Customer value grading evaluation method and system fusing ai and behavior portrait

Abstract

The application provides a customer value grading evaluation method and system fusing AI and behavior portraits, which comprises collecting customer behavior sequences in a collection period, setting behavior conversion benchmark weights after cleaning and coding, constructing a behavior correlation matrix to quantify behavior transition probabilities, using iterative augmented training until the model converges, splitting the behavior semantic blocks according to the matrix in each round, and generating a reordered sequence by block weight mean negative correlation probability replacement; matching the same portrait sample to extract high-value segments, embedding the target sequence according to the transition probability to obtain a fusion sequence; using the original sequence, the reordered sequence and the fusion sequence to form an augmented set to train the model, extracting behavior attribution scores to update the benchmark weights, and outputting customer scores by using the converged model after iteration, and completing customer value grading in combination with a preset threshold.

Description

Technical Field

[0001] This application belongs to the field of assessment, and in particular relates to a customer value grading assessment method and system that integrates AI and behavioral profiling. Background Technology

[0002] Customer value assessment has traditionally relied heavily on static statistical indicator systems. With the widespread application of artificial intelligence, data-driven models based on users' historical behavior sequences have gradually become mainstream. In real-world business scenarios, the activity levels of different customers vary significantly, and a large amount of customer behavior data generally suffers from high sparsity, long-tail distribution, and sample imbalance. For most low-to-medium frequency customers, their behavior sequences are typically short and have limited feature dimensions. If a value prediction model is directly trained based solely on raw, sparse time-series data, it is prone to overfitting, making it difficult to fully uncover the potential business value patterns behind the behavior, thus affecting the stability of customer value classification results. Therefore, how to achieve reasonable and high-quality sample expansion based on limited interaction data has become a key issue in improving e-commerce customer value assessment capabilities.

[0003] Existing evaluation methods typically enrich training samples using sequence data augmentation techniques. Common sequence augmentation methods often employ strategies such as random masking, global permutation, or simple pruning. However, these methods often fail to fully consider the temporal dependencies and business logic of user behavior in e-commerce scenarios. Excessive random perturbation can easily disrupt behavioral chains with strong conversion correlations in the original sequence, generating a large number of noisy samples that do not conform to real shopping logic, and even misleading model training. Furthermore, most existing augmentation mechanisms are static open-loop processes, making it difficult to incorporate the weight differences in the contribution of different behaviors to customer value during the augmentation process, and also difficult to fully utilize the high-value behavioral patterns of users with similar profiles for cross-sequence feature borrowing and fusion. These sample augmentation methods, which disconnect business semantics from model feedback, make it difficult for the model to accurately identify core behavioral segments with high discriminative power, limiting the generalization ability of customer value prediction models in e-commerce environments, and consequently making it difficult for customer value classification results to meet the needs of refined operations. Summary of the Invention

[0004] To improve the quality of expanding time-series samples of sparse long-tail customers in e-commerce and the generalization performance of customer value assessment models, this invention provides a customer value classification assessment method and system that integrates AI and behavioral profiling.

[0005] The first aspect of this invention provides a customer value grading and assessment method that integrates AI and behavioral profiling, comprising the following steps: Obtain customer behavior sequence data within a specified time period and obtain customer profile tags associated with customer identifiers. The behavior sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. Clean and encode the behavior sequence data to obtain a formatted set of user behavior sequences. Set an initial value for the conversion contribution based on the behavior type. Based on the formatted set of user behavior sequences, construct a behavior association matrix that quantifies the transfer probability between behaviors. Iterative augmentation training is performed until the customer value prediction model converges. Each iteration includes dividing the target sequence into behavioral semantic blocks based on the behavioral association matrix, performing block-level permutations to generate rearranged sequences, and the permutation probability of the behavioral semantic blocks is negatively correlated with the mean of the current weights within the blocks. Source sequences with the same profile tags are identified, high-value segments are extracted based on the current weights, and the high-value segments are embedded into the target sequence according to the connection points with the highest transition probability to generate fusion sequences. The augmentation set composed of the original sequence, rearranged sequences, and fusion sequences is used to train the customer value prediction model, extract behavioral attribution importance scores, and update the baseline value weights based on the behavioral attribution importance scores for use in the next iteration. The converged customer value prediction model is used to score target customers and value classification is completed based on preset thresholds.

[0006] Optionally, the behavior sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. The behavior sequence data is cleaned and encoded to obtain a formatted set of user behavior sequences, including: Extract the occurrence timestamp features from the behavioral sequence data, and remove invalid data records with abnormal occurrence timestamps to obtain a cleaned sequence; Construct a string attribute mapping table for behavior type labels, map each behavior type to an independent monotonically increasing positive integer number, and replace the strings in the cleaning sequence with the corresponding positive integer numbers; The customer identifier is used as the dividing key for classification, and each customer subset is sorted and connected in ascending order according to the time stamp of occurrence, and combined to generate a formatted set of user behavior sequences.

