Cross-channel marketing automation method and system based on real-time behavior triggering

Abstract

The invention discloses a cross-channel marketing automation method and system based on real-time behavior triggering. The method comprises the following steps: capturing a user behavior event flow in real time through a front-end acquisition unit; performing real-time mode matching on the event stream by utilizing a complex event processing engine, and generating a marketing trigger signal when a preset rule is matched; dynamically selecting an optimal channel for the user based on a real-time scoring model in response to the signal; the marketing content is sent to the selected channel through the unified execution module; and finally, tracking user feedback and updating model parameters and user preferences in real time to form closed-loop optimization. The corresponding system comprises a front-end data acquisition module, a streaming event processing module, a real-time intelligent decision-making module, a multi-channel execution module and a closed-loop optimization module which are used for implementing the steps. According to the method, millisecond-level intention perception, cost and experience optimal channel intelligent decision and data-driven adaptive optimization closed loop are realized, and the problems of delayed response, extensive channel and lack of self-adaption of traditional marketing are solved.

Description

Technical Field

[0001] This invention relates to the field of digital marketing technology, specifically to a cross-channel marketing automation method and system based on real-time behavior triggering. Background Technology

[0002] In today's digital economy era, the depth and breadth of interaction between enterprises and users are constantly increasing. Customer Relationship Management (CRM) and precision marketing have become core means for enterprises to enhance competitiveness and optimize return on investment (ROI). Traditional marketing automation systems are usually built on classic data warehouse and batch processing technology architectures. Their working mode can be summarized as "T+1" or periodic scheduled task mode. A typical existing technical solution usually includes: (1) Offline data collection and integration; the system collects behavioral data (such as page browsing, clicks, and purchases) generated by users at different touchpoints (such as websites and mobile applications) through log embedding and other methods. This data is not processed in real time, but is uploaded to a central data storage system (such as Hadoop distributed file system or relational database) for aggregation on a daily or fixed time interval basis. (2) Asynchronous user profile calculation; after the data collection is completed, the system starts offline batch processing calculation tasks during the off-peak business period (such as every night). The calculation delay of this process is usually several hours to a day. (3) Manual strategy configuration and audience selection: Marketing operations personnel manually combine tag conditions (such as "added to cart in the past 30 days" and "user level is VIP") in the management backend based on user profile tags generated the previous day to filter target audience. Then, they uniformly configure fixed marketing content (such as coupons) and preset outreach channels (such as SMS or email) for this audience and set a future delivery time. (4) Batch outreach execution and effect lag evaluation: At the scheduled time, the system pushes marketing instructions to the designated channels for centralized delivery. User behavior feedback needs to go through a full cycle before it can flow back to the system, resulting in a serious lag in strategy adjustment.

[0003] However, because the core processes rely on offline batch processing, the system cannot perceive and respond to users' real-time intentions. For example, when a user repeatedly browses the details page of a high-value product in the current session, showing a strong purchase intention but ultimately hesitating and not placing an order, the system cannot intervene at the "golden moment" when the user's intention is strongest. By the time the batch processing is completed the next day, the user may have already churned or completed the purchase on a competitor's platform, resulting in the best marketing retention opportunity being completely missed, and the marketing effect being greatly reduced. Furthermore, the channel selection in the existing solutions is usually targeted at specific marketing campaigns, using a single high-cost channel (such as SMS) to "cast a wide net" to all target user groups. This approach completely ignores the differences in channel preferences of individual users, as well as the differences in physical costs and reach efficiency. This results in wasted marketing budgets and reduced return on investment, and may also cause user resentment due to inappropriate channel reach, damaging the user experience. In addition, the strategy optimization of the existing system relies heavily on the experience of operations personnel and manual analysis. The cycle from effect feedback to strategy adjustment is long and inefficient, failing to form an automatic, real-time, and continuous optimization loop. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a cross-channel marketing automation method and system based on real-time behavior triggering. It overcomes the deficiencies of existing technologies, is rationally designed, and achieves millisecond-level intent perception, optimal channel intelligent decision-making in terms of cost and experience, and a data-driven adaptive optimization closed loop. It solves the problems of slow response, extensive channel management, and lack of adaptability in traditional marketing.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A cross-channel marketing automation method based on real-time behavior triggers includes the following steps:

[0007] S1: Real-time capture; Through the front-end data acquisition unit deployed on the user terminal, the user's native behavioral events on the digital interactive interface are captured in real time to form a continuous behavioral event stream; and the behavioral event stream is transmitted to the stream processing platform in real time via message middleware.

