Analysis methods, devices, equipment and media for screening target customer groups

Abstract

The present invention discloses an analysis method, device, equipment and medium for screening target customer groups, and relates to the technical field of customer data analysis. The method includes: constructing the registered user behavior data into a triple graph structure including user nodes, behavior nodes, and entity nodes, and injecting domain knowledge for semantic enhancement; dividing 24 hours into periodic time periods, generating time period embedding vectors, and modeling cross-time period behavior dependencies through time period-aware Transformer. The present invention constructs a triple graph structure and injects domain knowledge, combines periodic time period modeling with multimodal feature fusion, and deeply explores behavioral semantic associations and temporal dependencies to achieve the effect of capturing users' implicit needs; optimizes customer stratification through local multi-source data cross-validation, clarifies the four-level division and associates reinforcement learning with dynamic threshold decision-making to achieve the effect of improving stratification accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of customer data analysis, and in particular to an analysis method, device, equipment and medium for screening target customer groups. Background Art

[0002] Traditional user behavior analysis primarily relies on statistics from single behavioral sequences, lacking in-depth exploration of semantic connections and temporal dependencies between behaviors. For example, focusing solely on superficial metrics like browsing times and purchase frequency fails to understand the underlying intentions and needs of users across different behaviors. For example, browsing a competitor's product but not purchasing it may indicate a desire to wait for a promotion. This isolated behavioral analysis approach results in a superficial understanding of user needs, making it difficult to identify deeper business opportunities.

[0003] Existing customer segmentation methods often rely on static labels such as age, gender, and region, or simple activity metrics like login times and spending amounts. These methods ignore the dynamic nature of behavior and its potential commercial value. Furthermore, they fail to fully leverage local, multi-source data, such as cross-platform behavior and historical interaction data, for cross-validation, resulting in inaccurate segmentation results. Furthermore, customer segmentation results often exist as isolated steps, failing to effectively guide subsequent model parameter optimization and decision-making, creating logical gaps.

[0004] In addition, traditional models mostly use fixed thresholds to judge customer intentions, and do not design differentiated strategies for different stratifications, resulting in missed judgments of high-value customers or misjudgment of low-value customers; and the model parameter optimization is disconnected from the customer stratification results, and cannot be dynamically adjusted according to the business objectives of different levels, resulting in poor adaptability. Summary of the Invention

[0005] The object of the present invention is to provide an analysis method, device, equipment and medium for screening target customer groups, so as to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an analysis method for screening target customer groups, comprising the following steps:

[0007] The registered user behavior data is constructed into a triple graph structure containing user nodes, behavior nodes, and entity nodes, and domain knowledge is injected for semantic enhancement;

[0008] Divide 24 hours into periodic periods, generate period embedding vectors, and use period-aware Transformers to model cross-period behavioral dependencies.

[0009] Fuse behavior timing, entity semantics, and time period features to generate a comprehensive feature vector;

[0010] Calculate behavioral heat value, combine potential business value and potential business value offset to perform initial two-dimensional clustering to obtain preliminary customer stratification; introduce local multi-source data for cross-validation. Local multi-source data includes local cross-platform user behavior data and historical interaction data. Through data alignment and feature complementarity, optimize the preliminary customer stratification to obtain the final customer stratification results;

[0011] Based on the final customer stratification results, a state space containing business indicators for each stratification, an action space for parameter adjustment, and a reward function are constructed. Reinforcement learning parameters are then obtained by optimizing model parameters in real time through reinforcement learning.

[0012] Based on the final customer stratification results, behavioral heat value, offset of potential commercial value and reinforcement learning parameters, differentiated dynamic thresholds are used for different stratifications to output the probability of potential customer intention.

[0013] Preferably, the registered user behavior data is constructed into a triple graph structure including user nodes, behavior nodes, and entity nodes, and domain knowledge is injected for semantic enhancement, including:

[0014] Behavior nodes represent behavior types, and entity nodes are associated with category-level knowledge;

[0015] Semantic enhancement is achieved through graph convolutional neural networks, which inject domain knowledge about product category hierarchical relationships and interactive funnel conversion paths, aggregate the features of nodes and their adjacent nodes, and update node features based on the node's own features.

[0016] During the aggregation process, the node degree is regularized, including adjusting the feature weight according to the degree of the node and its adjacent nodes to balance the feature influence of nodes with different degrees and achieve semantic enhancement of node features.

[0017] Preferably, dividing 24 hours into periodic periods, generating period embedding vectors, and modeling cross-period behavior dependencies through period-aware Transformer include:

[0018] Process the time information in the behavioral data and divide 24 hours into several periodic periods;

[0019] For each behavioral event, an encoding method including sine and cosine functions is used to generate the corresponding time period embedding vector to capture the periodic characteristics of time;

[0020] Determine whether the dimension of the time period embedding vector matches the preset processing dimension. If so, it is directly input into the time period aware Transformer. If not, it is adjusted to the preset processing dimension through feature mapping and then input into the time period aware Transformer.

[0021] The behavior sequence is processed through the time-period-aware Transformer, cross-time behavior dependencies are modeled, and changes in users' behavior patterns in different time periods are captured.

[0022] Preferably, the fusion of behavior temporal sequence, entity semantics and time period features to generate a comprehensive feature vector includes:

[0023] Based on the triple graph structure and time period embedding vector, the temporal features of behaviors are extracted through a bidirectional long short-term memory network. The temporal features reflect the order and correlation of behaviors.

[0024] Extract entity semantic features and time period embedding features output by the graph convolutional neural network. Entity semantic features reflect the hierarchical association features of product categories.

[0025] After the temporal features, entity semantic features, and time period embedding features are spliced ​​together, they are input into a multi-layer perceptron for multimodal fusion to generate a comprehensive feature vector.

[0026] Preferably, the calculated behavior thermal value is combined with the potential commercial value and the offset of the potential commercial value to perform initial two-dimensional clustering to obtain preliminary customer stratification; local multi-source data is introduced for cross-validation, and the preliminary customer stratification is optimized through data alignment and feature complementarity to obtain the final customer stratification results, including:

[0027] The behavior heat value is calculated using the comprehensive feature vector. The behavior heat value comprehensively considers the time decay factor, the behavior conversion value weight, and the entity node semantic association score. The time decay factor reflects that recent behaviors are more important than long-term behaviors. The behavior conversion value weight reflects the difference in conversion value between different behaviors. The entity node semantic association score reflects the correlation between behaviors in the same category.

[0028] Calculate the user's potential commercial value and the offset of the potential commercial value. The potential commercial value is predicted based on historical consumption data, and the offset of the potential commercial value is obtained through knowledge graph reasoning, reflecting the deviation between the user's actual value and the predicted value.

[0029] Using behavioral heat value as the density indicator and potential commercial value as the value indicator, two-dimensional clustering is performed to obtain preliminary customer stratification. The preliminary customer stratification includes four levels: the high-activity and high-value level, where both the behavioral heat value and the potential commercial value are higher than the corresponding cluster center; the high-activity and low-value level, where the behavioral heat value is higher than the cluster center and the potential commercial value is lower than the cluster center; the low-activity and high-value level, where the behavioral heat value is lower than the cluster center and the potential commercial value is higher than the cluster center; and the low-activity and low-value level, where both the behavioral heat value and the potential commercial value are lower than the corresponding cluster center.

