A machine learning-based end-to-end marketing method and system
By collecting and modeling multi-touchpoint behavior logs, generating cross-touchpoint behavior event sequences, and using a multi-task neural network to output a comprehensive marketing state vector, the problem of missing user state feature extraction information and the inability to adjust marketing intervention actions in real time in existing technologies is solved, realizing closed-loop iteration of full-link marketing and precise marketing effects.
Patent Information
- Application Number
- CN202611295578.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-25
- Publication Date
- 2026-11-10
AI Technical Summary
Existing marketing systems cannot achieve unified merging of user behavior logs across windows and channels, resulting in missing information in the extraction of user status features, inability to adjust marketing interventions in real time, failure to achieve closed-loop feedback of marketing data across the entire chain, poor coordination of multi-touchpoint marketing, and a decline in long-term operational accuracy.
By collecting multi-touchpoint behavior logs from multiple independent marketing touchpoints, a cross-touchpoint behavior event sequence is generated. A unified model is then built using a multi-task neural network to output a comprehensive marketing state vector. Based on the user's marketing state classification results, candidate marketing intervention actions are generated and executed throughout the entire process.
It achieves unified merging of multi-touchpoint behaviors and full-link closed-loop iteration, improves the accuracy of user segmentation and classification, enhances the precision of marketing outreach, forms a complete marketing closed loop of perception-classification-execution-feedback, and continuously optimizes the intelligent marketing effect across the entire domain and link.
Smart Images

Figure CN123070992A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of next-generation information technology, and in particular to a machine learning-based end-to-end marketing method and system. Background Technology
[0002] With the diversification of online marketing channels, marketing scenarios have gradually differentiated into multiple independent marketing touchpoints such as apps, mini-programs, SMS, and H5 pages, leading to a continuous increase in fragmented user interactions across touchpoints. In the early stages, the industry mostly adopted a single-touchpoint independent behavior analysis model for user marketing operations, relying on single-channel behavior data to determine user marketing needs. Subsequently, machine learning technology was gradually integrated into the field of intelligent marketing, and neural network models were widely used in predicting and analyzing user conversion intentions and churn risks. By modeling behavioral sequences, potential marketing characteristics of users were mined to achieve user segmentation and targeted marketing pushes. At present, the industry is gradually exploring multi-touchpoint behavior fusion analysis, attempting to rely on multi-task learning networks to simultaneously process multi-dimensional marketing prediction tasks, connect user behavior data links across different channels, and adapt to the needs of refined marketing management across multiple scenarios and all user domains.
[0003] Currently, existing marketing systems suffer from fragmented data from multiple independent marketing touchpoints. Conventional technologies cannot achieve unified merging and modeling of user behavior logs across windows and channels, making it difficult to fully reconstruct the time-series behavioral patterns of users throughout their entire lifecycle. Furthermore, the extraction of user status features suffers from significant information gaps. Moreover, existing marketing interventions heavily rely on pre-set, fixed strategy libraries for matching and generation, failing to dynamically adapt and adjust to real-time user behavior. This results in defects such as generalized action matching, inconsistent touchpoint pushes, and a mismatch between intervention goals and user needs. Consequently, end-to-end marketing data cannot achieve closed-loop feedback and iteration, new user behavior after touchpoint execution cannot be updated in reverse, multi-touchpoint marketing collaboration is poor, and long-term operational accuracy continues to decline. Therefore, this paper proposes a machine learning-based end-to-end marketing method and system to address these issues. Summary of the Invention
[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution:
[0005] A machine learning-based end-to-end marketing approach includes:
[0006] S1: Collect multi-touchpoint behavior logs from multiple independent marketing touchpoints, and generate cross-touchpoint behavior event sequences based on customer identifier matching and time window session merging in the multi-touchpoint behavior logs;
[0007] S2: Adapt the cross-touchpoint behavioral event sequence to obtain the input tensor, and then pre-train a multi-task neural network containing a shared bottom-level encoding layer and two parallel task heads. The input tensor is received through the shared bottom-level encoding layer and the marketing feature encoding vector is output. The prediction tasks of two marketing objectives are executed simultaneously through the two parallel task heads, and the comprehensive marketing state vector is output.
