Customer relationship maintenance optimization system based on comparative learning

Abstract

The invention discloses a customer relationship maintenance optimization system based on comparative learning, and the system comprises a data collection and preprocessing module which is used for collecting and preprocessing multi-source behavior data associated with a target customer; the customer characterization module is used for training the standardized customer data set by adopting an improved MoCo-v3 contrast learning method to generate a customer characterization model; the customer state analysis module is used for carrying out feature coding, similarity calculation and state marking on the target customer; the maintenance instruction generation module is used for determining maintenance content and an execution mode based on the client state information; the model updating module is used for collecting actual response behavior data and performing incremental training on the customer characterization model; and the maintenance optimization module is used for outputting a customer relationship maintenance optimization result based on the updated customer characterization model. According to the invention, the improved comparative learning method is adopted to realize the intelligent maintenance of the customer relationship, and the method has the advantages of accurate feature representation, adaptive strategy optimization and high maintenance efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent customer relationship management, and particularly relates to a customer relationship maintenance optimization system based on contrast learning. BACKGROUND

[0002] The existing customer relationship management system is usually based on artificial experience or traditional supervised learning model, and the transaction records, communication frequency and historical behavior of customers are statistically analyzed to predict the customer churn risk and give maintenance suggestions. However, such methods rely on a large number of manually annotated training samples, are difficult to adapt to the dynamic changes of customer behavior patterns, and the generated maintenance strategies are mostly static rules or single risk level prompts, lacking multi-dimensional characterization of customer relationship urgency, life cycle stage and group characteristics, resulting in low pertinence and timeliness of maintenance decisions.

[0003] With the diversification of customer data sources and the complexity of interaction methods, the existing technology is difficult to realize deep representation and adaptive optimization of customer behavior characteristics, and the model often lacks generalization ability between different scenes, cannot adjust the strategy according to the real-time response behavior of the customer, and forms a closed-loop optimization mechanism. Therefore, there is an urgent need for a customer relationship maintenance system that can realize efficient learning of customer representation based on contrast learning, generate executable maintenance instructions combined with customer state labels, and have adaptive optimization capability to improve the intelligent and dynamic response level of customer retention.

[0004] Therefore, how to provide a customer relationship maintenance optimization system based on contrast learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a customer relationship maintenance optimization system based on contrast learning. The present application uses an improved contrast learning method to realize intelligent maintenance of customer relationship, and has the advantages of accurate feature representation, adaptive strategy optimization and high maintenance efficiency.

[0006] According to the customer relationship maintenance optimization system based on contrast learning of the present application, the system comprises:

[0007] A data acquisition and preprocessing module is configured to acquire and preprocess multi-source behavior data associated with a target customer and generate a standardized customer data set.

[0008] A customer representation module is configured to train the standardized customer data set using an improved MoCo-v3 contrast learning method and generate a customer representation model.

[0009] A customer state analysis module is configured to perform feature encoding and similarity calculation on the target customer based on the customer representation model, output customer group labels, relationship urgency labels and life cycle stage labels, and generate customer state information.

[0010] a maintenance instruction generation module configured to determine maintenance content and execution mode based on the customer state information, and generate a customer maintenance instruction;

[0011] a model updating module configured to collect actual response behavior data of the target customer, update the training sample set, and perform incremental training on the customer representation model;

[0012] a maintenance optimization module configured to perform adaptive optimization on the maintenance instructions of the customers in the same cluster based on the updated customer representation model, and output a customer relationship maintenance optimization result.

[0013] Optionally, the modules are implemented through the following methods:

[0014] a plurality of source behavior data associated with the target customer are collected, preprocessed, and standardized to generate a standardized customer data set;

[0015] the standardized customer data set is trained using an improved MoCo-v3 contrastive learning method, and a customer representation model is generated through multi-dimensional feature representation construction, sample similarity constraint, and model parameter iterative update;

[0016] the target customer in the maintenance state is processed based on the customer representation model, including feature encoding, similarity calculation, and state marking, and customer cluster marking, relationship urgency marking, and life cycle stage marking are outputted to generate customer state information;

[0017] strategy evaluation and action selection are performed based on the customer state information to determine the maintenance content and execution mode of the target customer, and a customer maintenance instruction is generated;

[0018] the customer maintenance instruction is executed, actual response behavior data of the target customer is collected and associated, the training sample set is updated, and the customer representation model is incrementally trained to output an updated customer representation model;

[0019] based on the updated customer representation model, maintenance instructions are generated for the customers in the same cluster as the target customer and are adaptively optimized to output a customer relationship maintenance optimization result.

[0020] Optionally, the plurality of source customer data includes transaction behavior data, communication record data, emotional feedback data, and social interaction data, and the preprocessing includes missing value filling, normalization, timestamp encoding, and feature alignment on the collected data.

