, marketing campaign history, transaction data). Real-time extraction and integration are achieved through a data interface adaptation layer. After preprocessing such as deduplication, completion, and filtering, the data undergoes standardization of amounts, unification of time formats, and digital conversion of customer IDs. Finally, a structured array is established based on the customer's digital ID, enabling unified storage and management of heterogeneous data. The specific processing flow includes:

[0067] 1. Deploy the data interface adaptation layer and connect it to the system to extract raw data; the specific steps are as follows:

[0068] (1) Deploy a data interface adaptation layer. This layer supports three connection methods: API interface, SDK integration, and direct database connection, establishing a data transmission channel between customer omnichannel interaction touchpoints and the enterprise's internal business systems. Customer omnichannel interaction touchpoints include social platforms, e-commerce platforms, and customer service systems. The enterprise's internal business systems include marketing activity management systems and financial transaction systems.

[0069] (2) Through the above data interface adaptation layer, real-time stream processing technology is used to extract two types of raw data based on the message middleware Kafka. The first type is customer omnichannel interaction data, which specifically includes customer-enterprise communication content, referring to the raw communication content data group composed of communication time, communication channel, communication topic, and interactive text or voice records; consumption records, referring to the raw consumption record data group composed of consumption time, consumed product ID, consumption amount, and payment method; and service feedback data, specifically the raw service feedback data group composed of feedback time, feedback type, feedback content, and satisfaction rating. The second type is enterprise business data, including marketing activity historical data, referring to the raw marketing activity data group composed of activity ID, activity topic, activity time, activity rules, and participating customer scope; and transaction data, referring to the raw transaction data group composed of transaction order number, transaction amount, transaction status, transaction completion time, and refund records.

[0070] 2. Perform data preprocessing on the extracted raw data to ensure data quality; the specific steps are as follows:

[0071] (1) Perform duplicate record removal operation: compare the data within a single original data group and delete completely duplicate data records.

[0072] (2) Define rules for missing values ​​in key fields and perform operations: text fields such as communication topic and feedback type are marked with the default value "not specified"; time fields such as consumption time and transaction completion time are filled with data extraction time.

[0073] (3) Define abnormal data identification rules and perform filtering operations: Consumption time, activity time, and transaction completion time that do not conform to the 24-hour system are identified as abnormal data; negative satisfaction scores and transaction amounts are identified as abnormal data; and consumption product IDs, activity IDs, and transaction order numbers that do not conform to ID compilation requirements are identified as abnormal data. Delete all abnormal data identification data.

[0074] 3. Convert the data format of the data after preprocessing; the specific steps are as follows:

[0075] (1) Convert all amounts to yuan;

[0076] (2) Align the time in the format of "year-month-day-hour-minute-second";

[0077] (3) The customer digital ID is a unique 12-digit code consisting of “year (4 digits) - month (2 digits) - day (2 digits) + communication sequence number (4 digits) of the first communication record”.

[0078] 4. Build structured arrays based on customer digital IDs to link customer omnichannel interaction data and enterprise business data.

[0079] II. Targeted Feature Extraction and Calculation Using Feature Engineering Techniques

[0080] The system analyzes and ranks customer purchases by category based on purchase records, identifying the top three categories and their corresponding discount thresholds. It also uses communication logs to identify the channel with the longest usage time to construct behavioral preferences. Combining RFM and FCI models, the system divides management cycles (potential customers, new customers, and returning customers), classifying high-value customers based on the top 30% of transaction amounts and marking customers at risk of churn based on average communication intervals. The system extracts contact person's job level, response time, discount level in transaction records, and the initial release date of product data from communication records to calculate transaction intervals and establish marketing response patterns. Finally, it integrates these three characteristics—behavioral preferences, management cycle stage, and marketing response patterns—by linking them to customer digital IDs to construct a multi-dimensional customer profile.