[0008] S3: Perform full user classification calculation on the comprehensive marketing status vector to obtain user marketing status classification results, and generate candidate marketing intervention actions based on user marketing status classification results;
[0009] S4: Assign candidate marketing intervention actions to the corresponding marketing touchpoint execution engine for execution, achieving end-to-end marketing.
[0010] The process of collecting multi-touchpoint behavior logs includes:
[0011] Deploy user behavior tracking points at multiple independent marketing touchpoints and capture user interaction logs at each touchpoint using the tracking SDK;
[0012] The behavior log data from different touchpoints is aggregated in real time through the message queue Kafka to form a multi-touchpoint behavior log set.
[0013] The process of generating a sequence of cross-touchpoint behavioral events includes:
[0014] Based on historical data, a cross-touchpoint user identifier mapping table is constructed to associate the user identifiers of the same user in the behavior logs of different touchpoints, and the session window is divided by time as the dimension. The duration threshold T of the session window is set according to the business scenario.
[0015] For each user's multi-touchpoint behavior log set, sort it in ascending order according to the timestamp t of the behavior occurrence to obtain a sorted behavior log sequence. Traverse the sorted sequence and group consecutive behavior log sequences with a time difference less than T into the same session window to generate a cross-touchpoint behavior event sequence.
[0016] The process of adapting cross-touchpoint behavioral event sequences includes:
[0017] The cross-touchpoint behavior event sequence within each session window is sorted by timestamp, and the behavior type in the cross-touchpoint behavior event sequence is encoded into a vector representation, and the business object identifier is mapped into a vector representation;
[0018] The behavior events and business object identifiers within the same session window are concatenated or stacked in chronological order to form a sequence of input tensors.
[0019] The implementation process of the shared underlying encoding layer in a multi-task neural network includes:
[0020] The shared underlying encoding layer incorporates a Transformer encoder, a multi-head self-attention mechanism, and a feedforward neural network.
[0021] The Transformer encoder adds temporal position encoding to the input tensor, outputting a feature tensor carrying temporal position information. Based on the feature tensor, a multi-head self-attention mechanism is used to calculate the association weights of each behavioral event in the sequence in parallel, outputting a global behavioral association feature matrix. Based on the global behavioral association feature matrix, a feedforward neural network is input to perform a multi-layer nonlinear mapping transformation, outputting a marketing feature encoding vector.
[0022] The process of simultaneously performing forecasting tasks for two marketing objectives includes:
[0023] The multi-task neural network contains two parallel task heads, which process the marketing feature encoding vectors of the shared encoding layer respectively, and perform two marketing objective prediction tasks: user purchase conversion rate prediction and user churn risk prediction.
[0024] User purchase conversion rate prediction obtains a fixed-dimensional feature vector by performing built-in global pooling and sequence-to-single-value transformation on the encoded feature sequence. Then, through mapping via a fully connected layer, the predicted purchase conversion rate is output.
[0025] User churn risk prediction involves processing the marketing feature encoding vector, performing average pooling, calculating the average value of features across all time steps, and selecting the maximum value among all time steps to obtain a fixed-dimensional feature vector. Then, the feature vector is activated by the Sigmoid activation function. Converted into a predicted probability value of churn risk;
[0026] The purchase conversion rate prediction, churn risk probability prediction, and marketing feature encoding vectors from the shared encoding layer are concatenated to obtain a comprehensive marketing state vector.
[0027] The process of obtaining user marketing status rating results includes:
[0028] The output comprehensive marketing state vector is subjected to multi-dimensional feature analysis to extract core marketing state features, including activity features. Preference characteristics With value characteristics ;
[0029] A hierarchical approach driven by both dynamic clustering and business rules is adopted to divide users into K levels. Based on the core marketing status, a full-scale user hierarchical calculation is performed to obtain the user marketing status hierarchical results. ,in, For high-value active users, This corresponds to growing active users.
[0030] The process of tiered calculation for all users includes:
[0031] If satisfied , For empty set, as well as The user marketing status classification result is: This corresponds to high-value active users;
[0032] in, The first-level threshold for activity level, For the core preference tag set, As the first-level threshold of value, For active user lifecycle stages, This refers to the current lifecycle stage of user u.