[0030] Introducing local multi-source data, which includes user behavior data and historical interaction data on different local platforms;

[0031] Perform format alignment and time dimension calibration on local multi-source data and the data on which the initial stratification is based. Format alignment includes unifying the timestamp format and behavior type definition, and time dimension calibration includes matching behaviors within the same time window.

[0032] Complementary features are extracted from local multi-source data. The complementary features include cross-platform behavioral preferences and are cross-validated with the preliminary customer stratification results: if a user's level in the preliminary stratification is consistent with the level reflected by the complementary features, the stratification result is retained; if not, the user's stratification level is adjusted in combination with the complementary features to obtain the final customer stratification result including the above four levels.

[0033] Preferably, the method of constructing a state space including business indicators of each stratification, an action space for parameter adjustment, and a reward function based on the final customer stratification result, and optimizing the model parameters in real time through reinforcement learning to obtain reinforcement learning parameters, includes:

[0034] Construct a state space that includes the precision, recall, and marketing ROI business indicators for each level in the final customer stratification results;

[0035] Construct an action space, which includes dynamic threshold parameters, behavior weights, and time segmentation granularity that can be adjusted for different layers.

[0036] Design a business value-oriented reward function based on the comprehensive calculation of the performance of each tier, including the return on investment of the high-value tier, the difference in F1 scores between different tiers, and the cost savings;

[0037] The reinforcement learning algorithm optimizes model parameters in real time, forming a self-evolutionary closed loop of "layered result feedback - parameter adjustment - layered accuracy improvement". The reinforcement learning algorithm includes a proximal strategy optimization algorithm.

[0038] Determine whether the reward value of reinforcement learning is lower than the preset performance threshold for a preset number of consecutive times. If so, trigger the parameter reset mechanism, which includes restoring to the historical optimal parameters; otherwise, continue iterative optimization;

[0039] Output optimized reinforcement learning parameters, which include dynamic threshold weight, behavior weight and time period division granularity parameters.

[0040] Preferably, the method of outputting the potential customer intention probability by adopting differentiated dynamic thresholds for different stratifications based on the final customer stratification results, behavioral heat value, potential commercial value offset, and reinforcement learning parameters includes:

[0041] Differentiated dynamic threshold calculation logic is set for each of the four levels in the final customer stratification results.

[0042] For high-activity, high-value tiers, dynamic threshold calculations focus on the offset of potential commercial value; for high-activity, low-value tiers, dynamic threshold calculations balance the behavioral heat value and the offset of potential commercial value; for low-activity, high-value tiers, dynamic threshold calculations reduce the weight of behavioral heat value; for low-activity, low-value tiers, dynamic threshold calculations increase the weight of behavioral heat value;

[0043] Based on differentiated calculation logic, the dynamic threshold is calculated by combining behavioral heat value, potential business value offset, and reinforcement learning parameters. The matching degree between the dynamic threshold and the comprehensive feature vector is then used to output the potential customer's intention probability. The higher the matching degree, the higher the intention probability. The intention probability of customers within the same tier is judged using a unified standard.

[0044] Among them, the calculation factors of dynamic thresholds include behavioral heat value, offset of potential commercial value and reinforcement learning parameters; the differentiation of dynamic threshold calculation logic is reflected in the weight distribution of each factor, and the weight distribution is determined based on the business objectives of each level.

[0045] The present invention also provides an analysis device for screening target customer groups, comprising:

[0046] The graph modeling module is used to construct the registered user behavior data into a triple graph structure containing user nodes, behavior nodes, and entity nodes, and inject domain knowledge for semantic enhancement;

[0047] The time series analysis module is used to divide the 24-hour period into periodic periods, generate period embedding vectors, and use the period-aware Transformer to model cross-period behavioral dependencies;

[0048] Feature fusion module, used to fuse behavior temporal features, entity semantic features, and time segment features to generate a comprehensive feature vector;

[0049] The stratification optimization module is used to calculate behavioral heat value and perform initial two-dimensional clustering based on potential business value and potential business value offset to obtain preliminary customer stratification. Local multi-source data is introduced for cross-validation, and the preliminary customer stratification is optimized through data alignment and feature complementarity to obtain the final customer stratification results.

[0050] The reinforcement learning module is used to construct a state space containing business indicators for each stratification, an action space for parameter adjustment, and a reward function based on the final customer stratification results. The reinforcement learning parameters are obtained by optimizing the model parameters in real time through reinforcement learning.

[0051] The prediction decision module is used to output the probability of potential customer intention based on the final customer stratification results, behavioral heat value, offset of potential commercial value and reinforcement learning parameters, using differentiated dynamic thresholds for different stratifications.

[0052] The present invention further provides an electronic device, wherein the electronic device is a physical device and comprises:

[0053] a processor and a memory, wherein the memory is communicatively connected to the processor;

[0054] The memory is used to store at least one executable instruction executed by the processor, and the processor is used to execute the executable instruction to implement the above-mentioned analysis method for screening the target customer group.

[0055] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the computer program implements the above-mentioned analysis method for screening target customer groups.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] By constructing a triplet graph structure and injecting domain knowledge, combining periodic time period modeling with multimodal feature fusion, we deeply explore behavioral semantic associations and temporal dependencies to capture users' implicit needs; through local multi-source data cross-validation, we optimize customer stratification, clarify the four-level division and associate reinforcement learning with dynamic threshold decision-making to improve stratification accuracy; through reinforcement learning, we form a self-evolutionary closed loop and differentiated dynamic threshold design to enhance model adaptability, reduce missed high-value customers and misjudgment of low-value customers, and ultimately improve target customer screening efficiency and marketing ROI. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A main flow chart of an analysis method for screening target customer groups provided by an embodiment of the present invention;

[0059] Figure 2 A schematic structural diagram of an analysis device for screening target customer groups provided by an embodiment of the present invention;

[0060] Figure 3 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0062] The execution subject of the method of this embodiment is a terminal, which can be a mobile phone, tablet computer, PDA, notebook or desktop computer, etc. Of course, it can also be other devices with similar functions, and this embodiment is not limited.

[0063] See also Figure 1 The present invention provides an analysis method for screening target customer groups. The method is based on the behavioral data of registered users. The behavioral data includes user behavioral event information. A behavioral event refers to a specific behavioral operation performed by a user on a specific entity at a specific time, and includes information such as the behavior type, associated entity, timestamp, and interaction duration, including:

[0064] Step 100, heterogeneous behavior semantic graph modeling: construct the registered user behavior data into a triple graph structure containing user nodes, behavior nodes, and entity nodes, and inject domain knowledge for semantic enhancement.