[0021] Optionally, the generation of the customer representation model includes the following specific steps:

[0022] The improved MoCo-v3 contrast learning method is used to construct a double-tower deep neural network structure, the initial parameters of the two-tower encoder are initialized with the same value, each customer sample in the standardized customer data set is represented as a feature sequence input into the double-tower deep neural network structure, the improved MoCo-v3 contrast learning method is based on the original double-tower momentum contrast learning mechanism, and the multi-dimensional feature disturbance processing is added to enhance the sample discrimination, the lightweight multi-head self-attention structure is added to optimize the feature correlation modeling, and the momentum coefficient is adjusted to be an adaptive dynamic updating mode to improve the parameter updating stability and convergence speed, and the structure includes a main tower encoder and a momentum tower encoder;

[0023] The feature disturbance processing is performed on each customer sample in the standardized customer data set, and the disturbance results are input into the main tower encoder and the momentum tower encoder respectively, the main tower encoder outputs a query vector, and the momentum tower encoder outputs a key vector, the query vector and the key vector from the same customer sample form a positive sample pair, and the query vector and the key vector from different customer samples are combined to form a negative sample pair;

[0024] The positive sample similarity is calculated for the query vector and the key vector of the same customer sample, and the negative sample similarity is calculated for the query vector and the key vector of different customer samples, and the similarity score is calculated by using the dot product similarity calculation method:

[0025] ;

[0026] Wherein, represents the similarity of the query vector and the key vector in the same feature space, represents the i-th component of the query vector, represents the i-th component of the key vector;

[0027] The similarity calculation contrast learning loss is used as an optimization target to perform back propagation update on the main tower encoder parameters:

[0028] ;

[0029] Wherein, represents the contrast learning loss value, represents the natural logarithm function, represents the exponential function, represents the query vector output by the main tower encoder, represents the key vector output by the momentum tower encoder, represents the dot product similarity of the positive sample pair, represents the dot product similarity of the negative sample pair, represents the temperature coefficient;

[0030] The momentum tower encoder parameters are updated based on the updating result of the main tower encoder parameters by using an exponential moving average strategy:

[0031] ;

[0032] wherein, represents the momentum tower encoder parameters at the current training moment, represents the current training time step, represents a momentum coefficient, represents the momentum tower encoder parameters at the last training moment, represents the main tower encoder parameters updated at the current training moment, represents the momentum tower encoder, represents the main tower encoder.

[0033] The above steps are repeatedly executed to iteratively train all customer samples in the standardized customer dataset until the contrastive learning loss function converges, and a customer representation model is generated.

[0034] Optionally, the generation of the customer state information comprises the following specific steps:

[0035] The standardized customer dataset of the target customer in the maintenance state is input into the customer representation model for feature encoding processing, and an embedding vector of the target customer is output, the embedding vector being composed of a plurality of numerical components, each component corresponding to a feature representation of the customer in the behavior, interaction and transaction dimensions;

[0036] The embedding vector of the historical customer is extracted, and a distance value between the embedding vector of the target customer and the embedding vector of each historical customer is calculated;

[0037] The similarity between the target customer and the historical customers is ranked according to the size of the distance value, a historical customer set within a preset range is selected, and a customer clustering label is generated;

[0038] The relationship closeness between the target customer and the historical customers is determined based on the concentration degree of the adjacent distance distribution, when the distance variance is less than a preset variance threshold, the target customer is marked as a high relationship urgency customer, when the distance variance is greater than or equal to the preset variance threshold, the target customer is marked as a low relationship urgency customer, and a corresponding relationship urgency label is output;

[0039] The customer state is divided into stages according to the transaction time interval, interaction frequency and active cycle length of the target customer, a life cycle stage label is output, and the customer clustering label, the relationship urgency label and the life cycle stage label are combined to generate the customer state information.

[0040] Optionally, the generation of the customer maintenance instruction comprises: joint analysis of the customer group label, the relationship urgency label and the life cycle stage label in the customer state information, weighted calculation according to the feature weight corresponding to each label to obtain the maintenance priority of the target customer, interval division of the maintenance priority result and screening of the corresponding strategy parameter set, strategy matching operation for the customer maintenance priority in different intervals, determination of the communication mode and the touch frequency parameter according to the relationship urgency label, determination of the resource investment level and the preferential range parameter according to the life cycle stage label, determination of the interaction channel and the communication content parameter according to the customer group label, combination and time sequence sorting of all the strategy parameters screened, and generation of the customer maintenance instruction including the communication mode, the touch opportunity, the resource investment level and the loss warning level.

[0041] Optionally, the obtaining of the updated customer representation model comprises: data association of the collected target customer actual response behavior data and the corresponding customer embedding vector, category coding and time sequence sorting of the response behavior to form a behavior response feature sequence, merging of the feature sequence and the customer embedding vector and the state label in the original training sample set to generate an updated training sample set, data standardization and feature alignment processing of the updated sample set, small-step gradient update training of the newly added data batch in the sample set, dynamic adjustment of the learning rate according to the change of the contrast learning loss function, layer-by-layer back propagation and weight update of the model parameters, momentum smoothing of the historical weight and prevention of overfitting, stopping of the training when the loss reduction amplitude is lower than a preset threshold in continuous several training rounds, and output of the updated customer representation model.