[0008] S2: Real-time triggering; Continuous monitoring and pattern matching of the behavioral event stream using a complex event processing engine; The complex event processing engine has at least one configurable marketing trigger rule pre-set, which defines the sequence of behavioral patterns to be matched and the associated context conditions; When the data in the behavioral event stream satisfies any of the marketing trigger rules, a marketing trigger signal containing user identifier, trigger rule identifier and real-time context information is generated;

[0009] S3: In response to the marketing trigger signal, initiate a real-time decision-making process to determine an optimal outreach channel from multiple alternative outreach channels for the current user; the decision-making process includes:

[0010] S301: Real-time acquisition or calculation of the estimated conversion rate of the current user for each alternative outreach channel. The expected revenue (R) of this marketing campaign, the cost per reach (C) of each channel, and the interference penalty value (D) based on the frequency of recent user contact.

[0011] S302: According to the formula Calculate a dynamic utility score S for each alternative outreach channel;

[0012] S303: Select the candidate outreach channel with the highest dynamic utility score S as the optimal outreach channel;

[0013] S4: Cross-channel execution; Based on the result of step S3, generate a standardized API request corresponding to the optimal reach channel, call the service interface of that channel, and send personalized marketing content to the current user;

[0014] S5: Closed-loop feedback learning; tracking and attributing the reach of the marketing content, collecting user feedback behavior data for this marketing campaign; based on the feedback behavior data, adjusting the data used to calculate the estimated conversion rate in real time using a streaming update method. Model parameters and / or user channel preference parameters.

[0015] Preferably, the native behavioral events include explicit events triggered by user active operations and implicit intent events inferred by the front-end data acquisition unit through analysis of user interaction sequences; the implicit intent events include at least one of the following: cumulative dwell time in a specific page area, non-linear page scrolling patterns, repeated viewing of specific interface elements, and heatmap data formed by mouse movement trajectories.

[0016] Preferably, the marketing triggering rule in step S2 supports logical association and state matching of event streams across multiple independent data sources; the context conditions include one or more of the following: real-time user attributes, dynamic business indicators, and environmental state conditions; wherein, the dynamic business indicators include real-time inventory quantity of goods, price change flags within the session, and the number of remaining seats in the activity.

[0017] Preferably, in step S3, the estimated conversion rate The conversion rate is calculated in real time through an online machine learning inference service. The feature vectors used include at least: user historical behavior features, current session context features, candidate channel historical performance features, and user-channel combination historical interaction features.

[0018] Preferably, the online machine learning inference service uses a gradient boosting tree model or a deep neural network model as the prediction model; the prediction model uses historical marketing outreach records as samples and labels whether the user completes the target conversion behavior within a preset time window, and performs offline training and periodic online updates.

[0019] Preferably, in step S3, the expected revenue R of this marketing campaign is dynamically calculated according to the following formula:

[0020] ;

[0021] Where P is the price of the core product or service associated with the current triggering scenario, M is the benchmark gross profit margin of the category to which the product or service belongs, and I is the incentive amount to be provided to the user.

[0022] Preferably, the interference penalty value D is calculated according to the following formula:

[0023] ;

[0024] Where N is the total number of marketing outreaches received by the current user within the most recent preset statistical period; As the baseline penalty weight, As a growth factor, and .

[0025] Preferably, in step S5, the user's preference weight for a specific channel k is updated using a time-decaying exponentially weighted moving average method. The updated formula is as follows:

[0026] If positive feedback is received from users through channel k, then:

[0027] ;

[0028] If no feedback is received or negative feedback is received during this outreach through channel k, then:

[0029] ;

[0030] in, The preference weights for channel k before the update; The preference weights for the updated channel k; This is the decay coefficient, with a value range of (0,1), used to control the rate at which historical weights are forgotten; For positive feedback gain factor, It is a negative feedback penalty factor.

[0031] This invention also discloses a cross-channel marketing automation system for implementing the above method. The system adopts a distributed architecture and includes the following modules that interact with data via an internal network or message middleware:

[0032] The front-end data acquisition unit is deployed in the form of a front-end software development kit or embedded monitoring code. The front-end data acquisition unit is configured to capture user interaction behavior on the interface in a non-blocking, asynchronous manner locally on the front end during the user session, and generate structured atomic event logs. The atomic event logs include at least the event type, timestamp, user anonymity identifier, current page context, and event association parameters. The front-end data acquisition unit is also used to push the atomic event logs to a specified message receiving endpoint on the back end in real time or near real time through a lightweight network protocol.