[0007] Optionally, constructing a behavior correlation matrix that quantifies the transition probabilities between behaviors based on the formatted set of user behavior sequences includes: Scan the formatted set of user behavior sequences and count the total frequency of any behavior type as the first occurrence of a single preceding event. In all sequences, find the second co-occurrence frequency of the preceding single event and any subsequent target event occurring immediately after each other without any other behavioral interference; Divide the second co-occurrence frequency by the first total occurrence frequency to calculate the transition probability from the preceding event to the following event; The transition probabilities between each behavior type pair are sequentially stored in a preset two-dimensional storage space to construct the behavior association matrix.

[0008] Optionally, the step of segmenting the target sequence into behavioral semantic blocks based on the behavioral association matrix includes: Following the order in which the target sequence occurs, the adjacent first and second lines are extracted sequentially; Query and extract the current transition probability value from the first behavior to the second behavior from the behavior association matrix; The current transition probability value is compared with a preset constant judgment threshold. If the current transition probability value is less than the judgment threshold, a split point marker is defined between the first action and the second action. The target sequence is truncated according to all the defined segmentation points to generate multiple independent continuous small sequences as the behavioral semantic blocks.

[0009] Optionally, the extraction of high-value segments based on the current weights includes: A fixed-length sliding window is used to extract subsequences containing at least three consecutive behavioral events from the source sequence, which are then used as candidate behavioral fragment combinations. Obtain the value weight of each independent behavioral event within each candidate behavioral fragment combination in the current iteration; The value weights of each behavioral event within the same candidate behavioral segment combination are summed and divided by the number of behavioral events within the combination to obtain the comprehensive value score of the segment. The segment with the highest comprehensive value score among all candidate behavioral segment combinations is selected and retained as the extracted high-value segment.

[0010] Optionally, the extraction of behavior attribution importance scores includes: The augmented set is mapped to continuous embedded vector features and input into the current customer value prediction model for forward propagation to obtain the prediction score parameters returned by the model. The backpropagation algorithm is used to calculate the partial derivative vector of the predicted score parameter with respect to each of the embedded vector features, and the norm of the partial derivative vector is calculated to obtain the prediction result bias caused by a single row feature. The prediction deviation values ​​of all independent behavioral features belonging to the same behavioral category are extracted and summed, and then divided by the total frequency of the category to calculate the average deviation value. The average deviation value for each category is output independently as the corresponding behavioral attribution importance score.

[0011] Optionally, the step of scoring target customers using the converged customer value prediction model and completing value classification based on a preset threshold includes: Obtain the target customer's value score output by the converged customer value prediction model; The value score is compared with a preset first-level threshold and a second-level threshold, wherein the first-level threshold is greater than the second-level threshold. If the value score is greater than the first level threshold, the target customer is determined to belong to the core value level. If the value score is less than or equal to the first level threshold and greater than the second level threshold, the target customer is determined to belong to the potential value level. If the value score is less than or equal to the second-level threshold, the target customer is determined to belong to the marginal value level.

[0012] A second aspect of this invention provides a customer value grading and assessment system that integrates AI and behavioral profiling, comprising the following modules: The module is used to acquire behavioral sequence data of customers within a specified time period and acquire customer profile tags associated with customer identifiers. The behavioral sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. The behavioral sequence data is cleaned and encoded to obtain a formatted set of user behavior sequences. An initial value of the baseline value weight is set for the conversion contribution based on the behavior type. Based on the formatted set of user behavior sequences, a behavior association matrix that quantifies the transfer probability between behaviors is constructed. The iterative module is used to perform iterative augmentation training until the customer value prediction model converges. Each iteration includes dividing the target sequence into behavioral semantic blocks based on the behavioral association matrix, performing block-level permutations to generate rearranged sequences, and the permutation probability of the behavioral semantic blocks is negatively correlated with the mean of the current weights within the blocks. Source sequences with the same profile tags are identified, high-value segments are extracted based on the current weights, and the high-value segments are embedded into the target sequence according to the connection points with the highest transition probability to generate a fusion sequence. The augmented set composed of the original sequence, the rearranged sequence, and the fusion sequence is used to train the customer value prediction model, extract the behavioral attribution importance score, and update the baseline value weight based on the behavioral attribution importance score for the next iteration. The scoring module is used to score target customers using the converged customer value prediction model and complete the value classification based on preset thresholds.

[0013] Furthermore, the behavior sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. The behavior sequence data is cleaned and encoded to obtain a formatted set of user behavior sequences, including: Extract the occurrence timestamp features from the behavioral sequence data, and remove invalid data records with abnormal occurrence timestamps to obtain a cleaned sequence; Construct a string attribute mapping table for behavior type labels, map each behavior type to an independent monotonically increasing positive integer number, and replace the strings in the cleaning sequence with the corresponding positive integer numbers; The customer identifier is used as the dividing key for classification, and each customer subset is sorted and connected in ascending order according to the time stamp of occurrence, and combined to generate a formatted set of user behavior sequences.

[0014] Furthermore, the step of constructing a behavior correlation matrix that quantifies the transition probabilities between behaviors based on the formatted set of user behavior sequences includes: Scan the formatted set of user behavior sequences and count the total frequency of any behavior type as the first occurrence of a single preceding event. In all sequences, find the second co-occurrence frequency of the preceding single event and any subsequent target event occurring immediately after each other without any other behavioral interference; Divide the second co-occurrence frequency by the first total occurrence frequency to calculate the transition probability from the preceding event to the following event; The transition probabilities between each behavior type pair are sequentially stored in a preset two-dimensional storage space to construct the behavior association matrix.