[0065] Specifically, step 100 includes:

[0066] Step 110, the behavior node represents the behavior type, and the entity node is associated with the category level knowledge;

[0067] Step 120: semantic enhancement is achieved through a graph convolutional neural network, which includes injecting domain knowledge about the hierarchical relationship between product categories and the interactive funnel conversion path, aggregating the features of the node and its adjacent nodes, and updating the node features based on the node's own features.

[0068] Step 130 , during the aggregation process, regularization is performed on the node degrees, including adjusting feature weights according to the degrees of the node and its adjacent nodes to balance the feature influences of nodes with different degrees and achieve semantic enhancement of node features.

[0069] The node types defined in the triple graph structure include:

[0070] Behavior nodes (browse / add to cart / repurchase / comment, with behavior duration attributes);

[0071] Entity node (product ID / page type / interactive component, associated category level knowledge);

[0072] User node (registered user ID, associated registration time and other basic time features);

[0073] The type is defined as:

[0074] Edge=(user node, behavior type, entity node, timestamp);

[0075] For example: (U001, browse, 3C category page, 2024-01-01 10:00) → (U001, add to cart, laptop A, 2024-01-01 10:05).

[0076] The domain knowledge injection layer accesses the product category knowledge graph (e.g., "laptops → 3C digital products → electronic products") and the interaction funnel model (e.g., "browse → add to cart → pay" conversion path);

[0077] The behavior node represents the behavior type, and the entity node is associated with category-level knowledge;

[0078] The semantic enhancement includes injecting domain knowledge of product category hierarchical relationships and interactive funnel conversion paths through a graph convolutional neural network to achieve semantic enhancement of node features. When calculating node embedding through a graph convolutional neural network, a degree-regularized adjacent node feature aggregation method is adopted. The calculation formula is:

[0079] ;

[0080] Where, For nodes In the graph convolutional neural network The feature vector of the layer (output features), is an activation function (such as ReLU, Sigmoid, etc.) used to introduce nonlinearity. For nodes All adjacent nodes of (ie with directly connected nodes), For nodes The degree of The number of connected edges), used for degree regularization, For adjacent nodes The degree, For the The weight matrix of the layer, used for feature transformation, For adjacent nodes In the The feature vector of the layer (input features), For the The bias matrix of the layer is used to adjust the feature offset, For nodes In the The layer's own feature vector (preserving the self-loop connection features).

[0081] By aggregating the features of adjacent nodes (weighted average) and combining them with their own features, the feature vector of node v is updated to achieve semantic enhancement.

[0082] Step 200, periodic time period attention mechanism: Divide 24 hours into periodic time periods, generate time period embedding vectors, and use the time period-aware Transformer to model cross-time period behavior dependencies.

[0083] Specifically, step 200 includes:

[0084] Step 210: Process the time information in the behavior data and divide 24 hours into several periodic time periods;

[0085] Step 220 , for each behavioral event, generate a corresponding time period embedding vector using an encoding method including sine and cosine functions to capture the periodic characteristics of time;

[0086] Step 230: Determine whether the dimension of the time period embedding vector matches the preset processing dimension. If so, directly input it into the time period-aware transformer. If not, adjust it to the preset processing dimension through feature mapping before inputting it into the time period-aware transformer.

[0087] In step 240 , the behavior sequence is processed through a time-period-aware Transformer to model cross-time-period behavior dependencies and capture changes in user behavior patterns in different time periods.

[0088] The embedding vector uses a learnable sinusoidal code to generate a time period embedding vector, and the formula is:

[0089] ;

[0090] Where, is the time period embedding vector corresponding to time t (used to encode the periodicity of time), is a time variable (such as a timestamp in minutes or hours), Period parameters related to the time period division granularity , is the dimension of the embedding vector (i.e., contains sine / cosine functions).

[0091] Time t is encoded by sinusoidal functions of different frequencies, capturing periodic patterns of time (such as day and night, weekday / weekend differences).

[0092] Current analysis methods rarely consider cyclical patterns in user behavior, such as behavioral differences between weekdays and weekends, or daytime and nighttime. In fact, users' behavioral motivations and consumption preferences may vary significantly at different times of the day. For example, "the conversion rate of add-to-cart behavior during leisure time is 2.3 times that of work time." Ignoring these cyclical patterns can lead to inappropriate timing and strategy selection for marketing activities, reducing marketing effectiveness.

[0093] The preset processing dimension in step 230 is a fixed dimension pre-set based on the structural design of the time-period-aware Transformer model and business scenario requirements (if encoding a 24-hour time period at hourly granularity, an embedding vector with a dimension of 24 may be generated; if encoding the three macro time periods of "work / leisure / sleep" may generate an embedding vector with a dimension of 3);

[0094] By comparing the values, determine whether the current dimension of the time period embedding vector (such as 24 or 3) is equal to the preset processing dimension (such as 64):

[0095] If the dimensions are equal (matched), the feature structure of the time period embedding vector meets the model input requirements and can be directly input into the time period-aware Transformer to model cross-time period behavior dependencies (for example, analyzing the correlation between users browsing at 8 AM on weekdays and adding items to a cart at 8 PM on weekends).

[0096] If the dimensions are not equal (mismatched): the dimensions need to be adjusted through feature mapping to preserve the temporal periodicity characteristics while adapting to the model input requirements.

[0097] When there is a mismatch, a linear mapping method is used to adjust the time period embedding vector to the preset processing dimension. The specific process is as follows:

[0098] Construct a dimension conversion matrix (shape is "preset processing dimension × original embedding vector dimension") and a bias term (shape is "preset processing dimension × 1");

[0099] The transformation is performed using the formula "adjusted vector = dimension transformation matrix × original time period embedding vector + bias term". The dimension transformation matrix and bias term are learnable parameters that are optimized during model training to ensure that the transformed vector adapts to the input dimension of the Transformer while preserving temporal periodic features (such as the periodic pattern captured by the sine / cosine functions).

[0100] After conversion, the vector dimension is consistent with the preset processing dimension, which can be stably input into the time-perceived Transformer for subsequent processing.

[0101] Step 300, semantic enhancement feature fusion: integrating behavior temporal sequence, entity semantics and time period features to generate a comprehensive feature vector.

[0102] Specifically, step 300 includes:

[0103] Step 310: Based on the triple graph structure and the time period embedding vector, the temporal features of the behaviors are extracted through a bidirectional long short-term memory network. The temporal features reflect the order and correlation of the behaviors.

[0104] Step 320: extracting entity semantic features and time period embedding features output by the graph convolutional neural network. The entity semantic features reflect the hierarchical association features of product categories.

[0105] In step 330 , the temporal features, entity semantic features, and time period embedding features are concatenated and input into a multi-layer perceptron for multimodal fusion to generate a comprehensive feature vector.

[0106] Among them, the multimodal feature fusion input layer splices three types of features:

[0107] Behavior sequence embedding: temporal features extracted by bidirectional long short-term memory network Bi-LSTM;

[0108] Entity semantic embedding: Product category-level features output by the graph convolutional neural network (GCN);

[0109] Period embedding: The periodic period coding vector is fused by the multi-layer perceptron and then input into the self-attention layer:

[0110] ;

[0111] Where, is the comprehensive feature vector after fusion, is the linear rectification activation function , is the weight matrix of the multilayer perceptron, is the bias vector of the multilayer perceptron, is the temporal feature output by the bidirectional long short-term memory network, is the semantic feature output by the graph convolutional neural network, Embed features for time periods.