[0033] The streaming event processing module is communicatively connected to the output of the front-end data acquisition unit and includes a unified event access gateway, a distributed message queue, and a complex event processing engine cluster.

[0034] The unified event access gateway receives event logs from different front-end sources, performs format verification, basic cleaning and standardization, and outputs a standardized event stream in a unified format. The distributed message queue is connected to the unified event access gateway and is used to buffer, store and distribute the standardized event stream to ensure high throughput and sequentiality of event data. The complex event processing engine cluster subscribes to event streams from the distributed message queue. The complex event processing engine cluster carries a dynamically loadable rule set, where each rule defines a complex pattern composed of multiple atomic events arranged in a specific logical and temporal order. The complex event processing engine cluster is configured to slide a time window on a continuous event stream, perform pattern matching and state calculation in real time, and when a complete pattern matching any rule is detected, a structured marketing trigger instruction is generated and published to the downstream decision message topic.

[0035] A real-time intelligent decision-making center, connected to the streaming event processing module, includes a real-time feature library, an online prediction model, and a decision engine. The real-time feature library connects to a user profile database, a behavior log database, and real-time computing storage. Upon receiving a decision request, it queries or calculates in real-time multi-dimensional feature vectors related to the current user, the current scenario, and all candidate channels in a low-latency manner. The online prediction model interfaces with the real-time feature library. The online prediction model receives the multi-dimensional feature vectors and infers in real-time the probability that the current user will achieve the target conversion through each candidate channel. The decision engine is configured to: receive the marketing trigger instruction; call the real-time feature library and online prediction model to obtain all dynamic parameters required for calculation; calculate the comprehensive utility score for each candidate channel according to the preset decision logic and scoring formula; execute the optimization logic based on the scoring results to generate the final channel selection decision and associated marketing content identifier;

[0036] The multi-channel adaptation and execution layer is connected to the output interface of the real-time intelligent decision center. This layer includes an instruction standardization and routing module and a channel adapter cluster. The instruction standardization and routing module receives the decision results output by the dynamic decision engine and encapsulates them into internally unified, channel-independent marketing execution instructions. The channel adapter cluster consists of multiple independent channel adapter microservices, each encapsulating the protocol logic for communicating with an external reach channel service. The channel adapter cluster subscribes to instructions published by the instruction standardization and routing module and, based on the channel type identifier in the instructions, converts the unified internal instructions into the protocol format and parameter structure required by the specific external channel and performs authentication. Then, it calls the channel's delivery interface to complete the actual delivery of marketing content.

[0037] The data loop and optimization module, connected to the multi-channel adaptation and execution layer and the real-time intelligent decision center, includes a full-link tracking and attribution unit, a real-time feedback stream processing pipeline, and a model and strategy optimizer. The full-link tracking and attribution unit assigns a globally unique tracking identifier to each outreach action when the multi-channel adaptation and execution layer generates execution instructions, embedding this identifier into a traceable link or event point. The attribution unit collects subsequent user interaction stream data and, by matching the tracking identifier, accurately attributes conversion behavior to specific marketing outreach events. The real-time feedback stream processing pipeline continuously handles the positive and negative feedback event streams after consumption attribution. The model and strategy optimizer is connected to the real-time feedback stream processing pipeline and the real-time intelligent decision center. The model and strategy optimizer is configured to: based on the feedback event stream, update the user dynamic preference features in the real-time feature library in an incremental or small-batch manner in real time; and periodically or triggerively initiate a retraining process to update the model parameters in the online prediction model using the accumulated new feedback data, completing the system's self-evolutionary loop.

[0038] Preferably, the reach channel types supported by the multi-channel adaptation and execution layer include: mobile push service, SMS gateway, email server, WebSocket in-site messaging service, and instant messaging application interface.

[0039] This invention provides a cross-channel marketing automation method and system based on real-time behavior triggering, which has the following beneficial effects:

[0040] (1) Real-time and precise intervention is achieved: Through front-end millisecond-level behavior collection and streaming complex event processing technology, the marketing response delay is shortened from the traditional hour or day level to the second or even millisecond level, which can reach users at the decision-making critical point when their intentions are strongest, greatly improving the instant conversion rate of high-intent users.

[0041] (2) Achieved intelligent channel decision-making with optimal cost and experience: By introducing a dynamic utility scoring model based on multi-dimensional real-time calculation, the model comprehensively calculates the estimated conversion rate, the expected revenue based on the scenario, and the interference penalty. This enables the selection of the most effective reach channel for each user, thereby significantly reducing the unit conversion cost while ensuring effectiveness.