[0015] Further, the step of segmenting the target sequence into behavioral semantic blocks based on the behavioral association matrix includes: Following the order in which the target sequence occurs, the adjacent first and second lines are extracted sequentially; Query and extract the current transition probability value from the first behavior to the second behavior from the behavior association matrix; The current transition probability value is compared with a preset constant judgment threshold. If the current transition probability value is less than the judgment threshold, a split point marker is defined between the first action and the second action. The target sequence is truncated according to all the defined segmentation points to generate multiple independent continuous small sequences as the behavioral semantic blocks.

[0016] Furthermore, the extraction of high-value segments based on the current weights includes: A fixed-length sliding window is used to extract subsequences containing at least three consecutive behavioral events from the source sequence, which are then used as candidate behavioral fragment combinations. Obtain the value weight of each independent behavioral event within each candidate behavioral fragment combination in the current iteration; The value weights of each behavioral event within the same candidate behavioral segment combination are summed and divided by the number of behavioral events within the combination to obtain the comprehensive value score of the segment. The segment with the highest comprehensive value score among all candidate behavioral segment combinations is selected and retained as the extracted high-value segment.

[0017] Furthermore, the extracted behavioral attribution importance score includes: The augmented set is mapped to continuous embedded vector features and input into the current customer value prediction model for forward propagation to obtain the prediction score parameters returned by the model. The backpropagation algorithm is used to calculate the partial derivative vector of the predicted score parameter with respect to each of the embedded vector features, and the norm of the partial derivative vector is calculated to obtain the prediction result bias caused by a single row feature. The prediction deviation values ​​of all independent behavioral features belonging to the same behavioral category are extracted and summed, and then divided by the total frequency of the category to calculate the average deviation value. The average deviation value for each category is output independently as the corresponding behavioral attribution importance score.

[0018] Furthermore, the step of using the converged customer value prediction model to score target customers and completing value classification based on preset thresholds includes: Obtain the target customer's value score output by the converged customer value prediction model; The value score is compared with a preset first-level threshold and a second-level threshold, wherein the first-level threshold is greater than the second-level threshold. If the value score is greater than the first level threshold, the target customer is determined to belong to the core value level. If the value score is less than or equal to the first level threshold and greater than the second level threshold, the target customer is determined to belong to the potential value level. If the value score is less than or equal to the second-level threshold, the target customer is determined to belong to the marginal value level.

[0019] This invention constructs a behavior correlation matrix to quantify the transition probabilities between behaviors and sets benchmark value weights, enabling the discovery of temporal correlation patterns in customer behavior and their true business contributions. During model training, an augmentation mechanism combining block-level permutation and cross-sequence high-value segment fusion is employed. Sequence reconstruction is performed based on the weight mean and the highest transition probability, expanding the high-quality sample space while maintaining the semantic coherence of behaviors. This alleviates the model overfitting problem caused by the scarcity of high-value behavior samples in e-commerce scenarios. By constructing a weight feedback update closed loop based on behavior attribution importance scores, the model can continuously optimize its perception of key business behaviors during iterative training. The synergistic effect of sequence augmentation and weight iteration mechanisms enhances the generalization ability of the customer value prediction model and improves the accuracy and reliability of target customer scoring and value classification results. Attached Figure Description

[0020] Figure 1 A flowchart of a customer value grading and assessment method that integrates AI and behavioral profiling; Figure 2 This is a diagram illustrating the probability of behavior transition. Figure 3 This is a diagram illustrating the value scores of candidate segments. Detailed Implementation

[0021] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Example 1 In Embodiment 1 of the present invention, as Figure 1 As shown, a customer value grading assessment method integrating AI and behavioral profiling includes the following steps: S1. Obtain customer behavior sequence data within a specified time period and obtain customer profile tags associated with customer identifiers. The behavior sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. Clean and encode the behavior sequence data to obtain a formatted set of user behavior sequences. Set an initial value for the conversion contribution based on the behavior type. Based on the formatted set of user behavior sequences, construct a behavior association matrix that quantifies the transfer probability between behaviors.

[0023] Extract user tracking logs from the past 90 days to obtain raw data containing user ID, behavior type, timestamp, and product ID. Simultaneously, retrieve profile features such as age stratification, spending level, activity level, membership level, or historical spending tier from customer profile tables, member information tables, or user stratification configuration tables, using user ID as the association key. Generate customer profile tags by mapping or clustering these features using rules. Remove invalid data records containing missing values. Group by user ID and sort in ascending order by timestamp. Discretize user IDs and product IDs into integers, transforming them into a formatted set of user behavior sequences readable by the model. Establish a dictionary mapping relationship, assigning initial baseline value weights to behavior types (e.g., browsing = 1, favorites = 3, adding to cart = 5, purchasing = 10). Traverse the formatted set of user behavior sequences, counting the co-occurrence frequency of adjacent behavior types to construct a two-dimensional co-occurrence count matrix. Perform Markov normalization on the co-occurrence count matrix row-wise to obtain a behavior association matrix where the sum of elements in each row is 1, quantifying the transition probability between different behaviors.