[0033] If satisfied , , as well as The user marketing status classification result is: This corresponds to a growing number of active users;
[0034] in, This is the secondary threshold for activity level. For a set of growth preference tags, This is the second-order threshold for value.
[0035] The process of generating candidate marketing interventions includes:
[0036] Obtain user marketing status classification results The corresponding purchase conversion rate prediction, churn risk probability prediction, and marketing feature encoding vector;
[0037] When the predicted purchase conversion rate is higher than the preset conversion threshold and the predicted churn risk probability is lower than the preset churn threshold, a conversion promotion action type is generated.
[0038] When the predicted purchase conversion rate is higher than the preset conversion threshold and the predicted churn risk probability is higher than the preset churn threshold, a churn recovery action type is generated.
[0039] When the predicted purchase conversion rate is lower than or equal to the preset conversion threshold, an activity boosting action type is generated directly.
[0040] Based on the marketing feature encoding vector corresponding to the user marketing status classification results, the action execution touchpoint and the business object to which the action is directed are determined. Combined with the user marketing status classification results, the action type, execution touchpoint, business object, and user marketing status classification results are structurally combined to form a complete candidate marketing intervention action.
[0041] A machine learning-based end-to-end marketing system includes:
[0042] Marketing data collection module: Collects multi-touchpoint behavior logs from multiple independent marketing touchpoints, and generates cross-touchpoint behavior event sequences based on customer identifier matching and time window session merging in the multi-touchpoint behavior logs;
[0043] State decision module: Adapts the cross-touchpoint behavioral event sequence to obtain the input tensor, and then pre-trains a multi-task neural network containing a shared bottom layer encoding layer and two parallel task heads. The shared bottom layer encoding layer receives the input tensor and outputs the marketing feature encoding vector. The two parallel task heads simultaneously execute the prediction tasks of two marketing objectives and output a comprehensive marketing state vector.
[0044] Marketing intervention module: Performs full-user classification calculation on the comprehensive marketing status vector to obtain user marketing status classification results, and generates candidate marketing intervention actions based on the user marketing status classification results;
[0045] Intervention Execution Module: Assigns candidate marketing intervention actions to the corresponding marketing touchpoint execution engine for execution, realizing end-to-end marketing.
[0046] The present invention has the following beneficial effects:
[0047] 1. Through the technical logic of unified merging of multi-touchpoint behaviors, multi-task shared neural network modeling, state-level direct-drive action generation, and full-link closed-loop iteration, it can fully integrate fragmented user behaviors from multiple independent touchpoints, accurately extract full-domain temporal behavior features, fully restore the user's real cross-touchpoint marketing status, and significantly improve the accuracy of user segmentation and classification.
[0048] 2. Secondly, by sharing the underlying encoding layer and using a dual-parallel task prediction architecture, the system synchronously outputs user metrics in both conversion and churn dimensions, resulting in higher feature reuse efficiency, lower system computational overhead, and a more comprehensive representation of the user's overall state vector.
[0049] 3. This solution does not rely on a pre-set marketing strategy library. It directly generates appropriate intervention actions based on the user's real-time marketing status classification results, effectively avoiding the generalization and adaptation problems of traditional strategies. The accuracy of marketing reach is significantly improved. Furthermore, it can realize the full feedback and iteration of touchpoint execution data, forming a complete marketing closed loop of perception-classification-execution-feedback, continuously optimizing the intelligent marketing effect across the entire domain and link, and improving user conversion, retention and full-cycle marketing value. Attached Figure Description
[0050] Figure 1 This is a flowchart illustrating the steps of a machine learning-based end-to-end marketing method proposed in this invention.
[0051] Figure 2 This is a system block diagram of a machine learning-based end-to-end marketing system proposed in this invention. Detailed Implementation
[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0053] Example 1:
[0054] like Figure 1 As shown, the present invention proposes a machine learning-based end-to-end marketing method, comprising:
[0055] S1: Collect multi-touchpoint behavior logs from multiple independent marketing touchpoints, and generate cross-touchpoint behavior event sequences based on customer identifier matching and time window session merging in the multi-touchpoint behavior logs;
[0056] Deploy user behavior tracking points at multiple independent marketing touchpoints and capture user interaction log data at each touchpoint through the tracking SDK;
[0057] Marketing touchpoints can include mobile apps, mini-programs, offline POS terminals, official website H5 pages, SMS outreach channels, and email outreach channels.