[0081] 1. Define and establish a customer behavior preference feature data set; the specific steps are as follows:

[0082] (1) Using "customer digital ID" as the grouping dimension, further subdivide each customer's consumption records by "product category ID" to calculate the total purchase value of each customer in each category, i.e., the total transaction amount, such as Figure 2As shown. The total purchase value is sorted in descending order of value, and the sum is used as the total purchase value. The categories ranked 1st, 2nd, and 3rd are extracted and labeled as Category 1, Category 2, and Category 3 of the customer's total purchase value, respectively. If the customer has fewer than 3 categories, the remaining unmet ranking positions are marked as none. For Categories 1, 2, and 3 of the customer's total purchase value, the discount data of all historical transactions under these categories is back-matched, and the average value is used as the discount threshold for Categories 1, 2, and 3.

[0083] (2) Extract the records of target customers’ interactions via telephone, online customer service, email and WeChat from the customer communication logs of the data base, calculate the cumulative usage time, and mark the communication channel ranked first as the customer’s preferred communication channel.

[0084] (3) Establish a customer behavior preference feature data group based on customer digital ID, including the total value of purchased goods; the first, second and third ranked categories and their corresponding discount thresholds; and preferred communication channels.

[0085] 2. Define and combine the transaction amount dimension of the RFM model with the interval analysis logic of the FCI model to divide the management cycle into stages; the specific steps are as follows:

[0086] (1) Extract communication records from the data base using customer digital ID, count all communication records in the past year, and mark customers with zero purchases in the transaction records as "potential customers"; mark customers with exactly one cumulative purchase as "new customers"; mark customers with more than one cumulative purchase as "old customers".

[0087] (2) Calculate the total amount of goods purchased by all customers in the past year using the RFM model and sort them. Mark the top 30% of customers in the sorting results as "high-value customers".

[0088] (3) Using the interval analysis logic of the FCI model, statistically analyze all communication records of all customers in the past year and calculate the time interval between two consecutive communications to obtain the average communication interval of the previous year. Calculate the time from the customer's most recent communication date to the current date based on the current date, and define the customer's silence period. Compare the customer's silence period with the average communication interval of the previous year. If the customer's silence period is greater than the average communication interval of the previous year, mark it as a "customer at risk of churn".

[0089] (4) Establish a customer management cycle stage data group based on the customer's digital ID, which includes customer stage label and customer status label. If the customer has no status label, mark it as none.

[0090] 3. Define the marketing response pattern; the specific steps are as follows:

[0091] (1) Extract communication records from the data base using the customer's digital ID. Compile all communication records within the past year and extract the initiation time, contact person's job level, and response time for each communication record. Calculate the time difference between the communication initiation time and the final response time, defining it as a single response cycle. Calculate the average of all single response cycles categorized by contact person (junior customer service, senior consultant, sales director) as the response cycle indicator for the contact person's job level corresponding to the user's digital ID.

[0092] (2) Extract transaction records from the data base using the customer's digital ID, and statistically analyze all transaction records within the past year. Extract the initial launch time of the corresponding product ID from each transaction record. For each transaction, calculate the interval between the transaction completion time and the initial launch time of the product, which is defined as the single new product response cycle. Calculate the average of all single new product response cycles based on the total value of the product categories purchased by the customer's digital ID for categories 1, 2, and 3, and use this as the new product response cycle indicator corresponding to the user's digital ID.

[0093] (3) Establish a marketing response pattern data group based on customer digital ID, including the response cycle of the contact person's job level and the new product response cycle.

[0094] A multi-dimensional customer feature profile data group is constructed based on customer data ID, which is composed of customer behavior preference feature data group, customer management cycle stage data group, and marketing response pattern data group.

[0095] Third, the random forest model outputs a tailored, multi-dimensional marketing strategy based on customer profile data.

[0096] Customer profile data sets are used as input features and correlated with recommended product categories, communication channels, and discount thresholds extracted from historical business data to construct a random forest model. In the data preprocessing stage, categorical variables are labeled and coded, and the training and test sets are divided in a 7:3 ratio for model training and validation. The core evaluation metrics are conversion rate, response rate, and repurchase rate. The trained model can automatically output the optimal multi-dimensional marketing strategy based on the characteristics of new customers.