[0024] To form the supervised training objective for the customer value prediction model, the system extracts real customer business results from the label observation window after the specified time period. These real business results include at least one of the following: payment amount, number of valid orders, repurchase frequency, refund rate, gross profit contribution, or changes in membership level. The refund rate, as a negative business indicator, undergoes reverse normalization or is assigned a negative weight. The real business results are then subjected to normalized weighting or threshold discretization to generate real customer value tags. For example, the payment amount, repurchase frequency, and gross profit contribution are normalized and then weighted to obtain a comprehensive value index. When the comprehensive value index is greater than a preset high-value threshold, the corresponding customer's real customer value tag is set to 1; otherwise, it is set to 0. This real customer value tag is bound to the customer identifier and stored, and is used as the target tag for calculating the loss function in subsequent training.

[0025] In one possible embodiment, the behavior sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. The behavior sequence data is cleaned and encoded to obtain a formatted set of user behavior sequences, including: Extract the occurrence timestamp features from the behavioral sequence data, and remove invalid data records with abnormal occurrence timestamps to obtain a cleaned sequence; Construct a string attribute mapping table for behavior type labels, map each behavior type to an independent monotonically increasing positive integer number, and replace the strings in the cleaning sequence with the corresponding positive integer numbers; The customer identifier is used as the dividing key for classification, and each customer subset is sorted and connected in ascending order according to the time stamp of occurrence, and combined to generate a formatted set of user behavior sequences.

[0026] The data cleaning module iterates through all collected behavior logs, extracts the timestamp field, and verifies it against a set valid time window, such as from 00:00 on January 1, 2023 to the current system time. If the timestamp is negative, null, or exceeds the valid time window, the corresponding data row is removed, resulting in a cleaned sequence. During the encoding phase, the module automatically scans and extracts a set of deduplicated behavior type strings, establishing a static hash mapping table (e.g., browsing 1, adding to cart 2, favorites 3, purchasing 4). A global replacement operation is performed to convert long text strings into positive integer numbers, reducing memory consumption and improving matrix operation efficiency. Using a distributed computing framework, the module employs a grouping aggregation operator, using the user's unique identifier as the key, to aggregate data from the same customer onto the same node. Within each customer's subset, a fast sorting algorithm is used to sort the data in ascending order based on the timestamp values. The discrete integer numbers are then concatenated chronologically, for example, generating a one-dimensional array like 1, 1, 2, 4, resulting in a formatted set of user behavior sequences for model use.

[0027] In one possible embodiment, constructing a behavior correlation matrix that quantifies the transition probabilities between behaviors based on the formatted set of user behavior sequences includes: Scan the formatted set of user behavior sequences and count the total frequency of any behavior type as the first occurrence of a single preceding event. In all sequences, find the second co-occurrence frequency of the preceding single event and any subsequent target event occurring immediately after each other without any other behavioral interference; Divide the second co-occurrence frequency by the first total occurrence frequency to calculate the transition probability from the preceding event to the following event; The transition probabilities between each behavior type pair are sequentially stored in a preset two-dimensional storage space to construct the behavior association matrix.

[0028] Initialize a two-dimensional array of dimension N×N in memory, where the parameter N represents the total number of behavior types in the mapping table, preferably between 5 and 20; assuming N=4 for the current business. Scan each formatted user behavior sequence one by one with a sliding window of size 1. For each pair of adjacent elements, use an array index mapping mechanism to increment the total frequency of the first occurrence of the preceding event by 1 in the first-level counter, and simultaneously increment the second co-occurrence frequency of the preceding event transitioning to the following event by 1 in the second-level hash mapper. After traversing all tens of millions of sequences, divide the second co-occurrence frequency of all records by the corresponding total frequency of the first occurrence to obtain a floating-point transition probability rounded to 4 decimal places; for example, if the total frequency of the add-to-cart event is 10,000 times and the co-occurrence frequency of the purchase event is 1,500 times, then the transition probability between them is 0.1500. Write the calculated probability values ​​one by one into the corresponding row and column coordinates of the two-dimensional array to generate a behavior association matrix representing coherence. The behavior transition probability is illustrated as follows: Figure 2 As shown.

[0029] S2, perform iterative augmentation training until the customer value prediction model converges. Each iteration includes dividing the target sequence into behavioral semantic blocks based on the behavioral association matrix, performing block-level permutations to generate rearranged sequences, and the permutation probability of the behavioral semantic blocks is negatively correlated with the mean of the current weights within the block. Identify source sequences with the same profile label, extract high-value segments based on the current weights, embed the high-value segments into the target sequence according to the connection points with the highest transition probability to generate a fusion sequence, and use the original sequence, rearranged sequence, and fusion sequence to form an augmentation set to train the customer value prediction model. Extract behavioral attribution importance scores, and update the baseline value weights based on the behavioral attribution importance scores for use in the next iteration.