[0042] (3) A data-driven automated closed loop has been constructed: Through the end-to-end tracking attribution and real-time feedback learning mechanism, the system can automatically update user preferences and prediction model parameters according to the marketing effect, realize the adaptive optimization of the strategy, reduce the dependence on human experience, and enable the marketing effect to be continuously improved. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of this invention or the prior art will be briefly introduced below.

[0044] Figure 1 The flowchart of the cross-channel marketing automation method based on real-time behavior triggering in this invention;

[0045] Figure 2 Overall architecture diagram of the cross-channel marketing automation system of this invention;

[0046] Figure 3 Flowchart of the closed-loop learning and optimization mechanism in this invention;

[0047] Figure 4 A simplified flowchart of multi-channel adaptation and execution in this invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0049] Example 1, as Figures 1 to 4 As shown, a cross-channel marketing automation method based on real-time behavior triggering is presented. Assume user A is browsing products on the webpage of an e-commerce platform. The system of this invention has been deployed on this platform, and all modules are operating normally.

[0050] Specifically, the following steps are included:

[0051] S1: Real-time capture; When a user searches on their device, the data collection unit deployed on the front end of the e-commerce platform website (usually integrated as a JavaScript SDK) monitors every interaction of user A in real time. When user A enters the product details page of a "burgundy high-waisted long-sleeved dress" priced at $39, the SDK begins to capture a series of native behavioral events, including:

[0052] Explicit events: "page view", "image click", "add to cart button click".

[0053] Implicit intent events: By analyzing the interaction sequence, the data collection unit inferred that "the cumulative time spent in the product specification parameter area reached 120 seconds", "there was repeated mouse hovering behavior in the price display area", and "after adding the product to the shopping cart, the user scrolled back to the delivery information area within 3 minutes and stayed there for a long time".

[0054] These events, which include timestamps, user IDs, page URLs, and event parameters, are encapsulated in real time and sent asynchronously to a backend message middleware (such as Apache Kafka) to form a continuous stream of behavioral events, providing a data source for subsequent real-time analysis.

[0055] S2: Real-time trigger;

[0056] The system utilizes a complex backend event processing engine (such as Esper or Flink CEP) to subscribe to the behavioral event streams in the aforementioned message middleware. The engine has a pre-defined configurable marketing trigger rule named "Hesitation After Adding to Cart for High-Priced Items." This rule defines the behavioral pattern sequence and contextual conditions as follows:

[0057] {Event A: Add item to cart} → {Within time window T (e.g., 5 minutes), Event B: Cumulative dwell time on the cart page or product key information page (e.g., price, shipping) > 90 seconds} AND {Contextual conditions: Product price > $30 and average order value > $100}

[0058] When the complex event processing engine detects that user A's behavioral event flow perfectly matches the rule (i.e., adding a $39 dress to their cart and hesitating on the relevant page for more than 90 seconds within the next 5 minutes), it immediately generates a marketing trigger signal. This signal includes: user identifier, trigger rule ID, and real-time context information (such as trigger product ID, price, current time, etc.).

[0059] Step S3: Dynamic Channel Decision-Making Process;

[0060] In response to the marketing trigger signal in step S2, the decision-making module initiates a real-time decision-making process for user A, with the goal of selecting the optimal channel from alternative channels (assuming "APP push," "SMS," and "email"). The decision-making process specifically includes:

[0061] S301: Real-time parameter acquisition and calculation;

[0062] 1) Estimated conversion rate ( The decision-making module calls the online machine learning inference service to request calculations of the probability that user A will complete the "add-to-cart and pay" conversion through various channels. Upon receiving the request, the inference service obtains the feature vectors needed for calculation from the real-time feature platform, for example:

[0063] User characteristics: User A's historical purchase frequency, preference score for clothing categories, and activity level in the past 7 days.

[0064] Contextual features: Current product price ($39), current device type (PC), current time period (evening).

[0065] Channel characteristics: average click-through rate of "APP push" in the "high-priced hesitant" scenario in the past hour, and average conversion rate of "SMS" in the same scenario.

[0066] Cross-feature: User A has a historical click-through rate of 2% for "SMS" marketing and 15% for "APP push notifications," but did not click on the most recent APP push notification.

[0067] Based on the above characteristics, the loaded gradient boosting tree model performs millisecond-level inference and outputs the predicted probability: , , .

[0068] 2) Expected return (R): According to the formula Dynamic calculation.