[0112] This comprehensive feature vector integrates the temporal, semantic, and time period information of the behavior, providing a more comprehensive feature representation for subsequent customer stratification and prediction.

[0113] Traditional customer analysis models typically use offline training and fixed parameters, making it difficult to respond to market changes and the dynamic evolution of user behavior in real time. When the market environment, user needs or business goals change, the model cannot be adjusted in time, resulting in a decrease in the accuracy and effectiveness of the analysis results. In addition, these models have high requirements for the integrity of user information. When user information is incomplete or missing, the model's performance will decline significantly and the generalization ability will be poor.

[0114] After splicing the temporal features, semantic features, and time period features, they are transformed nonlinearly through MLP to generate comprehensive features containing multi-dimensional information.

[0115] Step 400, business goal-sensitive dynamic stratification: Calculate behavioral heat value, combine potential business value and potential business value offset to perform initial two-dimensional clustering to obtain preliminary customer stratification; introduce local multi-source data for cross-validation, optimize preliminary customer stratification through data alignment and feature complementarity, and obtain the final customer stratification result.

[0116] Specifically, step 400 includes:

[0117] Step 410: Calculate the behavior heat value using the comprehensive feature vector. The behavior heat value comprehensively considers the time decay factor, the behavior conversion value weight, and the entity node semantic association score. The time decay factor reflects that recent behaviors are more important than long-term behaviors, the behavior conversion value weight reflects the difference in conversion value between different behaviors, and the entity node semantic association score reflects the correlation between behaviors in the same category.

[0118] Step 420: Calculate the user's potential commercial value and the offset of the potential commercial value. The potential commercial value is predicted based on historical consumption data, and the offset of the potential commercial value is obtained through knowledge graph reasoning, reflecting the deviation between the user's actual value and the predicted value.

[0119] Step 430: Using the behavior heat value as the density index and the potential business value as the value index, two-dimensional clustering is performed to obtain preliminary customer stratification. The preliminary customer stratification includes four levels:

[0120] The behavior heat value and potential commercial value are both higher than the high-activity and high-value level of the corresponding cluster center;

[0121] The high-activity, low-value layer has a higher behavioral heat value than the cluster center and a lower potential commercial value than the cluster center;

[0122] The low-activity high-value layer has a lower behavioral heat value than the cluster center and a higher potential commercial value than the cluster center;

[0123] The behavior heat value and potential commercial value are both lower than the low-activity and low-value level of the corresponding cluster center;

[0124] Step 440: Import local multi-source data, which includes user behavior data and historical interaction data on different local platforms;

[0125] Step 450: Format alignment and time dimension calibration are performed on the local multi-source data and the data based on the initial stratification. Format alignment includes unifying the timestamp format and behavior type definition. Time dimension calibration includes matching behaviors within the same time window.

[0126] Step 460: Extract complementary features from the local multi-source data. The complementary features include cross-platform behavioral preferences and perform cross-validation with the preliminary customer stratification results.

[0127] If the level of a user in the preliminary stratification is consistent with the level reflected by the complementary features, the stratification result is retained;

[0128] If there is any inconsistency, the user's stratification level is adjusted in combination with the complementary features to obtain the final customer stratification result including the above four levels.

[0129] Wherein, the behavioral thermal value calculation formula is:

[0130] ;

[0131] Where, For users Behavioral heat value (measures the activity and value of behavior), is the number of time steps in the time window (such as the number of behavior records in the last 30 days), is the time decay factor (decreasing over time, such as the last 7 days =1, decays by 0.7 every 7 days). The weight of the behavior conversion value (e.g. payment = 10, add to cart = 5, browsing = 1). The semantic association score of the entity node is calculated by the knowledge graph path length, such as "browse → add the same brand product" has a score of 2, and the cross-brand score is 1. The behavior heat value is used as the density indicator, combined with the potential commercial value ( Perform density-value dual-dimensional clustering, For users In time Behavior type (such as browsing, adding to cart), For behavior Related entities (such as products and pages).

[0132] Accumulate the user's behavioral value within the time window, combine time decay and semantic association, and quantify the comprehensive value of user behavior.

[0133] Density-value dual-dimensional clustering improves the DBSCAN algorithm, using HeatScore as the density indicator and introducing "potential commercial value" as the second dimension:

[0134] ;

[0135] A four-quadrant stratification of "high activity and high value / high activity and low value / low activity and high value / low activity and low value" is formed to solve the defect of traditional clustering that ignores commercial value.

[0136] In step 460, the complementary features of local multi-source data refer to features that can make up for the limitations of the single behavioral data on which the initial stratification is based. The core features include cross-platform behavioral preferences and historical interaction depth features. The specific extraction method is as follows:

[0137] Cross-platform behavior preferences: We analyze the behavioral data of the same user account on locally associated platforms (e.g., brand apps, mini-programs, official websites) to extract consistent or differentiated features of cross-platform behavior. For example, if a user frequently browses low-priced items in the app (which might be initially classified as "low value") but has a record of adding high-priced items to their cart in the mini-program, this high-value cross-platform behavior constitutes a complementary feature.

[0138] Historical interaction depth features: Extract long-term value signals from historical user interaction data (such as behavioral records in the past 6 months), including the stability of repurchase cycles (such as fixed purchases every quarter), category expansion trends (such as extending from basic models to high-end models), etc., to compensate for possible accidental deviations in short-term behavioral data.

[0139] During the extraction process, feature standardization (such as normalizing the behavior frequencies of different platforms to the [0,1] interval) is used to eliminate data format differences and ensure feature comparability.

[0140] Cross-validation aims to verify the rationality of the preliminary stratification results through complementary features. The specific process is as follows:

[0141] Feature Mapping: Map the extracted complementary features to the value assessment dimensions consistent with the initial stratification (i.e., the dual dimensions of "behavior heat" and "commercial value"). For example, high-value cross-platform purchases correspond to "potential commercial value improvement," while historical repurchase stability corresponds to "increased behavioral heat weight."

[0142] Consistency assessment: Calculate the degree of match between the preliminary stratification results and the complementary feature mapping results. If the two are completely consistent in their stratification (e.g., both are judged as "high activity, high value"), the preliminary stratification results will be retained directly; if there is a deviation (e.g., the preliminary stratification is "high activity, low value," but the complementary features show that its cross-platform behavior meets the "high value" characteristic), the stratification adjustment mechanism will be triggered;

[0143] Tier adjustment rules: Dynamic adjustments are made based on the weights of complementary features. For example, when the frequency of high-value cross-platform behaviors exceeds 30% (a preset threshold), the "High Activity, Low Value" tier is adjusted to "High Activity, High Value." When historical repurchase cycles are stable and cumulative spending exceeds the average for similar users, the behavior heat weight of the "Low Activity, High Value" tier is increased by 20% to more accurately reflect its long-term value.