[0042] Optionally, the output of the customer relationship maintenance optimization result comprises: batch processing and feature comparison of the customer embedding vectors belonging to the same group as the target customer based on the updated customer representation model, calculation of the feature similarity and the behavior deviation degree of each customer in the group, strategy adjustment of the customer samples with a deviation degree exceeding a preset deviation threshold, keeping of the original customer maintenance instruction for the customer samples with a deviation degree not exceeding the preset deviation threshold, adaptive weight update of the customer maintenance instructions of all the customer samples, dynamic correction of the communication mode, the touch opportunity, the resource investment level and the loss warning level according to the real-time collected actual response behavior data, and multi-round iteration verification of the optimized customer maintenance instruction, and output of the customer relationship maintenance optimization result when the strategy converges and meets the preset threshold condition in continuous verification rounds.

[0043] The beneficial effects of the present application are:

[0044] The application realizes deep representation of customer multi-dimensional characteristics and adaptive optimization of dynamic maintenance strategy by introducing an improved contrast learning mechanism in the customer relationship management process, and overcomes the shortcomings of traditional methods, such as dependence on artificial rules, limited feature expression capability, and inability to form a real-time feedback closed loop. The system collects multi-source behavior data such as transactions, communication, emotional feedback and social interaction at the data level, and performs standardized preprocessing to ensure the time consistency and feature comparability of the input data; at the model level, a double-tower deep neural network is established through an improved MoCo-v3 contrast learning structure to realize high-discrimination representation learning under unsupervised conditions, so that the customer has distinguishable embedded features in the behavior, interaction and transaction dimensions, providing a high-quality feature basis for subsequent customer grouping and state marking.

[0045] The application further introduces multi-dimensional feature disturbance, lightweight multi-head self-attention structure and adaptive momentum update mechanism in the model training process, so that the customer representation model can maintain stable feature alignment relationship between different samples, and continuously improve the learning convergence speed under similarity constraint. By constructing a customer grouping and relationship urgency determination mechanism based on distance distribution, accurate division of the association strength and life cycle stage between customers is realized, so that the system can identify the customer relationship network structure from a global perspective, dynamically generate customer state information, and provide a quantitative basis for decision-making of maintenance actions.

[0046] In addition, the application establishes a data backflow and model incremental update mechanism in the customer maintenance execution stage, associates the actual response behavior data of the customer with the embedded vector, realizes continuous learning of model parameters through small-step gradient optimization, so that the system can adjust the communication mode, touch time, resource allocation level and loss warning level in real time according to the latest behavior feedback of the customer. This design not only significantly improves the adaptability and execution accuracy of customer maintenance strategy, but also builds an intelligent closed loop of customer relationship management, realizes dynamic self-learning of the whole process from data perception, state recognition to strategy generation and optimization, and improves the automation and intelligence level of enterprises in customer retention and value mining. BRIEF DESCRIPTION OF DRAWINGS

[0047] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0048] Figure 1 A method flowchart of a customer relationship maintenance optimization system based on contrast learning is proposed for the application;

[0049] Figure 2 An improved MoCo-v3 contrast learning method for constructing a customer representation model of a customer relationship maintenance optimization system based on contrast learning is proposed for the application. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0051] refer to Figure 1 and Figure 2 A customer relationship maintenance and optimization system based on contrastive learning, comprising:

[0052] The data acquisition and preprocessing module is used to collect and preprocess multi-source behavioral data associated with target customers and generate standardized customer datasets.

[0053] The customer representation module is used to train a standardized customer dataset using an improved MoCo-v3 contrastive learning method to generate a customer representation model.

[0054] The customer status analysis module is used to perform feature encoding and similarity calculation on target customers based on the customer representation model, and output customer grouping labels, relationship urgency labels and life cycle stage labels to generate customer status information;

[0055] The maintenance instruction generation module is used to determine the maintenance content and execution method based on customer status information and generate customer maintenance instructions.

[0056] The model update module is used to collect actual response behavior data of target customers, update the training sample set, and perform incremental training on the customer representation model.

[0057] The maintenance optimization module is used to adaptively optimize maintenance instructions for the same customer segment based on the updated customer representation model, and output the customer relationship maintenance optimization results.

[0058] In this embodiment, the modules are connected through the following method:

[0059] Collect multi-source behavioral data associated with target customers, preprocess it, and generate a standardized customer dataset;

[0060] An improved MoCo-v3 contrastive learning method is used to train a standardized customer dataset. A customer representation model is generated through multi-dimensional feature representation construction, sample similarity constraints, and iterative updating of model parameters.

[0061] Based on the customer representation model, feature encoding, similarity calculation and status labeling are performed on target customers in the pending maintenance state, and customer group labels, relationship urgency labels and life cycle stage labels are output to generate customer status information;

[0062] Based on the customer state information, policy evaluation and action selection are performed to determine the maintenance content and execution mode of the target customer, and generate customer maintenance instructions;

[0063] The customer maintenance instructions are executed, actual response behavior data of the target customer is collected and associated, the training sample set is updated, and the customer representation model is incrementally trained, and an updated customer representation model is output;

[0064] Based on the updated customer representation model, maintenance instructions are generated for customers belonging to the same cluster as the target customer and are adaptively optimized, and a customer relationship maintenance optimization result is output.