[0030] A sequence prediction model based on a self-attention mechanism is instantiated as the customer value prediction model. An adaptive gradient optimizer and a binary cross-entropy loss function are configured. The outer loop is set to continuously iterate until the loss value on the validation set no longer decreases for five consecutive rounds, at which point the model is considered to have converged. The binary cross-entropy loss function takes the predicted score parameters output by the customer value prediction model and the real customer value labels as inputs, and is used to measure the deviation between the model's prediction results and the real customer value labels.

[0031] In each iteration, the target sequence is traversed, and the behavior association matrix is ​​queried. When the transition probability between two adjacent behaviors is less than 0.1, the sequence is broken at that position using a sequence truncation operation, dividing the sequence into multiple behavior semantic blocks. The arithmetic mean of the current weights of the behavior types contained in each behavior semantic block is calculated, and this mean is converted into a permutation probability parameter inversely proportional to its own size. Several semantic blocks are extracted according to this permutation probability, and the order of the semantic blocks is randomly shuffled and then reassembled to generate a rearranged sequence. The user's age, gender, and consumption level profile tags are clustered, and other user behavior sequences that are consistent with the target customer's real customer value tags or whose comprehensive value index difference is within a preset range are selected from the same cluster as the source sequence.

[0032] A sliding window of size five is used to traverse the source sequence, calculating the sum of the current weights of the behaviors within the window, and selecting the subsequence with the largest sum as the high-value segment. The behavior association matrix is ​​queried, and the transition probability products of the first and last behaviors of the high-value segment and each potential insertion point in the target sequence are calculated. The optimal connection point with the largest product is located, and the high-value segment is embedded into this optimal connection point to generate a fused sequence. The original sequence, rearranged sequence, and fused sequence are merged to construct batch data, which is then input into the customer value prediction model for forward computation and backpropagation to update the network parameters. The rearranged sequence and fused sequence serve as augmented samples for the target sequence, inheriting the true customer value labels corresponding to the customers in the target sequence. During training, the original sequence, rearranged sequence, and fused sequence are paired with the true customer value labels to form training sample pairs, ensuring that the augmented samples can participate in the supervised loss calculation. The gradient contribution of each behavior embedding vector in the input sequence to the prediction score parameters is calculated, and the norm of the gradient vector is taken as the behavior attribution importance score. Calculate the global mean of importance scores for the same behavior type, and normalize the global mean to a preset value range that is the same as the benchmark value weight. Using the exponential moving average (EMA) algorithm, the normalized global mean is weighted and summed with the benchmark value weight of the previous round at an update step size of 0.1 to obtain a new weight value, which is then passed to the next iteration.

[0033] In one possible embodiment, segmenting the target sequence into behavioral semantic blocks based on the behavioral association matrix includes: Following the order in which the target sequence occurs, the adjacent first and second lines are extracted sequentially; Query and extract the current transition probability value from the first behavior to the second behavior from the behavior association matrix; The current transition probability value is compared with a preset constant judgment threshold. If the current transition probability value is less than the judgment threshold, a split point marker is defined between the first action and the second action. The target sequence is truncated according to all the defined segmentation points to generate multiple independent continuous small sequences as the behavioral semantic blocks.

[0034] To identify natural breakpoints in the user behavior flow, starting from the beginning of the target sequence, the current behavior combination (i.e., the first and second behaviors) is extracted sequentially using two pointers. For this pair of integer encoded combinations, the encoded combinations are input as two-dimensional row and column indices into a pre-generated behavior association matrix to retrieve the current transition probability value in floating-point format. A preset constant judgment threshold is retrieved from the configuration library; this threshold is preferably set between 0.05 and 0.20, for example, a typical empirical value of 0.10. The retrieved current transition probability value is compared with 0.10 in a floating-point manner. If the retrieved probability value is only 0.03, lower than the judgment threshold of 0.10, it indicates a logical break between the two behaviors, and a split point marker is added at the corresponding index position in the original sequence. After the entire long sequence has been scanned, the sequence processing engine will use the sequence slicing method to truncate the original one-dimensional long array from the split point according to all the recorded split point identifiers, generating several small sequence arrays of varying lengths, such as 2 to 5 elements, with highly coherent internal logic. These small sequence outputs serve as behavioral semantic blocks for local substitution.

[0035] In one possible embodiment, the extraction of high-value segments based on the current weights includes: A fixed-length sliding window is used to extract subsequences containing at least three consecutive behavioral events from the source sequence, which are then used as candidate behavioral fragment combinations. Obtain the value weight of each independent behavioral event within each candidate behavioral fragment combination in the current iteration; The value weights of each behavioral event within the same candidate behavioral segment combination are summed and divided by the number of behavioral events within the combination to obtain the comprehensive value score of the segment. The segment with the highest comprehensive value score among all candidate behavioral segment combinations is selected and retained as the extracted high-value segment.