[0069] Where P is the product price ($39), M is the benchmark gross profit margin for the apparel category (assumed to be 20%), and I is the incentive amount to be provided to users (e.g., to promote conversion, a $2 coupon will be issued).

[0070] Therefore, the expected return is calculated to be R = $5.8.

[0071] 3) Cost per reach (C) for each channel: Read the fixed cost for each channel from the system configuration:

[0072] C(app) = $0, C(sms) = $0.05, C(email) = $0.001.

[0073] 4) Interference Penalty Value (D) Based on Recent User Outreach Frequency: To prevent multiple complex calculations (consuming excessive server resources) from being performed on the same user within a short period, a penalty value based on the recent request frequency of the user is introduced. The query shows that user A has received 2 marketing outreaches today. According to the formula... Calculation (assuming system preset) =0.1, =0.2); thus, the interference penalty value is obtained. Dollar.

[0074] S302: Calculate dynamic utility score;

[0075] Substitute the above parameters into the formula Calculate the ratings for each channel:

[0076] ;

[0077] ;

[0078] ;

[0079] S303: Optimal strategy selection;

[0080] Select dynamic utility rating The highest-scoring feature fusion strategy will be selected as the optimal feature fusion strategy for this retrieval. For example, by comparing the scores from the various channels mentioned above, The highest. Therefore, the output of the decision-making process is: select APP push as the optimal reach channel for this marketing campaign for user A, and associate it with the predefined "high-priced customer hesitant retention" marketing copy and a $2 coupon.

[0081] S4: Cross-channel execution;

[0082] Upon receiving the decision result, the unified execution module activates its internal "APP Push Channel Adapter," which converts internal instructions into the standardized API request format required by the corresponding mobile push service provider (such as Getui or JPush), filling in the device identifier, push title, text content, and redirect link. After a successful call, a push notification containing a $2 coupon is delivered to user A's phone in real time.

[0083] S5: Closed-loop feedback learning;

[0084] 1) Tracking and Attribution: A unique tracking parameter was embedded in the push notification link. 30 minutes later, user A clicked the push notification, returned to the app, used the coupon, and completed the $39 order. The closed-loop optimization module accurately attributed the conversion to this app push marketing campaign by matching the order with the tracking parameter.

[0085] 2) Streaming parameter updates: Based on this positive feedback, the system immediately initiates the update process:

[0086] Update user channel preference weights: Update user A's "APP push" channel preference weights using an exponentially weighted moving average method with time decay. Assuming the original weights System parameters , The new weights are: The weight has been increased.

[0087] Updated Model Features: This successful "user-channel-scenario" combination (user A, app push notification, hesitation due to high average order value) is treated as a positive sample and added to the real-time sample stream. This data is used for incremental updates or periodic retraining of the online machine learning model, enabling it to perform better in similar scenarios in the future. The forecasts are more accurate.

[0088] Thus, a complete automated marketing loop of "real-time perception - dynamic channel decision-making - precise reach - feedback optimization" is completed within minutes. Compared to the traditional T+1 model, this invention utilizes a Complex Event Processing (CEP) engine for streaming real-time matching. This allows the system to instantly identify moments of high intent revealed by users during interactions and initiate marketing intervention within seconds. Compared to existing technologies that delay for at least several hours or even a day, this invention shortens marketing response time from "days / hours" to "seconds," enabling precise reach during the "golden window" when users' purchasing intent is strongest. This effectively retains potential churned customers, directly converting real-time intent into purchasing behavior, and significantly improving the immediate conversion rate and ROI of marketing campaigns.

[0089] Furthermore, this invention introduces a dynamic utility scoring model based on multi-dimensional real-time calculation. This model not only considers the fixed costs of the channel but also integrates personalized predicted conversion rates derived through real-time machine learning reasoning. The system employs a dynamic, scenario-based calculation of expected returns (R) and a non-linear interference penalty (D) to prevent excessive disruption. This allows for the objective selection of the most cost-effective outreach channels for each user in a specific scenario, significantly reducing unit conversion costs while ensuring effectiveness. A complete data-driven closed loop is established. The system achieves accurate performance attribution through unique tracking identifiers and utilizes streaming update technology to process feedback data in real time. A time-decaying weighted average algorithm is used to update the user's preference weight for each channel in real time, enabling the system to quickly adapt to changes in user interests and habits. The entire "perception-decision-execution-feedback" process is fully automated. The system can automatically strengthen effective strategies and weaken or eliminate ineffective ones, thus achieving a leap from static configuration to dynamic self-optimization of marketing strategies, reducing reliance on human intervention and ensuring continuous improvement in long-term marketing effectiveness.