[0065] In the embodiment, the multi-source customer data includes transaction behavior data, communication record data, emotional feedback data, and social interaction data, and the preprocessing includes missing value filling, normalization, timestamp encoding, and feature alignment of the collected data.

[0066] In the embodiment, the generation of the customer representation model includes the following specific steps:

[0067] A double-tower deep neural network structure is constructed based on the improved MoCo-v3 contrastive learning method, the initial parameters of the two-tower encoder are initialized with the same value, and each customer sample in the standardized customer data set is represented as a feature sequence input into the double-tower deep neural network structure. The improved MoCo-v3 contrastive learning method adds multi-dimensional feature disturbance processing to enhance sample discrimination, adds a lightweight multi-head self-attention structure to optimize feature correlation modeling, adjusts the momentum coefficient to an adaptive dynamic update mode to improve parameter update stability and convergence speed. The structure includes a main tower encoder and a momentum tower encoder.

[0068] Each customer sample in the standardized customer data set is subjected to feature disturbance processing, and different disturbance results are input into the main tower encoder and the momentum tower encoder. The main tower encoder outputs a query vector, and the momentum tower encoder outputs a key vector. The query vector and the key vector from the same customer sample form a positive sample pair, and the query vector and the key vector from different customer samples are combined to form a negative sample pair.

[0069] The feature perturbation processing includes: performing additive noise perturbation and random masking processing on each customer sample in the standardized customer dataset in the numerical dimension, adding a small noise conforming to a zero mean and unit variance distribution to a continuous feature, randomly masking part of the feature values of a discrete feature and replacing them with feature values of a same-distribution sample, performing sliding window resampling and time series truncation in the time series dimension, truncating the customer behavior sequence by a preset window step and randomly adjusting the starting position to generate data views of different time segment combinations, performing feature permutation and vector rotation transformation in the semantic dimension, partially exchanging feature groups within the same customer sample and performing a small rotation in the embedding space to obtain diversified feature representations;

[0070] The output of the query vector includes: inputting the diversified feature representation into the main tower encoder, sequentially performing multi-layer convolution feature extraction, batch normalization processing and nonlinear activation operation, extracting local behavior pattern features, performing multi-head self-attention calculation on the extracted features, calculating the correlation weight between the features and generating a weighted feature representation, performing full connection mapping and layer normalization on the weighted feature representation, compressing the high-dimensional feature space and forming an embedding representation of fixed dimension, aggregating the time series features through an average pooling layer to generate a global feature vector, inputting the global feature vector into the output layer for linear transformation, and outputting the query vector corresponding to the customer sample;

[0071] The output of the key vector includes: inputting the diversified feature representation into the momentum tower encoder, sequentially performing convolution feature extraction, normalization and nonlinear activation operation to generate a feature matrix, performing multi-head self-attention calculation on the feature matrix, calculating the correlation weight between the features and generating a weighted feature matrix, performing full connection mapping and layer normalization on the weighted feature matrix to obtain an embedding matrix, performing average pooling operation based on the embedding matrix to aggregate the time series features to obtain a global feature vector, inputting the global feature vector into the output layer for linear transformation to generate the key vector;

[0072] The query vector and the key vector of the same customer sample are calculated for positive sample similarity, and the query vector and the key vector of different customer samples are calculated for negative sample similarity. The similarity score uses a dot product similarity calculation method:

[0073] ;

[0074] Wherein, represents the similarity of the query vector and the key vector in the same feature space, represents the i-th component of the query vector, represents the i-th component of the key vector;

[0075] Based on the similarity calculation, a contrastive learning loss is obtained, and the contrastive learning loss is used as an optimization target to perform back propagation update on the main tower encoder parameters:

[0076] ;

[0077] in, This represents the contrastive learning loss value. Represents the natural logarithm function. Represents an exponential function. This represents the query vector output by the main tower encoder. This represents the key vector output by the momentum tower encoder. This represents the dot product similarity of positive sample pairs. This represents the dot product similarity of negative sample pairs. Indicates the temperature coefficient;

[0078] The momentum tower encoder parameters are updated using an exponential moving average strategy based on the update results of the main tower encoder parameters:

[0079] ;

[0080] in, This represents the momentum tower encoder parameters at the current training moment. Indicates the current training time step. Indicates the momentum coefficient. This represents the momentum tower encoder parameters at the previous training time. This represents the updated parameters of the master encoder at the current training moment. This indicates a momentum tower encoder. Indicates the main tower encoder;

[0081] Repeat the above steps to iteratively train all customer samples in the standardized customer dataset until the contrastive learning loss function converges and a customer representation model is generated. The convergence condition is that the loss decreases below a preset threshold in several consecutive training rounds.