[0036] A fixed-length sliding window of length L is constructed, with the optimal range of parameter L being 3 to 5; in this example, L=3 is set. The window starts from index 0 of the source sequence and moves to the right, selecting subarrays consisting of three consecutive adjacent behaviors, such as browsing, adding to cart, and purchasing. These subarrays are then stored in a candidate pool to form candidate behavior fragment combinations. For each extracted candidate combination, the query engine retrieves the feature weight table updated in the current iteration to obtain the independent value weights for these three behavior events; for example, the current browsing weight is 0.2, adding to cart weight is 0.6, and purchasing weight is 1.5. The sum of these local weights is calculated as 2.3, and the result is divided by the fixed number of events in the combination (3) to calculate the average value of the fragment, 0.767, as the comprehensive value score. A maximum value selection strategy is used, comparing the comprehensive value scores of all candidate fragment combinations. The highest-scoring consecutive subsequence containing three elements is selected and copied, cached in an independent memory area, and retained as a high-value fragment for output. The candidate fragment value score is illustrated below. Figure 3 As shown.

[0037] In one possible embodiment, the extraction of behavioral attribution importance scores includes: The augmented set is mapped to continuous embedded vector features and input into the current customer value prediction model for forward propagation to obtain the prediction score parameters returned by the model. The backpropagation algorithm is used to calculate the partial derivative vector of the predicted score parameter with respect to each of the embedded vector features, and the norm of the partial derivative vector is calculated to obtain the prediction result bias caused by a single row feature. The prediction deviation values ​​of all independent behavioral features belonging to the same behavioral category are extracted and summed, and then divided by the total frequency of the category to calculate the average deviation value. The average deviation value for each category is output independently as the corresponding behavioral attribution importance score.

[0038] The current customer value prediction model is a deep neural network based on a self-attention mechanism. The input to the prediction model is a discrete sequence of user behavior IDs, and the output is a customer value prediction probability score between 0 and 1. The prediction model includes an embedding layer, a multi-head self-attention layer, a feedforward neural network layer, a pooling layer, and a fully connected output layer connected in sequence. Specifically, the embedding layer converts the discrete user behavior ID sequence into a continuous floating-point embedding vector feature matrix, so that each behavior event corresponds to a learnable vector representation. The multi-head self-attention layer calculates the degree of correlation between different behavior events and their contextual relationships, and generates a sequence feature representation that integrates global behavior dependencies. The feedforward neural network layer performs nonlinear transformations and feature enhancements on the sequence feature representation, and residual connections and normalization are used to maintain training stability. Subsequently, the pooling layer summarizes the sequence features from multiple time points to obtain a customer-level feature vector representing the overall behavioral characteristics of the target customer. Finally, the fully connected output layer maps the customer-level feature vector and normalizes it using the Sigmoid activation function to obtain a prediction score parameter between 0 and 1, which serves as the customer value prediction probability score.

[0039] The integer sequence within the augmented set is mapped to continuous floating-point embedding vector features, typically between 32 and 128 dimensions (e.g., 64 dimensions), via a lookup table in the embedding layer. These features are then input into the customer value prediction model described above for forward propagation calculations. The model returns predicted score parameters at the output layer. During the training phase, the real customer value label corresponding to the current customer identifier is read. The predicted score parameters and the real customer value label are input into the binary cross-entropy loss function to obtain the training loss, and backpropagation is performed to update the model parameters. To trace the contribution of behavior in reverse, the automatic differentiation engine is activated to execute the backpropagation algorithm. The gradient partial derivative of the predicted score with respect to each input feature vector is calculated using the chain rule. The L2 norm of this partial derivative vector is calculated, which is the square root of the sum of the squares of each element, yielding a scalar value of, for example, 0.018. This scalar value is set as the deviation of the single-behavior feature prediction result. For all events in the current training batch with the same category label (e.g., all being purchase behaviors), the independent bias values ​​are summed and divided by the total frequency of events in that batch to obtain the average bias value. This average bias value is then independently and persistently stored as the attribution importance score for that behavioral feature. The behavioral attribution importance score is extracted using a gradient norm attribution method to maintain consistency between the attribution calculation process and the aforementioned partial derivative norm step.

[0040] S3 uses the converged customer value prediction model to score target customers and completes value classification based on preset thresholds.

[0041] The model inference mode switching interface is invoked to switch the trained customer value prediction model to inference evaluation mode, and the automatic differentiation mechanism is disabled. The formatted user behavior sequence tensor of the target customer to be predicted is input into the model. After passing through a multi-head attention mechanism layer and a multi-layer perceptron layer, the Sigmoid activation function maps the output to a floating-point number between 0 and 1, which serves as the customer value score. The customer value score is discretized according to a list of boundary values ​​defined by the business. For example, scores less than or equal to 0.40 are classified as marginal value levels, scores greater than 0.40 and less than or equal to 0.80 are classified as potential value levels, and scores greater than 0.80 are classified as core value levels, thus completing the value classification. Customer identifiers and value classification labels are assembled into a data table and imported into an in-memory data storage system for downstream marketing systems to read.

[0042] In one possible embodiment, the step of scoring target customers using a converged customer value prediction model and completing value classification based on a preset threshold includes: Obtain the target customer's value score output by the converged customer value prediction model; The value score is compared with a preset first-level threshold and a second-level threshold, wherein the first-level threshold is greater than the second-level threshold. If the value score is greater than the first level threshold, the target customer is determined to belong to the core value level. If the value score is less than or equal to the first level threshold and greater than the second level threshold, the target customer is determined to belong to the potential value level. If the value score is less than or equal to the second-level threshold, the target customer is determined to belong to the marginal value level.