[0090] Example 2, as Figure 2 As shown, this invention also discloses a cross-channel marketing automation system for implementing the above method. The system adopts a distributed architecture and includes the following modules that interact with data through an internal network or message middleware (such as Kafka, gRPC):

[0091] Front-end data acquisition unit (data perception layer);

[0092] Deployed on the user client in the form of a front-end software development kit or embedded monitoring code, it consists of the following components:

[0093] 1) Lightweight Event Tracking SDK: Embedded in websites, mobile applications, or mini-programs as a library file or script. This SDK includes event listeners, local data handlers, and asynchronous senders. Event listeners monitor user native interaction events such as clicks, scrolling, hovering, page dwell, and foreground / background switching in a non-intrusive manner on the front end.

[0094] 2) Context capture engine: works in conjunction with listeners to capture and associate rich context information in real time when an event occurs, including but not limited to the current page URL or view identifier, product SKU, page scroll depth percentage, mouse coordinate trajectory relative to key elements, current network type and device model.

[0095] 3) Local Queue and Sending Strategy Manager: This manager is responsible for packaging event data and context into structured logs, initially storing them in a persistent queue on the client's local machine, and then, based on network conditions and throttling strategies, sending them asynchronously and in batches to the backend unified event access gateway via HTTPS or WebSocket long connections. This design ensures the reliability and real-time performance of data reporting without impacting the performance of the frontend main thread.

[0096] Streaming event processing module;

[0097] Communication connection with the output of the front-end data acquisition unit includes:

[0098] 1) Unified Event Access Gateway: Used to receive data from various front-end SDKs, perform authentication, format parsing, basic field validation and data anonymization, and output standardized user behavior event streams.

[0099] 2) Distributed Message Queue: Employing a streaming platform such as Apache Kafka or Pulsar, this queue persists and buffers the standardized event stream. It partitions events by topic, ensuring the sequential nature of events for the same user and providing downstream consumers with high-throughput, low-latency data access capabilities.

[0100] 3) Complex Event Processing Engine Cluster: As the core computing unit, it subscribes to event streams in the message queue. This engine loads a dynamic rule base, with each rule defined using a SQL-like or declarative language, describing the complex event patterns that need to be matched (e.g., event B did not occur within X seconds after event A, and event C occurred at least N times). The engine employs sliding time windows and stateful computation techniques to perform pattern matching on the stream in real time. When a rule is fully matched, a structured marketing trigger signal is immediately generated. This signal contains the rule ID, the matched event sequence, the user ID, and a complete snapshot of the trigger moment context, and is published to a dedicated "Decision Request" topic.

[0101] Real-time intelligent decision-making center;

[0102] Connected to the streaming event processing module, it is used to make optimal decisions within milliseconds after receiving a trigger signal, and consists of the following components:

[0103] 1) Real-time Feature Library: Connects distributed caches (such as Redis) and real-time data warehouses (such as Druid). When a decision request arrives, this real-time feature library performs hundreds of feature queries and real-time calculations in parallel. For example, it obtains user behavior statistics over the past hour, current session path, historical conversion rates of each channel over the past 30 days, real-time RFM tiers for users, and price ranges for current products, and assembles them into the real-time feature vector required for decision-making within hundreds of milliseconds.

[0104] 2) Online Prediction Model: Loaded with a high-performance prediction model (such as an optimized XGBoost or TensorFlowServing model). This online prediction model receives feature vectors from the feature service, performs real-time forward inference, and outputs the estimated probability that the user will complete the target conversion (such as click, purchase) through each candidate channel (SMS, push, email, etc.). The model supports A / B testing and canary releases.

[0105] 3) Decision Engine: Acts as the coordinator of the decision-making process. Its workflow is as follows: a) Listens to the "Decision Request" topic and obtains marketing trigger events. b) Synchronously calls the aforementioned feature services and model inference services to obtain all necessary parameters. c) Dynamically obtains business parameters such as the current cost (C) of each channel and the current profit margin (M) of the product from the configuration center. d) Based on the formula... The real-time utility score (S) for each channel is calculated in parallel. e) The optimization logic is executed to select the channel corresponding to the highest score, and the preset marketing copy template is associated to generate the final cross-channel marketing execution instruction.

[0106] Multi-channel adaptation and execution layer;

[0107] Connected to the output interface of the real-time intelligent decision-making center, it accurately delivers decision instructions to users; the key lies in decoupling and fault tolerance; specifically including:

[0108] 1) Command Standardization and Routing Module: This module receives execution commands from the decision engine and routes them to the corresponding channel adapter microservice based on the channel type (e.g., mobile push service, SMS gateway, email server, WebSocket in-site messaging service, instant messaging application interface). This module is responsible for command priority sorting, flow control, and initial format validation.