[0082] In this embodiment, the generation of customer status information includes the following specific steps:

[0083] The standardized customer dataset of the target customers in the unmaintained state is input into the customer representation model, and feature encoding processing is performed to output the embedding vector of the target customer. The embedding vector consists of several numerical components, each of which corresponds to the feature representation of the customer in the dimensions of behavior, interaction and transaction.

[0084] The output of the embedding vector includes: inputting standardized feature data of the target customer into the customer representation model; performing multi-layer feature encoding and non-linear mapping on the input data; sequentially performing convolutional feature extraction, batch normalization, and activation function operations; extracting local correlation information of customer behavior, interaction, and transaction features; performing multi-head self-attention calculation on the extracted feature matrix; calculating the correlation weights between each feature and generating a weighted feature matrix; performing fully connected mapping and layer normalization on the weighted feature matrix; compressing the high-dimensional feature space to form a fixed-dimensional feature representation; aggregating different temporal features through average pooling to obtain a global feature vector; inputting the global feature vector into the output layer to perform a linear transformation; and outputting the embedding vector of the target customer.

[0085] Extract the embedding vectors of historical customers, calculate the distance between the target customer embedding vector and each historical customer embedding vector. The distance calculation is to sum the squared differences between the target customer and historical customers on the same feature dimension, and then perform a square root operation on the sum. The result is used as the distance metric between the two.

[0086] The similarity between target customers and historical customers is sorted according to the distance value. A set of historical customers whose distance values ​​are within a preset range is selected to generate customer grouping tags. The customer grouping tags are used to characterize the association level and category of target customers in the customer relationship network.

[0087] The degree of concentration of the adjacent distance distribution determines the closeness of the relationship between the target customer and the historical customer. When the distance variance is less than the preset variance threshold, the target customer is marked as a high relationship urgency customer. When the distance variance is greater than or equal to the preset variance threshold, the target customer is marked as a low relationship urgency customer. The corresponding relationship urgency label is output. The relationship urgency label is used to characterize the association strength and maintenance priority between the target customer and the historical customer in the embedded vector space.

[0088] Customer status is divided into stages based on the target customer's transaction time interval, interaction frequency, and active cycle length. Lifecycle stage markers are output, and customer group markers, relationship urgency markers, and lifecycle stage markers are combined to generate customer status information.

[0089] The output of the lifecycle stage marker includes: collecting the transaction time series, interaction record series and active behavior logs of the target customer; calculating the mean and standard deviation of the transaction time interval; performing cumulative statistics on the interaction frequency within a fixed time window; performing sliding window aggregation on the active cycle length; and performing threshold determination and interval mapping based on the combined characteristics of time interval, interaction frequency and active cycle length to complete the stage division of customer status. Customers with short transaction intervals and high interaction frequency are marked as active stage, customers with moderate transaction intervals and stable interaction frequency are marked as stable stage, and customers with extended transaction intervals and decreased interaction frequency are marked as churn warning stage. The corresponding lifecycle stage marker is then output.

[0090] The lifecycle stage markers are used to characterize the classification of the target customer's status stage within the transaction and interaction activity cycle;

[0091] The classification identifiers include active phase markers, stabilization phase markers, and churn warning phase markers. The active phase markers reflect the state of frequent customer transactions and intensive interactions, the stabilization phase markers reflect the state of regular customer transactions and stable behavior, and the churn warning phase markers reflect the state of extended customer transaction intervals and decreased interaction frequency.

[0092] In this embodiment, the generation of customer maintenance instructions includes: jointly analyzing customer segmentation markers, relationship urgency markers, and lifecycle stage markers in customer status information; performing weighted calculations based on the feature weights corresponding to each marker to obtain the maintenance priority of the target customer; dividing the maintenance priority results into intervals and filtering the corresponding set of strategy parameters; performing strategy matching operations for customer maintenance priorities within different intervals; determining communication method and reach frequency parameters based on relationship urgency markers; determining resource allocation level and discount parameters based on lifecycle stage markers; determining interaction channels and communication content parameters based on customer segmentation markers; and combining and chronologically sorting all the filtered strategy parameters to generate customer maintenance instructions that include communication method, reach timing, resource allocation level, and churn warning level.

[0093] In this embodiment, obtaining the updated customer representation model includes: associating the collected actual response behavior data of target customers with the corresponding customer embedding vectors; classifying and sorting the response behaviors by time series to form a behavior response feature sequence; merging the feature sequence with the customer embedding vectors and state labels in the original training sample set to generate an updated training sample set; performing data standardization and feature alignment on the updated sample set; performing small-step gradient update training on batches of newly added data in the sample set; dynamically adjusting the learning rate based on the changes in the contrastive learning loss function; performing layer-by-layer backpropagation and weight update on the model parameters; maintaining the momentum smoothness of historical weights and preventing overfitting; stopping training when the loss decreases below a preset threshold in several consecutive training rounds; and outputting the updated customer representation model. The actual response behaviors include reply behavior, repurchase behavior, rejection behavior, complaint behavior, and churn behavior.