[0144] Cross-validation of local multi-source data can effectively avoid the one-sidedness of single behavioral data. For users whose short-term behavior conflicts with long-term value (e.g., frequent browsing but stable long-term purchases), historical interaction characteristics can be used to correct stratification bias. For users whose cross-platform behavior has value differences, cross-platform preference characteristics can be used to supplement the information blind spots of a single platform.

[0145] The final output of four levels (high activity and high value, high activity and low value, low activity and high value, and low activity and low value) will reflect both the intensity of short-term behavior and the potential for long-term value, thereby enhancing the business guidance of customer stratification.

[0146] Step 500, reinforcement learning closed-loop optimization: Based on the final customer stratification results, construct a state space containing the business indicators of each stratification, an action space for parameter adjustment, and a reward function, and optimize the model parameters in real time through reinforcement learning to obtain reinforcement learning parameters.

[0147] Specifically, step 500 includes:

[0148] Step 510: construct a state space, which includes business indicators such as precision, recall, and marketing ROI for each level in the final customer stratification result;

[0149] Step 520: construct an action space, which includes dynamic threshold parameters, behavior weights, and time period division granularity that can be adjusted for different layers;

[0150] Step 530 : Design a business value-oriented reward function. The reward function is calculated based on the comprehensive performance of each layer, including the return on investment of the high-value layer, the difference in F1 scores of different layers, and the cost savings.

[0151] Step 540: Optimize model parameters in real time using a reinforcement learning algorithm to form a self-evolutionary closed loop of "layered result feedback - parameter adjustment - layered accuracy improvement". The reinforcement learning algorithm includes a proximal strategy optimization algorithm.

[0152] Step 550: Determine whether the reward value of the reinforcement learning is lower than the preset performance threshold for a preset number of consecutive times. If so, trigger a parameter reset mechanism, which includes restoring to the historical optimal parameters; if not, continue iterative optimization;

[0153] Step 560: Output optimized reinforcement learning parameters, which include dynamic threshold weights, behavior weights, and time period division granularity parameters.

[0154] Among them, the state space includes six business indicators such as accuracy, recall rate, and marketing ROI for each customer group;

[0155] Action space: Adjust five configurable parameters, including dynamic threshold parameters, behavior weights, and time segmentation granularity;

[0156] The action space can adjust parameters such as dynamic threshold parameters, behavior weights, and time period division granularity. The reward function is:

[0157] ;

[0158] Where, is the reward value of reinforcement learning (measures the commercial value of model decisions), 、 and is the weight coefficient, is the return on investment (ratio of marketing benefits to costs), The F1 score (a measure of precision and recall) for high-value customer groups, is the F1 score of the low-value customer group, Cost savings (e.g., cost reduction from reducing invalid reach);

[0159] in, Focus on the two high-value tiers of "high activity and high value" and "low activity and high value" in the final customer stratification, and calculate the ratio of their marketing input to output. For example, the input includes marketing resources for the high-value tier (such as coupon costs and reach channel costs); the output includes the actual consumption amount of customers in this tier and the incremental revenue brought by repeat purchases; this indicator has the highest weight in the reward function ( Usually set to 0.5-0.6), prioritizing marketing efficiency for high-value customers.

[0160] To quantify the difference in recognition accuracy between "high-value customer groups" (high activity, high value, low activity, high value) and "low-value customer groups" (high activity, low value, low activity, low value), the F1 score is the harmonic average of precision and recall, reflecting the degree of match between the stratification results and the actual value;

[0161] When the F1 score of the high-value customer group is significantly higher than that of the low-value customer group (for example, the difference is ≥ 0.2), it means that the model can effectively distinguish high-value and low-value customers, and this part contributes positive rewards;

[0162] Weight Set to 0.3-0.4 to encourage the model to strengthen its ability to distinguish different value layers.

[0163] This is mainly due to the reduction of ineffective reach to the “low-activity, low-value” level. Quantification methods include:

[0164] Compare the marketing costs for low-value segments before and after optimization (e.g., reduced SMS push fees and ineffective advertising costs);

[0165] Combined with the increase in conversion rate for low-value segments (e.g., from 1% to 3%), calculate the reduction in resource waste due to precise screening;

[0166] Weight Set it to 0.1-0.2 to balance the goals of cost control and high-value customer mining.

[0167] 、 and The value is adjusted dynamically according to business objectives (such as increasing Prioritize ROI and increase new customer acquisition To strengthen customer segmentation), and adjust the logic and link it with the business indicators of the final customer segmentation (for example, when the ROI of the high-value segmentation continues to decline, it will automatically increase weight).

[0168] Furthermore, the self-evolutionary closed loop of "stratification result feedback - parameter adjustment - stratification accuracy improvement" is the core mechanism that dynamically associates customer stratification results with model parameters through reinforcement learning. The specific process is as follows:

[0169] Feedback on stratification results: The final customer stratification results (four tiers, such as high activity and high value, high activity and low value) and their corresponding business indicators (such as accuracy and marketing ROI for each tier) are input into the reinforcement learning model as environmental states, providing real-time feedback on the current stratification effect.

[0170] Parameter adjustment: Based on state feedback, the reinforcement learning algorithm is used to adjust key model parameters (dynamic threshold weight, behavior weight, and time period granularity). For example, if the recall rate of a high-activity, high-value layer is lower than the preset threshold, the behavior weight parameter corresponding to that layer is increased to enhance sensitivity to high-value behaviors.

[0171] Improved stratification accuracy: The adjusted parameters are applied to the next round of customer stratification calculations. Through iterative optimization, the match between the stratification results and business objectives (such as maximizing the recall rate of high-value customers) is gradually improved, forming a closed-loop iteration.

[0172] The reinforcement learning algorithm uses a proximal policy optimization algorithm. Its core advantage is to achieve stable training by limiting the amplitude of parameter updates. Specific applications include:

[0173] Policy network design: Build a neural network as a policy function, with the input being the business indicators of the current customer tier (such as accuracy and cost savings at each tier), and the output being the probability distribution of parameter adjustments (i.e., the recommended probabilities of different parameter adjustment plans). For example, for the high-activity, low-value tier, the output is the probability of adjustment plans such as "increasing the granularity of time segmentation" or "reducing the weight of dynamic thresholds."

[0174] Stable training mechanism: By constraining the difference between the new and old strategies, we can avoid fluctuations in stratification results caused by excessive parameter updates. For example, we can control the probability ratio of the new strategy to the old strategy within a reasonable range to ensure that parameter adjustments are effective without deviating too much from the historical optimal state.

[0175] Multi-round iterative training: Each round of iteration uses the latest customer stratification data to update the policy network. Through multiple rounds of iteration, the parameters gradually converge to the optimal solution, balancing the exploration of new parameter combinations and the use of historical optimal experience.

[0176] The reward value judgment and parameter reset mechanism include:

[0177] Reward Value Calculation: Reward values ​​are dynamically generated based on the business performance of each tier, comprehensively reflecting the business value of the tiered results. For example, an increase in the ROI of highly active and high-value tiers, a widening of the F1 score differences between different tiers, and a reduction in the cost of ineffective engagement for low-value tiers will all positively impact reward values.