[0058] The types of behaviors in the interaction behavior log include page browsing, button clicks, adding products to cart, order payment, customer service inquiries, coupon redemption, and activity participation. Each interaction behavior log record must include at least a user identifier, a timestamp of the behavior occurrence, and a behavior type code.
[0059] During the data collection process, behavioral log data from different touchpoints are aggregated in real time via the Kafka message queue, forming a multi-touchpoint behavioral log set. ;
[0060] Where n is the quantity index, and the set of behavior logs. Each log entry The structure is , The user identifier for the nth log entry. The timestamp of the nth log entry. The behavior type of the nth log entry;
[0061] Based on historical data, a cross-touchpoint user identifier mapping table is constructed to associate the user identifiers of the same user in the behavior logs of different touchpoints. For example, this can be achieved through a binding relationship table between device ID and member ID. Association table between device ID and mobile phone number, and correspondence table between member ID;
[0062] Subsequently, the session windows were divided according to time, and the duration threshold T of each session window was set based on the business scenario. The multi-touchpoint behavior log sets for each user were then sorted in ascending order according to the timestamp t of the behavior occurrence, resulting in a sorted behavior log sequence. After traversing the sorted sequence, consecutive behavior log sequences with a time difference less than T are grouped into the same session window to generate cross-touchpoint behavior event sequences. ;
[0063] Within each session window, the behavioral events are arranged in chronological order. This represents the k-th cross-touchpoint behavior event of the business object with user identifier u.
[0064] S2: Adapt the cross-touchpoint behavioral event sequence to obtain the input tensor, and then pre-train a multi-task neural network containing a shared bottom-level encoding layer and two parallel task heads. The input tensor is received through the shared bottom-level encoding layer and the marketing feature encoding vector is output. The prediction tasks of two marketing objectives are executed simultaneously through the two parallel task heads, and the comprehensive marketing state vector is output.
[0065] Each cross-touchpoint behavior event in the cross-touchpoint behavior event sequence S is converted into the input format of the neural network, and a multi-task neural network is pre-trained.
[0066] The multi-task neural network includes an input layer, a bottom-level encoding layer, task head 1, task head 2, and a fully connected layer.
[0067] The bottom layer is a multi-layer neural network Transformer encoder used to encode features of the cross-touchpoint behavioral event sequence S;
[0068] The parallel task header contains two parallel task headers, which correspond to two prediction tasks related to marketing objectives: user purchase conversion rate prediction and user churn risk prediction.
[0069] Adaptation processing for cross-touchpoint behavior event sequences:
[0070] The sequence of cross-touchpoint behavior events within each session window is sorted by timestamp, and the behavior type is encoded into a vector representation, such as mapping discrete behavior types into low-dimensional dense vectors through an embedding layer.
[0071] Map business object identifiers to vector representations, such as through an embedding layer for products / activities / pages;
[0072] Then, the behavior events and business object identifiers within the same session window are concatenated or stacked in chronological order to form a sequence-form input tensor X, which serves as the input to the shared underlying encoding layer.
[0073] The implementation process of the shared underlying encoding layer includes:
[0074] The shared bottom-level encoding layer receives the input tensor X and, through nonlinear transformation and feature extraction of a multi-layer neural network, captures the cross-touchpoint behavioral correlation features of behavioral event sequences, including:
[0075] The shared underlying encoding layer incorporates a Transformer encoder, a multi-head self-attention mechanism, and a feedforward neural network.
[0076] The input tensor is encoded with temporal position based on the Transformer encoder, which fully preserves the sequential temporal logic of cross-touchpoint behavior event sequences and outputs a feature tensor carrying temporal position information.
[0077] Then, the feature tensor is called, and the association weights of each behavioral event in the sequence are calculated in parallel through the multi-head attention mechanism to capture the interdependence between preceding and subsequent behavioral events over a long time span and output the global behavioral association feature matrix.