[0094] In this embodiment, the output of the customer relationship maintenance optimization result includes: based on the updated customer representation model, batch processing and feature comparison of customer embedding vectors belonging to the same subgroup as the target customer are performed; the feature similarity and behavioral deviation of each customer in the subgroup are calculated; strategy adjustment is performed on customer samples whose deviation exceeds a preset deviation threshold; the original customer maintenance instructions are maintained on customer samples whose deviation does not exceed the preset deviation threshold; adaptive weight update is performed on the customer maintenance instructions of all customer samples; communication methods, contact timing, resource allocation level and churn warning level are dynamically corrected based on real-time collected actual response behavior data; multiple rounds of iterative verification are performed on the optimized customer maintenance instructions; and the customer relationship maintenance optimization result is output when the strategies in the continuous verification rounds converge and meet the preset threshold conditions.

[0095] Example 1:

[0096] To verify the feasibility of this invention in practice, it was applied to the customer relationship management scenario of a large-scale e-commerce platform. This platform has tens of millions of active customers, generating multi-dimensional behavioral data daily, including transactions, browsing, inquiries, and after-sales service. Traditional customer maintenance methods rely on manual grouping and fixed rule-based push notifications, which suffer from problems such as delayed feature updates, inflexible strategy adjustments, and uneven resource allocation. This is especially problematic during periods of significant fluctuation in customer activity, easily leading to the churn of high-value customers.

[0097] In implementing this invention, the system first collects and preprocesses multi-source behavioral data associated with target customers in the platform database, including transaction records, communication logs, social interactions, and emotional feedback content. Missing value imputation, normalization, and timestamp encoding are performed to generate a standardized customer dataset. Subsequently, an improved MoCo-v3 contrastive learning method is used to construct a customer representation model. The main tower and momentum tower of a dual-tower deep neural network structure are used to perturb the samples' features, perform contrastive learning training, and update momentum parameters, thereby enabling the customer embedding vectors to possess higher temporal stability and feature discriminative power.

[0098] After generating the customer representation model, the system inputs the customer data to be maintained into the model, outputs customer embedding vectors, calculates the distance to historical customers, and automatically generates customer segmentation tags, relationship urgency tags, and lifecycle stage tags to form customer status information. Based on this status information, the system assesses maintenance priorities, generates customer maintenance instructions that include communication methods, contact timing, resource allocation levels, and churn warning levels, and automatically executes them through multi-channel systems.

[0099] During maintenance execution, the system continuously collects actual customer response behavior data, including responses, repeat purchases, rejections, complaints, and churn behaviors, and updates this data in association with customer embedding vectors to achieve dynamic incremental optimization of the training sample set. The updated customer representation model is further used for adaptive strategy optimization of similar customer groups, automatically adjusting maintenance parameters for customers with significant behavioral deviations within the group, while maintaining the original maintenance instructions for customers with stable behavior. After multiple rounds of training and iteration, the system finally outputs the optimized customer relationship maintenance results, achieving intelligent closed-loop adjustment of customer maintenance strategies.

[0100] Application results show that the system can accurately capture customer lifecycle stages and relationship change trends, improve the accuracy of customer segmentation and the response speed of maintenance strategies, effectively reduce the risk of high-value customer churn, realize the transformation of customer relationship management from experience-based decision-making to intelligent self-learning decision-making, and verify the feasibility and superiority of the invention in the field of customer relationship maintenance and optimization.

[0101] Table 1. Performance Comparison of the Invention and Traditional Customer Relationship Maintenance Optimization Methods

[0102] Indicator category Traditional method Inventive method Customer cluster accuracy (%) 82.4 93.7 Relationship closeness recognition accuracy (%) 78.1 91.5 Life cycle stage division accuracy (%) 80.6 92.8 Customer maintenance response rate (%) 64.2 81.9 Resource investment utilization rate (%) 72.5 89.4 Model convergence rounds (%) 48 29 High-value customer attrition rate reduction (%) — 18.4 Customer maintenance strategy update delay (%) 16.8 7.3 System running stability (annual downtime length / hour) 11.5 3.2

[0103] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0104] In terms of customer segmentation accuracy, the method of this invention achieves 93.7%, which is 13.7% higher than the 82.4% of the traditional method. This improvement is mainly due to the introduction of the dual-tower structure and the multi-dimensional feature perturbation mechanism, which makes the customer embedding representation more discriminative and can more finely distinguish the differences in customer features, thus avoiding the problem of misjudgment of similar features in traditional rule-based segmentation methods.

[0105] In terms of accuracy in identifying relationship urgency, the method of this invention improves by 17.1%, from 78.1% to 91.5%. This is because the improved MoCo-v3 method combines adaptive momentum updates and a multi-head attention structure in similarity calculation, enabling the model to capture more complex nonlinear relationships between customer behaviors, thereby accurately identifying potentially high-risk customers. The accuracy of lifecycle stage segmentation improves from 80.6% to 92.8%, an increase of 15.1%, indicating that the system performs better in time-series behavioral pattern recognition and can dynamically perceive the stage changes of customers from active to churn.