[0097] 1. Define the model's input features and output label dataset; the specific steps are as follows:

[0098] (1) Use the multi-dimensional customer feature profile data set as the input feature set for the random forest model. The principle diagram of the random forest model is shown below. Figure 3 As shown.

[0099] (2) Extract the model output tag set: 3 recommended product categories, communication channels, interaction frequency, discount threshold, and contact person's job level.

[0100] (3) Construct a “feature-tag” association dataset: Using the customer’s digital ID as the association key, establish a correspondence between the input features of step (1) and the output tags of step (2) to form a structured “customer feature-marketing strategy tag” dataset.

[0101] 1. Preprocess the input features to adapt them for model training; the specific steps are as follows:

[0102] (1) For the categorical variables in the input features, the label encoding method is used to convert them into numerical data; the specific encoding rules are as follows:

[0103] Total value of purchased items in categories 1, 2, and 3: Category 3 → 0, Category 2 → 1, Category 1 → 2.

[0104] Management cycle stages: Potential customers → 0, New customers → 1, Existing customers → 2, High-value customers → 3, Customers at risk of churn → 4;

[0105] Communication channel types: WeChat → 0, telephone customer service → 1, shopping platform → 2, SMS → 3;

[0106] Contact person's job level: Junior Customer Service → 0, Senior Consultant → 1, Sales Director → 2.

[0107] After encoding, a mapping table between categorical variables and numerical values ​​is generated, which is used for reverse decoding during model inference.

[0108] (2) For continuous variables in the input features, including discount threshold, customer response cycle based on contact person's job level, and new product response cycle, extract the discount threshold (denoted as A) within the multi-dimensional customer feature profile data group and create an array. The Min-Max normalization method is used to compress it to the [0,1] interval. The calculation formula is as follows:

[0109] In the formula express The minimum value in the array; express The maximum value in the array.

[0110] Similarly, the normalized calculation results for the customer response cycle and new product response cycle at the contact person's job level were obtained.

[0111] 3. Divide the dataset into training and testing sets and configure the model parameters. The specific steps are as follows:

[0112] (1) The preprocessed “feature-label” dataset was divided into training set and test set in a 7:3 ratio using stratified sampling.

[0113] (2) Random forest model parameter configuration, specifically including: assigning a value of 100 to the number of decision trees; assigning a value of 10 to the maximum tree depth; assigning a value of 5 to the minimum number of split samples; assigning a value of 5 to the minimum number of leaf node samples; and selecting the Gini coefficient as the feature selection method. The parameter experiment process is as follows: Figure 4 As shown.

[0114] 4. Train the random forest model and verify its performance. The specific steps are as follows:

[0115] (1) Input the input features and output labels of the training set into the random forest model to generate the model prediction results.

[0116] (2) Input the input features of the test set into the trained model to obtain the predicted marketing strategy labels, compare them with the real labels of the test set, and calculate the core evaluation indicators:

[0117] Customer conversion rate metric indicates the percentage of customers who make their first purchase as a result of the prediction strategy;

[0118] The strategy response rate metric represents the customer's click / engagement rate with the predictive strategy;

[0119] The customer repurchase rate metric represents the percentage of customers who repurchase within 30 days after the prediction strategy is implemented.

[0120] When the average of the three indicators is ≥80%, the model is considered to have met the performance standards. If it does not meet the standards, the model parameters are adjusted and retrained until the standards are met. The model training results are as follows: Figure 5 As shown.

[0121] 5. Automated output of full-dimensional marketing strategies: Input the multi-dimensional feature profile of the target customers who need to plan marketing strategies into the model. Based on the "customer feature-marketing strategy" correlation rules learned during the training phase, the model outputs full-dimensional marketing strategies, including recommended categories, communication channels, discount thresholds, and contact person job level.