[0078] Based on the global behavior association feature matrix, the feedforward neural network performs multi-layer nonlinear mapping transformations, simultaneously completing feature vector dimension normalization and dimension unification adjustment. Finally, the marketing feature encoding vector is output from the shared bottom encoding layer. ;
[0079] Specifically, the core of the encoding process is to transform the original behavioral event features into more discriminative sequence features through attention mechanisms or recurrent structures, providing a shared underlying feature representation for subsequent task heads;
[0080] A multi-task neural network contains two parallel task heads, which encode marketing feature vectors for a shared encoding layer, respectively. Process and execute the forecasting task for two marketing objectives;
[0081] Among them, the two marketing objectives are user purchase conversion rate prediction and user churn risk prediction. The user purchase conversion rate prediction task corresponds to task head 1, and the user churn risk prediction task corresponds to task head 2.
[0082] Task 1: The encoded feature sequence is processed using built-in global pooling and a sequence-to-single-value conversion to obtain a fixed-dimensional feature vector. Then, through mapping via a fully connected layer, the predicted purchase conversion rate is output. ;
[0083] Task Header 2: Encode Marketing Feature Vectors The process involves performing average pooling to calculate the average of the features across all time steps, and then selecting the maximum value among all features to obtain a fixed-dimensional feature vector. Then, the feature vector is activated by the Sigmoid activation function. Converted to predicted churn risk probability value ;
[0084] The predicted purchase conversion rate output from Task Header 1 The predicted output loss risk value of Task Head 2 Marketing feature encoding vectors with shared encoding layers By concatenating the vectors, we obtain the comprehensive marketing state vector. ;
[0085] Specifically, this vector integrates information such as user behavior characteristics, purchase conversion tendency, and churn risk tendency, and can comprehensively represent the user's current marketing-related status as input data for S3.
[0086] S3: Perform full user classification calculation on the comprehensive marketing status vector to obtain user marketing status classification results, and generate candidate marketing intervention actions based on user marketing status classification results;
[0087] The process of obtaining user marketing status classification results by performing full user classification calculation includes:
[0088] The output of the comprehensive marketing state vector Perform multi-dimensional feature analysis to extract core marketing status features:
[0089] From the overall marketing status vector Extract the interaction frequency and average page dwell time over the past 7 days to quantify the user's recent activity level in marketing campaigns and obtain activity characteristics. u represents the corresponding user identifier;
[0090] From the overall marketing status vector Extract product category, content format, and channel preference tag sets to represent users' long-term interest and acquisition preference characteristics. :
[0091] From the overall marketing status vector Extract historical purchase amount, repurchase rate, and average order value to measure the value characteristics of users' historical contributions. ;
[0092] A hierarchical method driven by both dynamic clustering and business rules is adopted to divide users into K levels, where K is a preset number of levels in the business. The hierarchical calculation is performed on all users, and a unique user status level identifier is output. The hierarchical rules are defined as follows:
[0093] If satisfied , For empty set, as well as The user marketing status classification result is: This corresponds to high-value active users;
[0094] in, The first-level threshold for activity level, For the core preference tag set, As the first-level threshold of value, For active user lifecycle stages, This refers to the current lifecycle stage of user u.
[0095] Furthermore, the core preference tag set By extracting behavioral tags related to product categories, content formats, and channel preferences from the multi-touchpoint behavior log L over a long period of 90 days, and performing frequency statistics on these behavioral tags, tags related to high-value behaviors are selected.
[0096] Active user lifecycle stages By extracting user interaction frequency and time interval from cross-touchpoint behavioral event sequences, and obtaining them based on the active user determination threshold preset by the business side:
[0097] Interaction frequency threshold: Interactions per 7 days, meaning at least 5 interactions in the last 7 days;
[0098] Time interval threshold: "Tian" means that the most recent interaction was ≤3 days ago.
[0099] If the user's behavioral characteristics meet the interaction frequency And it takes a long time If so, the user is determined to be in the active user stage, and the corresponding stage is identified as follows: ;
[0100] If satisfied , , as well as The user marketing status classification result is: This corresponds to a growing number of active users;
[0101] in, This is the secondary threshold for activity level. For a set of growth preference tags, This is the second-order threshold for value.