[0109] 2) Channel Adapter Cluster: This consists of multiple independent channel adapters, each encapsulating the interaction details with a specific external channel (such as AWS Cloud SMS, SendGrid email, OneSignal push, Facebook Messenger, etc.). The internal implementation of the adapter includes:

[0110] Protocol converter: Converts internal unified instructions into specific protocols (such as HTTP / HTTPS, SMTP, TCP private protocols) and data formats (such as JSON, XML) required by external APIs.

[0111] Authentication and Security Manager: Manages authentication information such as keys and tokens required by the channel, and is responsible for request signing.

[0112] Retry and degradation mechanism: When an external channel call fails, it will be retried according to the policy; when the channel is unavailable, degradation logic can be triggered (such as email being downgraded to internal message).

[0113] Response handler: Parses responses from external channels, converts them into internal statuses (success, failure, restricted), and writes them back to the log.

[0114] 3) Real-time Reach Status Tracker: Simultaneously with the issuance of the command, a globally unique trace_id is generated and embedded into the SMS link or push notification. This tracker records all association information between the trace_id and the marketing event, preparing for subsequent attribution.

[0115] Data closed-loop and optimization module;

[0116] Connected to multi-channel adaptation and execution layers and a real-time intelligent decision-making center, including:

[0117] 1) End-to-end tracking and attribution: By monitoring the user's subsequent behavior flow (such as clicking a link with trace_id, completing an order), the conversion behavior is accurately attributed to the specific marketing touch event through trace_id or message_id, forming a closed-loop record of "touch-feedback".

[0118] 2) Real-time feedback stream processing pipeline: used to write attribution results (positive: click, conversion; negative: ignore, unsubscribe) as a new real-time event stream into the message queue.

[0119] 3) Model and Policy Optimizer: Used to subscribe to feedback streams. Upon receiving user feedback, it immediately updates the user's real-time characteristics in the distributed cache, such as "preference weight for SMS channels" and "recent marketing sensitivity." This sub-second update allows the next decision to be based on the user's latest changes in interest.

[0120] 4) Offline Model Training and Deployment Platform: This platform is used to periodically (e.g., daily) combine accumulated feedback data with historical feature data to initiate offline training pipelines, producing new, higher-performing prediction models. After training, the new model is smoothly updated to the online machine learning inference service using blue-green or canary deployment strategies, completing the iterative evolution of the model.

[0121] This invention utilizes millisecond-level reporting from the front-end data acquisition unit and real-time CEP matching from the streaming event processing module. The system can identify users' micro-level intentions such as "hesitation," "price comparison," and "deep interest" at the kernel level and initiate intervention within seconds or even milliseconds. This effectively solves the severe lag problem caused by the traditional "T+1" batch processing mode, allowing it to exert influence at the user's decision-making critical point, thereby significantly improving the conversion rate of high-intent users and reducing order loss due to slow response. The online prediction model in the real-time intelligent decision center comprehensively considers real-time estimated conversion rate, channel physical costs, user historical preferences, and real-time disturbance level, automatically selecting the most cost-effective outreach method through a multi-objective optimization algorithm. Compared to the traditional reliance on manual experience and a "broad-based" approach to channel selection, this avoids ineffective and experience-damaging operations such as sending expensive SMS messages to users who dislike SMS. While achieving the same or higher conversion goals, marketing channel costs are reduced by more than 20%, while user satisfaction and long-term lifetime value are improved by reducing unnecessary disturbance to users.

[0122] This invention decouples complex marketing automation processes into single-responsibility, independently deployable, and scalable services. The multi-channel adaptation and execution layer uses an adapter pattern to uniformly manage heterogeneous channel interfaces; adding a new channel requires only the development of a new adapter, without altering the core decision-making logic, resulting in extremely high system scalability. Simultaneously, asynchronous message communication and fault-tolerant design between modules ensure high system availability when facing traffic spikes or component failures.