[0043] Once the model's loss function reaches the convergence criterion set early stopping, the deployed online inference module will feed the real-time captured target customer sequence data into the customer value prediction model. Through forward inference, it will calculate the user's value score within the 0-1 range. Simultaneously, the business configuration center has pre-injected global hyperparameters for rating interception: the first-level threshold is preferably set between 0.75 and 0.90 (e.g., a hard-coded value of 0.80), and the second-level threshold is preferably set between 0.30 and 0.50 (e.g., a hard-coded value of 0.40). The system executes nested logical validation branches: If the target customer's score is 0.85, meeting the condition of being greater than 0.80, the rule engine classifies the customer into the core value level and allocates the highest marketing budget; if the target's score is 0.60, falling within the range of greater than 0.40 and less than or equal to 0.80, the target is marked as belonging to the potential value level, and new customer acquisition resources are appropriately allocated; conversely, if the score drops to 0.20, less than or equal to 0.40, the engine immediately determines that the user belongs to the marginal value level, executes the churn warning strategy, and completes the complete business loop from quantitative scoring to qualitative classification.

[0044] Example 2 Embodiment 2 of the present invention proposes a customer value grading and assessment system that integrates AI and behavioral profiling, including the following modules: The module is used to acquire behavioral sequence data of customers within a specified time period and acquire customer profile tags associated with customer identifiers. The behavioral sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. The behavioral sequence data is cleaned and encoded to obtain a formatted set of user behavior sequences. An initial value of the baseline value weight is set for the conversion contribution based on the behavior type. Based on the formatted set of user behavior sequences, a behavior association matrix that quantifies the transfer probability between behaviors is constructed. The iterative module is used to perform iterative augmentation training until the customer value prediction model converges. Each iteration includes dividing the target sequence into behavioral semantic blocks based on the behavioral association matrix, performing block-level permutations to generate rearranged sequences, and the permutation probability of the behavioral semantic blocks is negatively correlated with the mean of the current weights within the blocks. Source sequences with the same profile tags are identified, high-value segments are extracted based on the current weights, and the high-value segments are embedded into the target sequence according to the connection points with the highest transition probability to generate a fusion sequence. The augmented set composed of the original sequence, the rearranged sequence, and the fusion sequence is used to train the customer value prediction model, extract the behavioral attribution importance score, and update the baseline value weight based on the behavioral attribution importance score for the next iteration. The scoring module is used to score target customers using the converged customer value prediction model and complete the value classification based on preset thresholds.

[0045] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A customer value grading and assessment method integrating AI and behavioral profiling, characterized in that, Includes the following steps: Obtain customer behavior sequence data within a specified time period and obtain customer profile tags associated with customer identifiers. The behavior sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. Clean and encode the behavior sequence data to obtain a formatted set of user behavior sequences. Set an initial value for the conversion contribution based on the behavior type. Based on the formatted set of user behavior sequences, construct a behavior association matrix that quantifies the transfer probability between behaviors. Iterative augmentation training is performed until the customer value prediction model converges. Each iteration includes dividing the target sequence into behavioral semantic blocks based on the behavioral association matrix, performing block-level permutations to generate rearranged sequences, and the permutation probability of the behavioral semantic blocks is negatively correlated with the mean of the current weights within the blocks. Source sequences with the same profile tags are identified, high-value segments are extracted based on the current weights, and the high-value segments are embedded into the target sequence according to the connection points with the highest transition probability to generate fusion sequences. The augmentation set composed of the original sequence, rearranged sequences, and fusion sequences is used to train the customer value prediction model, extract behavioral attribution importance scores, and update the baseline value weights based on the behavioral attribution importance scores for use in the next iteration. The converged customer value prediction model is used to score target customers and value classification is completed based on preset thresholds.

2. The method according to claim 1, characterized in that, The behavior sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. The behavior sequence data is cleaned and encoded to obtain a formatted set of user behavior sequences, including: Extract the occurrence timestamp features from the behavioral sequence data, and remove invalid data records with abnormal occurrence timestamps to obtain a cleaned sequence; Construct a string attribute mapping table for behavior type labels, map each behavior type to an independent monotonically increasing positive integer number, and replace the strings in the cleaning sequence with the corresponding positive integer numbers; The customer identifier is used as the dividing key for classification, and each customer subset is sorted and connected in ascending order according to the time stamp of occurrence, and combined to generate a formatted set of user behavior sequences.

3. The method according to claim 2, characterized in that, The process of constructing a behavior correlation matrix that quantifies the transition probabilities between behaviors based on the formatted set of user behavior sequences includes: Scan the formatted set of user behavior sequences and count the total frequency of any behavior type as the first occurrence of a single preceding event. In all sequences, find the second co-occurrence frequency of the preceding single event and any subsequent target event occurring immediately after each other without any other behavioral interference; Divide the second co-occurrence frequency by the first total occurrence frequency to calculate the transition probability from the preceding event to the following event; The transition probabilities between each behavior type pair are sequentially stored in a preset two-dimensional storage space to construct the behavior association matrix.