[0102] Furthermore, the set of growth preference tags By extracting user behavior tags from cross-touchpoint behavior tracking logs over the past 90 days, an initial preference tag pool is formed. The frequency of user behavior corresponding to the tags is statistically analyzed over time. If it increases by more than 50% compared to the past 90 days, it indicates that the behavior corresponding to the tag has strong growth potential. Tags with strong growth potential in behavior frequency are selected to form a set of growth preference tags.
[0103] This process continues until all feature combinations and the entire user lifecycle are fully covered, completing the full user classification calculation and ultimately outputting accurate user marketing status classification results. ;
[0104] The process of generating candidate marketing interventions includes:
[0105] User marketing status classification results Corresponding purchase conversion rate forecast , churn risk probability prediction value Marketing feature encoding vector As direct input, the type of action to be performed. Decision operation:
[0106] When the purchase conversion rate is predicted The conversion rate is higher than the preset conversion threshold and the predicted probability of churn is high. When the churn rate falls below the preset threshold, a conversion-promoting action type is directly generated.
[0107] When the purchase conversion rate is predicted The conversion rate is higher than the preset conversion threshold and the predicted probability of churn is high. When the churn rate exceeds the preset threshold, a churn recovery action is directly generated.
[0108] When the purchase conversion rate is predicted When the conversion threshold is lower than or equal to the preset threshold, an activity-boosting action type is directly generated.
[0109] Subsequently, based on the user marketing status classification results Corresponding marketing feature encoding vector This directly identifies the action execution touchpoint and the business object to which the action is directed, and then combines this with the user marketing status classification results. The results of the action type, execution touchpoint, business target, and user marketing status classification are structured and combined to form a complete set of candidate marketing intervention actions. ,in, The marketing feature encoding vector output by the shared underlying encoding layer mentioned above includes execution touchpoints and business objects. Action types include conversion promotion actions, churn recovery actions, and activity enhancement actions;
[0110] Specifically, marketing feature encoding vector Output and action type from the S2 shared underlying coding layer The results are derived from full-scale user segmentation calculations, without any external independent data sources. Each parameter can be traced back to the original cross-touchpoint user behavior logs, resulting in user marketing status segmentation. User operation priority and benefit matching level were locked, and marketing feature encoding vector was locked. Built-in user touchpoint preferences, business object preferences, and action types By clearly defining the core objectives of marketing intervention, the three types of information are structured and integrated into a unified action format, fully covering all core marketing elements such as intervention objectives, execution channels, target audiences, and user level permissions, providing a standardized and unified basis for downstream compliance verification, conflict resolution, and touchpoint implementation.
[0111] S4: Assign candidate marketing intervention actions to the corresponding marketing touchpoint execution engine for execution, achieving end-to-end marketing;
[0112] The candidate marketing intervention actions output by S3 Based on the user marketing status classification results built into the action Marketing Feature Encoding Vector The bound attributes are precisely assigned to the corresponding independent marketing touchpoint execution engines, and each touchpoint execution engine is based on the action type. Complete marketing intervention and outreach operations according to a standardized instruction format;
[0113] Specifically, after receiving the assigned candidate marketing intervention actions, the execution engines at each marketing touchpoint will categorize them according to action type. The requirement is to perform the corresponding operation, for example:
[0114] The coupon distribution engine distributes coupons to users based on action parameters, such as coupon value, validity period, and distribution conditions; the push engine sends activity messages to users based on action parameters, such as push content, push time, and push touchpoints.
[0115] The product recommendation engine displays recommended products at corresponding touchpoints based on action parameters, such as the list of recommended products, the recommended location, and the recommendation algorithm.
[0116] Once execution is complete, the execution engine will record the action execution results to form an execution log, which will be used as part of the user behavior log in the subsequent S1 and participate in the next round of user behavior sequence generation and strategy optimization, thereby realizing a closed loop of full-link marketing.