[0122] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0123] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A novel customer relationship management method, characterized in that, Includes the following processes: S1 establishes a data foundation composed of customer omnichannel interaction data and enterprise business data. The customer omnichannel interaction data includes communication content, consumption records, and service feedback; the enterprise business data includes marketing campaign history and transaction data. S2, based on a data foundation, constructs customer profiles including behavioral preferences, management cycle stages, and marketing response patterns: By combining consumption records, transaction data, marketing campaign history, and communication content, customer behavioral preference characteristics are defined based on feature engineering methods. Based on communication content and consumption records, management cycle stages are defined using feature engineering, RFM, and FCI methods. By combining communication content, marketing campaign history, and transaction data, marketing response patterns are defined based on feature engineering methods. S3 uses customer profiles as input parameters for the random forest model, combines historical data to train the random forest model, and outputs marketing strategies including recommended product categories, communication channels, interaction frequency, discount thresholds, and contact person job levels.

2. The novel customer relationship management method as described in claim 1, characterized in that: The customer omnichannel interaction data specifically includes: communication content between customers and enterprises, which refers to the original communication content data group consisting of communication time, communication channel, communication topic, and interactive text or voice records; consumption records, which refers to the original consumption record data group consisting of consumption time, consumed product ID, consumption amount, and payment method; and service feedback data, which specifically consists of the original service feedback data group consisting of feedback time, feedback type, feedback content, and satisfaction rating. The enterprise business data includes marketing campaign historical data, which refers to the original marketing campaign data set consisting of campaign ID, campaign theme, campaign time, campaign rules, and participating customer scope; and transaction data, which refers to the original transaction data set consisting of transaction order number, transaction amount, transaction status, transaction completion time, and refund record.

3. The novel customer relationship management method as described in claim 1, characterized in that: Preprocess the data in S1, including performing duplicate record removal, defining and performing rules for missing values ​​in key fields, and defining and performing filtering rules for identifying abnormal data. Simultaneously, the preprocessed data undergoes data format conversion; including: Convert all amounts to yuan; Align the time in the format "year-month-day-hour-minute-second"; The customer's digital ID is a unique 12-digit code consisting of the "year-month-day plus the communication sequence number of the day" from the first communication record.

4. The novel customer relationship management method as described in claim 1, characterized in that: In S2, customer behavior preference features are defined based on feature engineering methods by combining consumption records, transaction data, marketing activity history, and communication content. Specifically, these features refer to the 1st, 2nd, and 3rd categories of total purchase value and their corresponding discount thresholds, as well as the 1st communication channel in terms of usage time. Based on communication content and consumption records, and using feature engineering, RFM, and FCI methods, define management cycle stages, including: defining customer stage labels using feature engineering methods based on communication content and consumption records, including potential customer stage, new customer stage, and existing customer stage; High-value customers, defined by RFM method with consumption records as input, are those whose total transaction amount ranks in the top 30% of all customers; and customers at risk of churn, defined by FCI method with communication content as input, are those whose silence period exceeds the average communication interval of the previous year. By combining communication content, marketing campaign history, and transaction data, marketing response patterns are defined based on feature engineering methods. These patterns include the job level of the communication contact person and the response cycle, as well as the time interval between the transaction product and its initial release.

5. A novel customer relationship management method as described in claim 1 or 4, characterized in that: The process of defining customer behavior preference characteristics is as follows: 1) Using "customer digital ID" as the grouping dimension, further subdivide each customer's consumption records by "product category ID" to calculate the total purchase value of each customer in each category, i.e., the total transaction amount; Sort the total purchase value in descending order and sum them up to get the total purchase value. Extract the categories ranked 1st, 2nd, and 3rd and label them as Category 1, Category 2, and Category 3 of the total purchase value for the customer, respectively. If the customer has fewer than 3 categories, mark the remaining unmet ranking positions as none. For Category 1, 2nd, and 3 of the total purchase value for the customer, backtrack and match the discount data that the customer has enjoyed in all historical transactions under this category, and calculate the average value as the discount threshold corresponding to Category 1, 2nd, and 3. 2) Extract records of interactions between target customers via telephone, online customer service, email, and WeChat from the customer communication logs of the data base, calculate the cumulative usage time, and mark the communication channel ranked first as the customer's preferred communication channel; 3) Establish a customer behavior preference feature data group based on customer digital ID, including the total value of purchased goods; the first, second, and third ranked categories and their corresponding discount thresholds; and preferred communication channels.