[0106] In terms of business performance, the method of this invention increased the customer maintenance response rate from 64.2% to 81.9%, an improvement of 27.6%, indicating that the customer maintenance instructions output by the model are more in line with customer behavior characteristics and preferences. At the same time, the resource allocation utilization rate increased by 23.3%, from 72.5% to 89.4%, indicating that the allocation of maintenance resources among customer groups is more reasonable, avoiding the problems of resource concentration or waste in traditional strategies.

[0107] In terms of training and system performance, the number of model convergence epochs decreased from 48 to 29, improving the convergence speed by 39.6%. This is attributed to the introduced exponential moving average update mechanism, which smooths the changes in momentum tower parameters and avoids gradient oscillations. The system latency during customer maintenance strategy updates decreased from 16.8 seconds to 7.3 seconds, an improvement of 56.5%, significantly enhancing the model's responsiveness in real-time scenarios. The system's annual downtime was reduced from 11.5 hours to 3.2 hours, improving operational stability by 72.2%, indicating that the model architecture and data management strategy of this invention are more robust in high-concurrency environments.

[0108] Furthermore, this invention reduces the churn rate of high-value customers by 18.4%. This is because the system can automatically complete customer status monitoring, strategy evaluation, and maintenance instruction generation, and achieve adaptive optimization of maintenance strategies through incremental learning, thereby reducing reliance on manual decision-making.

[0109] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A customer relationship maintenance and optimization system based on contrastive learning, characterized in that, include: The data acquisition and preprocessing module is used to collect and preprocess multi-source behavioral data associated with target customers and generate standardized customer datasets. The customer representation module is used to train a standardized customer dataset using an improved MoCo-v3 contrastive learning method to generate a customer representation model. The customer status analysis module is used to perform feature encoding and similarity calculation on target customers based on the customer representation model, and output customer grouping labels, relationship urgency labels and life cycle stage labels to generate customer status information; The maintenance instruction generation module is used to determine the maintenance content and execution method based on customer status information and generate customer maintenance instructions. The model update module is used to collect actual response behavior data of target customers, update the training sample set, and perform incremental training on the customer representation model. The maintenance optimization module is used to adaptively optimize maintenance instructions for the same customer segment based on the updated customer representation model, and output the customer relationship maintenance optimization results.

2. The customer relationship maintenance and optimization system based on contrastive learning according to claim 1, characterized in that, The modules are connected in the following way: Collect multi-source behavioral data associated with target customers, preprocess it, and generate a standardized customer dataset; An improved MoCo-v3 contrastive learning method is used to train a standardized customer dataset. A customer representation model is generated through multi-dimensional feature representation construction, sample similarity constraints, and iterative updating of model parameters. Based on the customer representation model, feature encoding, similarity calculation and status labeling are performed on target customers in the pending maintenance state, and customer group labels, relationship urgency labels and life cycle stage labels are output to generate customer status information; Based on customer status information, conduct strategy evaluation and action selection, determine the maintenance content and execution method for target customers, and generate customer maintenance instructions; Execute customer maintenance instructions, collect and correlate actual response behavior data of target customers, update the training sample set, incrementally train the customer representation model, and output the updated customer representation model. Based on the updated customer representation model, maintenance instructions are generated and adaptively optimized for customers belonging to the same segment of the target customers, and the optimized customer relationship maintenance results are output.

3. The customer relationship maintenance and optimization system based on contrastive learning according to claim 2, characterized in that, The multi-source customer data includes transaction behavior data, communication record data, emotional feedback data, and social interaction data. The preprocessing includes filling missing values, normalizing, timestamp encoding, and feature alignment of the collected data.

4. The customer relationship maintenance and optimization system based on contrastive learning according to claim 2, characterized in that, The generation of the customer representation model includes the following specific steps: A dual-tower deep neural network structure is constructed based on the improved MoCo-v3 contrastive learning method. The initial parameters of the two tower encoders are initialized with the same value. Each customer sample in the standardized customer dataset is represented as a feature sequence and input into the dual-tower deep neural network structure. The improved MoCo-v3 contrastive learning method adds multi-dimensional feature perturbation processing to enhance sample discrimination on the basis of the original dual-tower momentum contrastive learning mechanism, adds a lightweight multi-head self-attention structure to optimize feature correlation modeling, and adjusts the momentum coefficient to an adaptive dynamic update method to improve parameter update stability and convergence speed. The structure includes a main tower encoder and a momentum tower encoder. For each customer sample in the standardized customer dataset, feature perturbation processing is performed. Different perturbation results are input into the main tower encoder and the momentum tower encoder respectively. The main tower encoder outputs a query vector, and the momentum tower encoder outputs a key vector. Query vectors and key vectors from the same customer sample are combined to form positive sample pairs, and query vectors and key vectors from different customer samples are combined in pairs to form negative sample pairs. Positive sample similarity is calculated between the query vector and key vector of the same customer sample, and negative sample similarity is calculated between the query vector and the key vector of different customer samples. The similarity score is calculated using the dot product similarity method. ； in, This indicates the degree of similarity between the query vector and the key vector in the same feature space. This represents the i-th component of the query vector. This represents the i-th component of the key vector; Based on similarity calculation, a contrastive learning loss is used as the optimization objective to perform backpropagation updates on the main tower encoder parameters: ； in, This represents the contrastive learning loss value. Represents the natural logarithm function. Represents an exponential function. This represents the query vector output by the main tower encoder. This represents the key vector output by the momentum tower encoder. This represents the dot product similarity of positive sample pairs. This represents the dot product similarity of negative sample pairs. Indicates the temperature coefficient; The momentum tower encoder parameters are updated using an exponential moving average strategy based on the update results of the main tower encoder parameters: ； in, This represents the momentum tower encoder parameters at the current training moment. Indicates the current training time step. Indicates the momentum coefficient. This represents the momentum tower encoder parameters at the previous training time. This represents the updated parameters of the master encoder at the current training moment. This indicates a momentum tower encoder. Indicates the main tower encoder; Repeat the above steps to iteratively train all customer samples in the standardized customer dataset until the contrastive learning loss function converges, generating a customer representation model.