[0117] Example 2:
[0118] like Figure 2 As shown, the present invention proposes a machine learning-based end-to-end marketing system, comprising:
[0119] Marketing data collection module: Collects multi-touchpoint behavior logs from multiple independent marketing touchpoints, and generates cross-touchpoint behavior event sequences based on customer identifier matching and time window session merging in the multi-touchpoint behavior logs;
[0120] State decision module: Adapts the cross-touchpoint behavioral event sequence to obtain the input tensor, and then pre-trains a multi-task neural network containing a shared bottom layer encoding layer and two parallel task heads. The shared bottom layer encoding layer receives the input tensor and outputs the marketing feature encoding vector. The two parallel task heads simultaneously execute the prediction tasks of two marketing objectives and output a comprehensive marketing state vector.
[0121] Marketing intervention module: Performs full-user classification calculation on the comprehensive marketing status vector to obtain user marketing status classification results, and generates candidate marketing intervention actions based on the user marketing status classification results;
[0122] Intervention Execution Module: Assigns candidate marketing intervention actions to the corresponding marketing touchpoint execution engine for execution, realizing end-to-end marketing.
[0123] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.
[0124] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples 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 implemented in hardware or software depends on the specific application and design constraints of the technical solution.
[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A machine learning-based end-to-end marketing method, characterized in that, include: S1: Collect multi-touchpoint behavior logs from multiple independent marketing touchpoints, and generate cross-touchpoint behavior event sequences based on customer identifier matching and time window session merging in the multi-touchpoint behavior logs; S2: Adapt the cross-touchpoint behavioral event sequence to obtain the input tensor, and then pre-train a multi-task neural network containing a shared bottom-level encoding layer and two parallel task heads. The input tensor is received through the shared bottom-level encoding layer and the marketing feature encoding vector is output. The prediction tasks of two marketing objectives are executed simultaneously through the two parallel task heads, and the comprehensive marketing state vector is output. S3: Perform full user classification calculation on the comprehensive marketing status vector to obtain user marketing status classification results, and generate candidate marketing intervention actions based on user marketing status classification results; S4: Assign candidate marketing intervention actions to the corresponding marketing touchpoint execution engine for execution, achieving end-to-end marketing.
2. The end-to-end marketing method based on machine learning according to claim 1, characterized in that, The process of collecting multi-touchpoint behavior logs includes: Deploy user behavior tracking points at multiple independent marketing touchpoints and capture user interaction logs at each touchpoint using the tracking SDK; The behavior log data from different touchpoints is aggregated in real time through the message queue Kafka to form a multi-touchpoint behavior log set.
3. The end-to-end marketing method based on machine learning according to claim 1, characterized in that, The process of generating a sequence of cross-touchpoint behavioral events includes: Based on historical data, a cross-touchpoint user identifier mapping table is constructed to associate the user identifiers of the same user in the behavior logs of different touchpoints, and the session window is divided by time as the dimension. The duration threshold T of the session window is set according to the business scenario. For each user's multi-touchpoint behavior log set, sort it in ascending order according to the timestamp t of the behavior occurrence to obtain a sorted behavior log sequence. Traverse the sorted sequence and group consecutive behavior log sequences with a time difference less than T into the same session window to generate a cross-touchpoint behavior event sequence.
4. The end-to-end marketing method based on machine learning according to claim 3, characterized in that, The process of adapting cross-touchpoint behavioral event sequences includes: The cross-touchpoint behavior event sequence within each session window is sorted by timestamp, and the behavior type in the cross-touchpoint behavior event sequence is encoded into a vector representation, and the business object identifier is mapped into a vector representation; The behavior events and business object identifiers within the same session window are concatenated or stacked in chronological order to form a sequence of input tensors.
5. The end-to-end marketing method based on machine learning according to claim 4, characterized in that, The implementation process of the shared underlying encoding layer in a multi-task neural network includes: The shared underlying encoding layer incorporates a Transformer encoder, a multi-head self-attention mechanism, and a feedforward neural network. The Transformer encoder adds temporal position encoding to the input tensor, outputting a feature tensor carrying temporal position information. Based on the feature tensor, a multi-head self-attention mechanism is used to calculate the association weights of each behavioral event in the sequence in parallel, outputting a global behavioral association feature matrix. Based on the global behavioral association feature matrix, a feedforward neural network is input to perform a multi-layer nonlinear mapping transformation, outputting a marketing feature encoding vector.