6. A novel customer relationship management method as described in claim 1 or 4, characterized in that: The defined management cycle phase, specifically, is as follows: 1) Extract communication records from the data base using customer digital IDs. Collect all communication records within the past year. Customers with zero purchases in their transaction records are labeled as "potential customers". Customers with exactly one purchase are labeled as "new customers". Customers with more than one purchase are labeled as "existing customers". 2) Calculate the total amount of goods purchased by all customers in the past year using the RFM model and sort them. Mark the top 30% of customers in the sorting results as "high-value customers". 3) Using the interval analysis logic of the FCI model, statistically analyze all communication records of all customers in the most recent year, and calculate the time interval between two consecutive communications to obtain the average communication interval duration of the previous year. Calculate the duration from the customer's last communication date to the current date based on the current date, and define the customer's silence period; compare the customer's silence period with the average communication interval of the previous year. If the customer's silence period is greater than the average communication interval of the previous year, mark it as a "customer at risk of churn". 4) Establish a customer management cycle stage data group based on the customer's digital ID, which includes customer stage tags and customer status tags. If the customer has no status tag, mark it as none.

7. A novel customer relationship management method as described in claim 1 or 4, characterized in that: The defined marketing response pattern, specifically, is as follows: 1) Extract communication records from the data base using the customer's digital ID, compile all communication records within the past year, and extract the initiation time, contact person's job level, and response time for each communication record; The time difference between the initiation time of communication and the final response time of this communication is calculated and defined as a single response cycle; the average value of all single response cycles is statistically analyzed by classifying the contact person as junior customer service, senior consultant, and sales director, and is used as the response cycle indicator for the contact person's job level corresponding to the user's digital ID; 2) Extract transaction records from the data base using the customer's digital ID, and statistically analyze all transaction records within the past year. Extract the first release time of the corresponding product ID from each transaction record. For each transaction, calculate the interval between the transaction completion time and the first release time of the product, which is defined as the single new product response cycle. Calculate the average of all single new product response cycles based on the total value of the product categories purchased by the customer's digital ID in categories 1, 2, and 3, and use this as the new product response cycle indicator corresponding to the user's digital ID. 3) Establish a marketing response pattern data group based on customer digital IDs, including the response cycle of the contact person's job level and the new product response cycle.

8. The novel customer relationship management method as described in claim 1, characterized in that: The S3 section, which combines historical data to train a random forest model, specifically includes: Random forest model parameter configuration, including: assigning a value of 100 to the number of decision trees; assigning a value of 10 to the maximum tree depth; assigning a value of 5 to the minimum number of split samples; assigning a value of 5 to the minimum number of leaf node samples; and selecting the Gini coefficient as the feature selection method. Input the input features and output labels of the training set into the random forest model to generate the model's prediction results; Input the features from the test set into the trained model to obtain the predicted marketing strategy labels, compare them with the true labels from the test set, and calculate the core evaluation metrics: Customer conversion rate metric indicates the percentage of customers who make their first purchase as a result of the prediction strategy; The strategy response rate metric represents the customer's click / engagement rate with the predictive strategy; The customer repurchase rate metric represents the percentage of customers who repurchase within 30 days after the prediction strategy is implemented. If the average of the three indicators is ≥80%, the model is considered to have met the performance standards; if it does not meet the standards, the model parameters are adjusted and the model is retrained until it meets the standards.

Images