[0178] Preset performance threshold setting: The preset performance threshold is a certain proportion of the historical optimal reward value (which can be adjusted according to business goals). If the reward value is lower than this threshold for multiple consecutive times (for example, three times), it is determined that the model parameters may have fallen into a local optimum and cannot be improved through regular iterations;

[0179] Parameter reset operation: When the reset mechanism is triggered, the current parameters are restored to the historical optimal parameters (that is, the parameter combination with the highest reward value in the past period of time). After the restoration, the parameter exploration range is appropriately increased to avoid falling into the same local optimal state repeatedly, ensuring the model's ability to continue to evolve.

[0180] The output reinforcement learning parameters include three key parameters, and their application scenarios are as follows:

[0181] Dynamic threshold weights: used to adjust the dynamic threshold calculation logic for different tiers. For example, assigning a higher potential business value offset weight to the high-activity, high-value tier, and a higher behavioral heat value weight to the low-activity, low-value tier.

[0182] Behavior Weight: Adjust the importance of different behaviors in the behavior heat value calculation. For example, increase the weight of high-conversion behaviors such as "repurchase" and "add to cart" and reduce the weight of shallow behaviors such as "browsing";

[0183] Time segmentation granularity: Optimize the time segmentation strategy. For example, use finer granularity for peak user activity periods (such as evening leisure time) and coarser granularity for low-activity periods, balancing computational efficiency and feature capture capabilities.

[0184] Step 600, dynamic threshold decision: Based on the final customer stratification results, behavioral heat value, potential business value offset and reinforcement learning parameters, differentiated dynamic thresholds are used for different stratifications to output the potential customer intention probability.

[0185] Specifically, step 600 further includes:

[0186] Step 610: Setting differentiated dynamic threshold calculation logic for each of the four levels in the final customer stratification result;

[0187] Step 620: For the high-activity, high-value layer, the dynamic threshold calculation focuses on the offset of potential commercial value; for the high-activity, low-value layer, the dynamic threshold calculation balances the behavior heat value and the offset of potential commercial value; for the low-activity, high-value layer, the dynamic threshold calculation reduces the weight of the behavior heat value; for the low-activity, low-value layer, the dynamic threshold calculation increases the weight of the behavior heat value;

[0188] Step 630: Based on the differentiated calculation logic in step 610, the dynamic threshold is calculated by combining the behavioral heat value, the offset of the potential commercial value, and the reinforcement learning parameters. The dynamic threshold is then matched with the comprehensive feature vector to output the potential customer's intention probability. The higher the match, the higher the intention probability. The intention probability of customers at the same level is determined uniformly.

[0189] Among them, the calculation factors of dynamic thresholds include behavioral heat value, offset of potential commercial value and reinforcement learning parameters; the differentiation of dynamic threshold calculation logic is reflected in the weight distribution of each factor, and the weight distribution is determined based on the business objectives of each level.

[0190] In step 610, differentiated dynamic thresholds are set for the four levels of final customer stratification. The essence of this is to adjust the judgment criteria according to the business objectives of each level:

[0191] For the high-activity, high-value tier, priority must be given to ensuring that high-potential customers are not missed, so the focus is on potential business value deviations that can reflect long-term value fluctuations.

[0192] The high-activity, low-value tier requires distinguishing between "short-term active but valueless" and "active with potential value to be released" customers. Therefore, it is necessary to balance the behavioral heat value (short-term activity) with the potential business value offset (long-term value signal).

[0193] The low-activity, high-value tier needs to tap into customers with low behavioral frequency but high value, so the weight of the behavioral heat value (activity) is reduced to avoid misjudgment due to short-term silence;

[0194] The low-activity and low-value tiers need to control marketing costs and reduce ineffective reach, so the weight of behavioral heat value should be increased, and only customers with significantly improved activity should be retained.

[0195] In step 620, the calculation logic of the dynamic thresholds at each level is as follows:

[0196] High-activity, high-value tier: When calculating dynamic thresholds, the potential commercial value offset accounts for 60%-70%. For example, if a customer's historical consumption records indicate that their potential commercial value is "high," but their behavioral heat value has recently decreased slightly due to a decrease in promotional activities, the potential commercial value offset (such as knowledge graph reasoning indicating potential demand for new product categories) will dominate the threshold judgment. Even if the behavioral heat value is slightly lower, they may still be judged as a high-intent customer, avoiding missed judgments.

[0197] High Activity, Low Value Tier: The dynamic threshold uses a balanced weighting of "50% each for Behavioral Heat Value and Potential Commercial Value Offset." For example, if a customer browses frequently but never purchases (high Behavioral Heat Value, low Current Value), and the potential commercial value offset indicates that their recently browsed products are highly similar to those of historical high-value customers, they will be included in the "potential customer to be converted" through a balanced calculation, rather than being directly classified as having low intent.

[0198] Low-activity, high-value tier: The weight of behavioral heat value is reduced to 30%-40%, while the weight of potential commercial value offset is increased to 60%-70%. For example, a customer who purchases once a quarter but has a very high average order value (low behavioral heat value, high value) may still be considered a high-intent customer even if they haven't purchased for three consecutive months. If the potential commercial value offset indicates that their customer base typically repurchases every four months, they will still be considered a high-intent customer, thus avoiding misscreening due to short-term low activity.

[0199] Low-activity, low-value tier: The weight of the behavioral heat value is increased to 70%-80%. Only when activity increases significantly (e.g., behavior frequency increases by more than 50% compared to the historical average) will the customer pass the threshold. For example, a customer who has been browsing low-priced items infrequently for a long time will be considered a potential customer only if they recently purchase a higher-priced item (a sudden increase in the behavioral heat value), thus reducing ineffective outreach to stable, low-value customers.

[0200] In step 630, the dynamic threshold is calculated based on the behavior heat value and the offset of the potential commercial value, and is dynamically adjusted in combination with reinforcement learning parameters (dynamic threshold weight, behavior weight, etc.):

[0201] Reinforcement learning parameters are optimized in real time based on the historical judgment results of each layer. For example, if the missed judgment rate of the high-activity and high-value layer increases, the weight coefficient of the potential commercial value offset will be automatically increased;

[0202] The matching degree between the threshold and the comprehensive feature vector is calculated through feature similarity. The higher the matching degree, the more consistent the customer characteristics are with the high-intention characteristics of this level, and the higher the output intention probability is (for example, a matching degree of 90% corresponds to an intention probability of 0.9).

[0203] The probability of customer intention within the same level adopts a unified judgment standard (just like the mapping relationship between matching degree and probability within the level is consistent), ensuring horizontal comparability within the level and providing a clear basis for subsequent marketing resource allocation (just like the top 30% of customers with intention probability within the same level will receive priority for accurate push notifications).