6. The end-to-end marketing method based on machine learning according to claim 5, characterized in that, The process of simultaneously performing forecasting tasks for two marketing objectives includes: The multi-task neural network contains two parallel task heads, which process the marketing feature encoding vectors of the shared encoding layer respectively, and perform two marketing objective prediction tasks: user purchase conversion rate prediction and user churn risk prediction. User purchase conversion rate prediction obtains a fixed-dimensional feature vector by performing built-in global pooling and sequence-to-single-value transformation on the encoded feature sequence. Then, through mapping via a fully connected layer, the predicted purchase conversion rate is output. User churn risk prediction involves processing the marketing feature encoding vector, performing average pooling, calculating the average value of features across all time steps, and selecting the maximum value among all time steps to obtain a fixed-dimensional feature vector. Then, the feature vector is activated by the Sigmoid activation function. Converted into a predicted probability value of churn risk; The purchase conversion rate prediction, churn risk probability prediction, and marketing feature encoding vectors from the shared encoding layer are concatenated to obtain a comprehensive marketing state vector.
7. The end-to-end marketing method based on machine learning according to claim 6, characterized in that, The process of obtaining user marketing status rating results includes: The output comprehensive marketing state vector is subjected to multi-dimensional feature analysis to extract core marketing state features, including activity features. Preference characteristics With value characteristics ; A hierarchical approach driven by both dynamic clustering and business rules is adopted to divide users into K levels. Based on the core marketing status, a full-scale user hierarchical calculation is performed to obtain the user marketing status hierarchical results. ,in, For high-value active users, This corresponds to growing active users.
8. The end-to-end marketing method based on machine learning according to claim 7, characterized in that, The process of tiered calculation for all users includes: If satisfied , For empty set, as well as The user marketing status classification result is: ;in, The first-level threshold for activity level, For the core preference tag set, As the first-level threshold of value, For active user lifecycle stages, This refers to the current lifecycle stage of user u. If satisfied , , as well as The user marketing status classification result is: ;in, This is the secondary threshold for activity level. For a set of growth preference tags, This is the second-order threshold for value.
9. A machine learning-based end-to-end marketing method according to claim 8, characterized in that, The process of generating candidate marketing interventions includes: Obtain user marketing status classification results The corresponding purchase conversion rate prediction, churn risk probability prediction, and marketing feature encoding vector; When the predicted purchase conversion rate is higher than the preset conversion threshold and the predicted churn risk probability is lower than the preset churn threshold, a conversion promotion action type is generated. When the predicted purchase conversion rate is higher than the preset conversion threshold and the predicted churn risk probability is higher than the preset churn threshold, a churn recovery action type is generated. When the predicted purchase conversion rate is lower than or equal to the preset conversion threshold, an activity boosting action type is generated directly. Based on the marketing feature encoding vector corresponding to the user marketing status classification results, the action execution touchpoint and the business object to which the action is directed are determined. Combined with the user marketing status classification results, the action type, execution touchpoint, business object, and user marketing status classification results are structurally combined to form a complete candidate marketing intervention action.
10. A machine learning-based end-to-end marketing system, implemented according to the method of any one of claims 1-9, characterized in that, include: Marketing data collection module: Collects multi-touchpoint behavior logs from multiple independent marketing touchpoints, and generates cross-touchpoint behavior event sequences based on customer identifier matching and time window session merging in the multi-touchpoint behavior logs; State decision module: Adapts the cross-touchpoint behavioral event sequence to obtain the input tensor, and then pre-trains a multi-task neural network containing a shared bottom layer encoding layer and two parallel task heads. The shared bottom layer encoding layer receives the input tensor and outputs the marketing feature encoding vector. The two parallel task heads simultaneously execute the prediction tasks of two marketing objectives and output a comprehensive marketing state vector. Marketing intervention module: Performs full-user classification calculation on the comprehensive marketing status vector to obtain user marketing status classification results, and generates candidate marketing intervention actions based on the user marketing status classification results; Intervention Execution Module: Assigns candidate marketing intervention actions to the corresponding marketing touchpoint execution engine for execution, realizing end-to-end marketing.