5. A customer relationship maintenance and optimization system based on contrastive learning according to claim 2, characterized in that, The generation of the customer status information includes the following specific steps: The standardized customer dataset of the target customers in the unmaintained state is input into the customer representation model, and feature encoding processing is performed to output the embedding vector of the target customer. The embedding vector consists of several numerical components, each of which corresponds to the feature representation of the customer in the dimensions of behavior, interaction and transaction. Extract the embedding vectors of historical customers and calculate the distance between the target customer's embedding vector and the embedding vectors of each historical customer; Based on the distance value, the similarity between target customers and historical customers is sorted, and a set of historical customers whose distance values ​​are within a preset range is selected to generate customer segmentation tags; The degree of concentration of adjacent distance distribution is used to determine the closeness of the relationship between target customers and historical customers. When the distance variance is less than the preset variance threshold, the target customer is marked as a high relationship urgency customer. When the distance variance is greater than or equal to the preset variance threshold, the target customer is marked as a low relationship urgency customer. The corresponding relationship urgency label is output. Customer status is divided into stages based on the target customer's transaction time interval, interaction frequency, and active cycle length. Lifecycle stage markers are output, and customer group markers, relationship urgency markers, and lifecycle stage markers are combined to generate customer status information.

6. A customer relationship maintenance and optimization system based on contrastive learning according to claim 2, characterized in that, The generation of customer maintenance instructions includes: jointly analyzing customer segmentation markers, relationship urgency markers, and lifecycle stage markers in customer status information; performing weighted calculations based on the feature weights corresponding to each marker to obtain the maintenance priority of the target customer; dividing the maintenance priority results into intervals and filtering the corresponding set of strategy parameters; performing strategy matching operations for customer maintenance priorities within different intervals; determining communication methods and reach frequency parameters based on relationship urgency markers; determining resource allocation levels and discount parameters based on lifecycle stage markers; determining interaction channels and communication content parameters based on customer segmentation markers; and combining and chronologically sorting all the filtered strategy parameters to generate customer maintenance instructions that include communication methods, reach timing, resource allocation levels, and churn warning levels.

7. A customer relationship maintenance and optimization system based on contrastive learning according to claim 2, characterized in that, The updated customer representation model is obtained by: associating the collected actual response behavior data of the target customers with the corresponding customer embedding vectors; classifying and sorting the response behavior by time series to form a behavior response feature sequence; merging the feature sequence with the customer embedding vectors and state labels in the original training sample set to generate an updated training sample set; performing data standardization and feature alignment on the updated sample set; performing small-step gradient update training on batches of newly added data in the sample set; dynamically adjusting the learning rate based on the changes in the contrastive learning loss function; performing layer-by-layer backpropagation and weight update on the model parameters; maintaining the momentum smoothness of historical weights and preventing overfitting; stopping training when the loss decreases below a preset threshold in several consecutive training rounds; and outputting the updated customer representation model.

8. A customer relationship maintenance and optimization system based on contrastive learning according to claim 2, characterized in that, The output of the customer relationship maintenance optimization results includes: based on the updated customer representation model, batch processing and feature comparison of customer embedding vectors belonging to the same subgroup as the target customer; calculating the feature similarity and behavioral deviation of each customer within the subgroup; performing strategy adjustment on customer samples with deviation exceeding a preset deviation threshold; maintaining the original customer maintenance instructions for customer samples with deviation not exceeding the preset deviation threshold; adaptively updating the customer maintenance instructions for all customer samples; dynamically correcting the communication method, contact timing, resource allocation level, and churn warning level based on real-time collected actual response behavior data; performing multiple rounds of iterative verification on the optimized customer maintenance instructions; and outputting the customer relationship maintenance optimization results when the strategies in the continuous verification rounds converge and meet the preset threshold conditions.

Images