[0204] In this embodiment, the present invention constructs a triplet graph structure and injects domain knowledge, combines periodic time period modeling with multimodal feature fusion, and deeply explores behavioral semantic associations and temporal dependencies to capture users' implicit needs; optimizes customer stratification through local multi-source data cross-validation, clarifies the four-level division and associates reinforcement learning with dynamic threshold decision-making to improve stratification accuracy; forms a self-evolution closed loop and differentiated dynamic threshold design through reinforcement learning to enhance model adaptability, reduce missed judgments of high-value customers and misjudgments of low-value customers, and ultimately improve target customer screening efficiency and marketing ROI.

[0205] Based on the above embodiments, Figure 2 As shown, the present invention further provides an analysis device for screening a target customer group, which is used to support the analysis method for screening a target customer group in the above embodiment. The analysis device for screening a target customer group includes:

[0206] A graph modeling module 11 is used to construct registered user behavior data into a triple graph structure containing user nodes, behavior nodes, and entity nodes, and inject domain knowledge for semantic enhancement;

[0207] The time series analysis module 12 is used to divide the 24-hour period into periodic periods, generate period embedding vectors, and model cross-period behavior dependencies through period-aware transformers;

[0208] A feature fusion module 13 is used to fuse behavior temporal features, entity semantic features, and time segment features to generate a comprehensive feature vector;

[0209] The stratification optimization module 14 is used to calculate the behavioral heat value and perform initial two-dimensional clustering based on the potential business value and the offset of the potential business value to obtain preliminary customer stratification. Local multi-source data is introduced for cross-validation, and the preliminary customer stratification is optimized through data alignment and feature complementarity to obtain the final customer stratification results.

[0210] The reinforcement learning module 15 is used to construct a state space containing business indicators of each stratification, an action space for parameter adjustment, and a reward function based on the final customer stratification results, and to obtain reinforcement learning parameters by optimizing model parameters in real time through reinforcement learning;

[0211] The prediction decision module 16 is used to output the probability of potential customer intention based on the final customer stratification results, behavioral heat value, offset of potential commercial value and reinforcement learning parameters, using differentiated dynamic thresholds for different stratifications.

[0212] In this embodiment, the present invention constructs a triplet graph structure and injects domain knowledge, combines periodic time period modeling with multimodal feature fusion, and deeply explores behavioral semantic associations and temporal dependencies to capture users' implicit needs; optimizes customer stratification through local multi-source data cross-validation, clarifies the four-level division and associates reinforcement learning with dynamic threshold decision-making to improve stratification accuracy; forms a self-evolution closed loop and differentiated dynamic threshold design through reinforcement learning to enhance model adaptability, reduce missed judgments of high-value customers and misjudgments of low-value customers, and ultimately improve target customer screening efficiency and marketing ROI.

[0213] Furthermore, the analysis device for screening the target customer group may run the above-mentioned analysis method for screening the target customer group. For specific implementation, please refer to the method embodiment, which will not be described in detail here.

[0214] Based on the above embodiments, Figure 3 As shown, the present invention also provides an electronic device, the electronic device comprising:

[0215] At least one processor 22, at least one memory 21, a communication interface 23 and a communication bus 24, wherein the processor 22 is communicatively connected to the memory 21;

[0216] In this embodiment, the memory 21 can be implemented in any appropriate manner, for example, the memory 21 can be a read-only memory, a mechanical hard disk, a solid-state drive, or a USB flash drive, etc. The memory 21 is used to store at least one executable instruction executed by the processor;

[0217] In this embodiment, the processor 22 can be implemented in any appropriate manner. For example, the processor 22 can take the form of a microprocessor or a processor and a computer-readable medium storing computer-readable program code (such as software or firmware) that can be executed by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, etc.; the processor is used to execute the executable instructions to implement the above-mentioned analysis method for screening the target customer group.

[0218] Based on the above embodiments, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned analysis method for screening target customer groups is implemented.

[0219] Those skilled in the art will appreciate that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0220] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices, equipment and modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0221] In the several embodiments provided in this application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple modules or units can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or equipment, which can be electrical, mechanical or other forms.

[0222] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network elements. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0223] In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0224] If the functions are implemented as software modules and sold or used as standalone products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for causing a computer device (such as a personal computer, server, or network device) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program instructions, such as USB flash drives, mobile hard drives, read-only storage servers, random access storage servers, magnetic disks, or optical disks.

[0225] In addition, it should be noted that the combination of the various technical features in this case is not limited to the combination described in the claims of this case or the combination described in the specific embodiments. All technical features recorded in this case can be freely combined or combined in any way unless there is a contradiction between them.

[0226] It should be noted that the above examples are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples, and many similar variations are possible. All variations directly derived from or associating with the present invention by those skilled in the art are intended to fall within the scope of protection of the present invention.

[0227] The above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An analysis method for screening target customer groups, characterized in that: The following steps are involved: The registered user behavior data is constructed into a triple graph structure containing user nodes, behavior nodes, and entity nodes, and domain knowledge is injected for semantic enhancement; Divide 24 hours into periodic periods, generate period embedding vectors, and use period-aware Transformers to model cross-period behavioral dependencies. Fuse behavior timing, entity semantics, and time period features to generate a comprehensive feature vector; Calculate behavioral heat value, combine potential business value and potential business value offset to perform initial two-dimensional clustering to obtain preliminary customer stratification; introduce local multi-source data for cross-validation. Local multi-source data includes local cross-platform user behavior data and historical interaction data. Through data alignment and feature complementarity, optimize the preliminary customer stratification to obtain the final customer stratification results; Based on the final customer stratification results, a state space containing business indicators for each stratification, an action space for parameter adjustment, and a reward function are constructed. Reinforcement learning parameters are then obtained by optimizing model parameters in real time through reinforcement learning. Based on the final customer stratification results, behavioral heat value, offset of potential commercial value and reinforcement learning parameters, differentiated dynamic thresholds are used for different stratifications to output the probability of potential customer intention.

2. The method according to claim 1, characterized in that The registered user behavior data is constructed into a triple graph structure containing user nodes, behavior nodes, and entity nodes, and domain knowledge is injected for semantic enhancement, including: Behavior nodes represent behavior types, and entity nodes are associated with category-level knowledge; Semantic enhancement is achieved through graph convolutional neural networks, which inject domain knowledge about product category hierarchical relationships and interactive funnel conversion paths, aggregate the features of nodes and their adjacent nodes, and update node features based on the node's own features. During the aggregation process, the node degree is regularized, including adjusting the feature weight according to the degree of the node and its adjacent nodes to balance the feature influence of nodes with different degrees and achieve semantic enhancement of node features.

3. The method according to claim 1, characterized in that The 24-hour period is divided into periodic periods, period embedding vectors are generated, and cross-period behavior dependencies are modeled using a period-aware Transformer, including: Process the time information in the behavioral data and divide 24 hours into several periodic periods; For each behavioral event, an encoding method including sine and cosine functions is used to generate the corresponding time period embedding vector to capture the periodic characteristics of time; Determine whether the dimension of the time period embedding vector matches the preset processing dimension. If so, it is directly input into the time period aware Transformer. If not, it is adjusted to the preset processing dimension through feature mapping and then input into the time period aware Transformer. The behavior sequence is processed through the time-period-aware Transformer, cross-time behavior dependencies are modeled, and changes in users' behavior patterns in different time periods are captured.