4. The method according to claim 1, characterized in that, The step of segmenting the target sequence into behavioral semantic blocks based on the behavioral association matrix includes: Following the order in which the target sequence occurs, the adjacent first and second lines are extracted sequentially; Query and extract the current transition probability value from the first behavior to the second behavior from the behavior association matrix; The current transition probability value is compared with a preset constant judgment threshold. If the current transition probability value is less than the judgment threshold, a split point marker is defined between the first action and the second action. The target sequence is truncated according to all the defined segmentation points to generate multiple independent continuous small sequences as the behavioral semantic blocks.

5. The method according to claim 1 or 4, characterized in that, The extraction of high-value segments based on the current weights includes: A fixed-length sliding window is used to extract subsequences containing at least three consecutive behavioral events from the source sequence, which are then used as candidate behavioral fragment combinations. Obtain the value weight of each independent behavioral event within each candidate behavioral fragment combination in the current iteration; The value weights of each behavioral event within the same candidate behavioral segment combination are summed and divided by the number of behavioral events within the combination to obtain the comprehensive value score of the segment. The segment with the highest comprehensive value score among all candidate behavioral segment combinations is selected and retained as the extracted high-value segment.

6. The method according to claim 1, characterized in that, The extracted behavior attribution importance score includes: The augmented set is mapped to continuous embedded vector features and input into the current customer value prediction model for forward propagation to obtain the prediction score parameters returned by the model. The backpropagation algorithm is used to calculate the partial derivative vector of the predicted score parameter with respect to each of the embedded vector features, and the norm of the partial derivative vector is calculated to obtain the prediction result bias caused by a single row feature. The prediction deviation values ​​of all independent behavioral features belonging to the same behavioral category are extracted and summed, and then divided by the total frequency of the category to calculate the average deviation value. The average deviation value of each category is output independently as the corresponding behavioral attribution importance score.

7. The method according to claim 1, characterized in that, The step of using a converged customer value prediction model to score target customers and completing value classification based on preset thresholds includes: Obtain the target customer's value score output by the converged customer value prediction model; The value score is compared with a preset first-level threshold and a second-level threshold, wherein the first-level threshold is greater than the second-level threshold. If the value score is greater than the first level threshold, the target customer is determined to belong to the core value level. If the value score is less than or equal to the first level threshold and greater than the second level threshold, the target customer is determined to belong to the potential value level. If the value score is less than or equal to the second-level threshold, the target customer is determined to belong to the marginal value level.

8. A customer value grading and assessment system integrating AI and behavioral profiling, characterized in that, Includes the following modules: The module is used to acquire behavioral sequence data of customers within a specified time period and acquire customer profile tags associated with customer identifiers. The behavioral sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. The behavioral sequence data is cleaned and encoded to obtain a formatted set of user behavior sequences. An initial value of the baseline value weight is set for the conversion contribution based on the behavior type. Based on the formatted set of user behavior sequences, a behavior association matrix that quantifies the transfer probability between behaviors is constructed. The iterative module is used to perform iterative augmentation training until the customer value prediction model converges. Each iteration includes dividing the target sequence into behavioral semantic blocks based on the behavioral association matrix, performing block-level permutations to generate rearranged sequences, and the permutation probability of the behavioral semantic blocks is negatively correlated with the mean of the current weights within the blocks. Source sequences with the same profile tags are identified, high-value segments are extracted based on the current weights, and the high-value segments are embedded into the target sequence according to the connection points with the highest transition probability to generate fusion sequences. The augmented set composed of the original sequence, rearranged sequences, and fusion sequences is used to train the customer value prediction model, extract behavioral attribution importance scores, and update the baseline value weights based on the behavioral attribution importance scores for use in the next iteration. The scoring module is used to score target customers using the converged customer value prediction model and complete the value classification based on preset thresholds.

9. The system according to claim 8, characterized in that, The behavior sequence data includes customer identifier, behavior type, occurrence timestamp, and associated object identifier. The behavior sequence data is cleaned and encoded to obtain a formatted set of user behavior sequences, including: Extract the occurrence timestamp features from the behavioral sequence data, and remove invalid data records with abnormal occurrence timestamps to obtain a cleaned sequence; Construct a string attribute mapping table for behavior type labels, map each behavior type to an independent monotonically increasing positive integer number, and replace the strings in the cleaning sequence with the corresponding positive integer numbers; The customer identifier is used as the dividing key for classification, and each customer subset is sorted and connected in ascending order according to the time stamp of occurrence, and combined to generate a formatted set of user behavior sequences.

10. The system according to claim 8, characterized in that, The process of constructing a behavior correlation matrix that quantifies the transition probabilities between behaviors based on the formatted set of user behavior sequences includes: Scan the formatted set of user behavior sequences and count the total frequency of any behavior type as the first occurrence of a single preceding event. In all sequences, find the second co-occurrence frequency of the preceding single event and any subsequent target event occurring immediately after each other without any other behavioral interference; Divide the second co-occurrence frequency by the first total occurrence frequency to calculate the transition probability from the preceding event to the following event; The transition probabilities between each behavior type pair are sequentially stored in a preset two-dimensional storage space to construct the behavior association matrix.

Images