4. The method according to claim 1, wherein The fusion of behavior timing, entity semantics, and time period features to generate a comprehensive feature vector includes: Based on the triple graph structure and time period embedding vector, the temporal features of behaviors are extracted through a bidirectional long short-term memory network. The temporal features reflect the order and correlation of behaviors. Extract entity semantic features and time period embedding features output by the graph convolutional neural network. Entity semantic features reflect the hierarchical association features of product categories. After the temporal features, entity semantic features, and time period embedding features are spliced ​​together, they are input into a multi-layer perceptron for multimodal fusion to generate a comprehensive feature vector.

5. The method according to claim 1, wherein The calculated behavior heat value is combined with the potential business value and the offset of the potential business value to perform initial two-dimensional clustering to obtain preliminary customer stratification; Introducing local multi-source data for cross-validation, optimizing preliminary customer stratification through data alignment and feature complementarity, and obtaining the final customer stratification results, including: The behavior heat value is calculated using the comprehensive feature vector. The behavior heat value comprehensively considers the time decay factor, the behavior conversion value weight, and the entity node semantic association score. The time decay factor reflects that recent behaviors are more important than long-term behaviors. The behavior conversion value weight reflects the difference in conversion value between different behaviors. The entity node semantic association score reflects the correlation between behaviors in the same category. Calculate the user's potential commercial value and the offset of the potential commercial value. The potential commercial value is predicted based on historical consumption data, and the offset of the potential commercial value is obtained through knowledge graph reasoning, reflecting the deviation between the user's actual value and the predicted value. Using behavioral heat value as the density indicator and potential commercial value as the value indicator, two-dimensional clustering is performed to obtain preliminary customer stratification. The preliminary customer stratification includes four levels: the high-activity and high-value level, where both the behavioral heat value and the potential commercial value are higher than the corresponding cluster center; the high-activity and low-value level, where the behavioral heat value is higher than the cluster center and the potential commercial value is lower than the cluster center; the low-activity and high-value level, where the behavioral heat value is lower than the cluster center and the potential commercial value is higher than the cluster center; and the low-activity and low-value level, where both the behavioral heat value and the potential commercial value are lower than the corresponding cluster center. Introducing local multi-source data, which includes user behavior data and historical interaction data on different local platforms; Perform format alignment and time dimension calibration on local multi-source data and the data on which the initial stratification is based. Format alignment includes unifying the timestamp format and behavior type definition, and time dimension calibration includes matching behaviors within the same time window. Complementary features are extracted from local multi-source data. The complementary features include cross-platform behavioral preferences and are cross-validated with the preliminary customer stratification results: if a user's level in the preliminary stratification is consistent with the level reflected by the complementary features, the stratification result is retained; if not, the user's stratification level is adjusted in combination with the complementary features to obtain the final customer stratification result including the above four levels.

6. The method according to claim 1, characterized in that Based on the final customer stratification results, the state space including the business indicators of each stratification, the action space for parameter adjustment, and the reward function are constructed, and the reinforcement learning parameters are obtained by real-time optimization of the model parameters through reinforcement learning, including: Construct a state space that includes the precision, recall, and marketing ROI business indicators for each level in the final customer stratification results; Construct an action space, which includes dynamic threshold parameters, behavior weights, and time segmentation granularity that can be adjusted for different layers; Design a business value-oriented reward function based on the comprehensive calculation of the performance of each tier, including the return on investment of the high-value tier, the difference in F1 scores between different tiers, and the cost savings; Real-time optimization of model parameters is achieved through reinforcement learning algorithms, forming a self-evolutionary closed loop of "layered result feedback - parameter adjustment - layered accuracy improvement." Reinforcement learning algorithms include proximal strategy optimization algorithms. Determine whether the reward value of reinforcement learning is lower than the preset performance threshold for a preset number of consecutive times. If so, trigger the parameter reset mechanism, which includes restoring to the historical optimal parameters; otherwise, continue iterative optimization; Output optimized reinforcement learning parameters, which include dynamic threshold weight, behavior weight and time period division granularity parameters.

7. The method according to claim 1, characterized in that According to the final customer stratification results, behavioral heat value, potential business value offset and reinforcement learning parameters, differentiated dynamic thresholds are used for different stratifications to output the potential customer intention probability, including: Differentiated dynamic threshold calculation logic is set for each of the four levels in the final customer stratification results. For high-activity, high-value tiers, dynamic threshold calculations focus on the offset of potential commercial value; for high-activity, low-value tiers, dynamic threshold calculations balance the behavioral heat value and the offset of potential commercial value; for low-activity, high-value tiers, dynamic threshold calculations reduce the weight of behavioral heat value; for low-activity, low-value tiers, dynamic threshold calculations increase the weight of behavioral heat value; Based on differentiated calculation logic, the dynamic threshold is calculated by combining behavioral heat value, potential business value offset, and reinforcement learning parameters. The matching degree between the dynamic threshold and the comprehensive feature vector is then used to output the potential customer's intention probability. The higher the matching degree, the higher the intention probability. The intention probability of customers within the same tier is judged using a unified standard. Among them, the calculation factors of dynamic thresholds include behavioral heat value, offset of potential commercial value and reinforcement learning parameters; the differentiation of dynamic threshold calculation logic is reflected in the weight distribution of each factor, and the weight distribution is determined based on the business objectives of each level.

8. An analysis device for screening target customer groups, characterized in that: include: The graph modeling module is used to construct the registered user behavior data into a triple graph structure containing user nodes, behavior nodes, and entity nodes, and inject domain knowledge for semantic enhancement; The time series analysis module is used to divide the 24-hour period into periodic periods, generate period embedding vectors, and use the period-aware Transformer to model cross-period behavioral dependencies; Feature fusion module, used to fuse behavior temporal features, entity semantic features, and time segment features to generate a comprehensive feature vector; The stratification optimization module is used to calculate behavioral heat value and perform initial two-dimensional clustering based on potential business value and potential business value offset to obtain preliminary customer stratification. Local multi-source data is introduced for cross-validation, and the preliminary customer stratification is optimized through data alignment and feature complementarity to obtain the final customer stratification results. The reinforcement learning module is used to construct a state space containing business indicators for each stratification, an action space for parameter adjustment, and a reward function based on the final customer stratification results. The reinforcement learning parameters are obtained by optimizing the model parameters in real time through reinforcement learning. The prediction decision module is used to output the probability of potential customer intention based on the final customer stratification results, behavioral heat value, offset of potential commercial value and reinforcement learning parameters, using differentiated dynamic thresholds for different stratifications.

9. An electronic device, characterized in that: The electronic device comprises: a processor and a memory, wherein the memory is communicatively connected to the processor; The memory is used to store at least one executable instruction executed by the processor, and the processor is used to execute the executable instruction to implement the analysis method for screening target customer groups according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the analysis method for screening a target customer group according to any one of claims 1 to 7 is implemented.

Images