[0123] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time behavior-triggered cross-channel marketing automation, characterized in that: The method comprises the following steps: S1: Real-time capture; through a front-end data acquisition unit deployed on a user terminal, real-time capture of native behavior events of a user on a digital interaction interface to form a continuous behavior event stream; and real-time transmission of the behavior event stream to a stream processing platform via a message middleware; S2: Real-time triggering; continuous monitoring and pattern matching of the behavior event stream by using a complex event processing engine; at least one configurable marketing triggering rule is pre-stored in the complex event processing engine, and the marketing triggering rule defines a behavior pattern sequence to be matched and associated context conditions; when data in the behavior event stream meets any of the marketing triggering rules, a marketing triggering signal containing a user identifier, a triggering rule identifier, and real-time context information is generated; S3: In response to the marketing triggering signal, a real-time decision-making process is started to determine an optimal touch channel for the current user from multiple alternative touch channels; the decision-making process comprises: S301: Real-time acquisition or calculation of the estimated conversion rate of the current user for each alternative touch channel the expected revenue R of this marketing, the single-touch cost C of each channel, and the interference penalty value D based on the user's recent touch frequency S302: Calculate a dynamic utility score S for each alternative reach channel according to the formula S = (a * R + b * C + c * D + d * E + e * F + f * G + h * H + i * I + j * J + k S303: Selecting an alternative touch channel with the highest dynamic utility score S as the optimal touch channel; S4: Cross-channel execution; according to the result of step S3, a standardized API request corresponding to the optimal touch channel is generated, a service interface of the channel is called, and personalized marketing content is sent to the current user; S5: closed-loop feedback learning; tracking and attribution of the reach effect of the marketing content, collecting feedback behavior data of the user for the marketing, and based on the feedback behavior data, adjusting the model parameters and / or the user channel preference parameters used to calculate the estimated conversion rate in a streaming update manner . 2.The real-time behavior triggered cross-channel marketing automation method of claim 1, wherein: The native behavior events include explicit events triggered by user active operations and implicit intent events inferred by the front-end data acquisition unit through analysis of user interaction sequences. The implicit intent events include at least one of the following: cumulative stay time in a specific page area, non-linear page scrolling pattern, repeated viewing of specific interface elements, and heat map data formed by mouse movement tracks. 3.The real-time behavior-triggered cross-channel marketing automation method of claim 1, wherein: The marketing triggering rule in step S2 supports logical association and state matching of event streams across multiple independent data sources; the context conditions include one or more of user real-time attributes, dynamic business indicators, and environmental state conditions; wherein the dynamic business indicators include real-time inventory quantity of goods, session price change flag, and number of remaining seats for activities.

4. The method of claim 1, wherein the method further comprises: In step S3, the estimated conversion rate is obtained by real-time computation via an online machine learning inference service; the estimated conversion rate is computed The feature vector used includes at least: user historical behavior features, current session context features, candidate channel historical performance features, and user-channel combination historical interaction features.

5. The method of claim 4, wherein the method further comprises: The prediction model used by the online machine learning inference service is a gradient boosting tree model or a deep neural network model; the prediction model takes historical marketing touch records as samples and whether the user completes the target conversion behavior within a preset time window as a label for offline training and periodic online updating.

6. The method of claim 1, wherein the method further comprises: In step S3, the expected revenue R of this marketing is dynamically calculated according to the following formula: ； Where P is the price of the core product or service associated with the current triggering scenario, M is the benchmark gross profit rate of the product or service, and I is the incentive amount to be provided to the user.

7. The method of claim 1, wherein the method further comprises: The interference penalty value D is calculated according to the following formula: ； N is the total number of marketing touches received by the current user in a recent preset statistical period; is a baseline penalty weight, is a growth factor, and . 8.The real-time behavior-triggered cross-channel marketing automation method of claim 1, wherein: In step S5, the preference weight of the user for the specific channel k is updated using the exponentially weighted moving average method with time decay ; the update formula is as follows: If the user provides positive feedback through channel k this time, then: ； If the user does not provide feedback or provides negative feedback through channel k this time, then: ； wherein, is the preference weight of channel k before update; is the preference weight of channel k after update; is the decay coefficient, with a value range of (0, 1), used to control the forgetting speed of the historical weight; is the positive feedback gain factor, is the negative feedback penalty factor.

9. A cross-channel marketing automation system for implementing the method of any one of claims 1 to 8, characterized by: The system adopts a distributed architecture and comprises the following modules that interact with each other through internal networks or message middleware: A front-end data collection unit in the form of a front-end software development kit or embedded monitoring code is deployed, which is configured to capture user interaction behavior on the interface in an asynchronous manner without blocking in a local front-end during a user session, and generate a structured atomic event log; The atomic event log at least contains event type, timestamp, user anonymous identification, current page context, and event associated parameters; The front-end data collection unit is also used to push the atomic event log to a back-end designated message receiving end in real time or near real time through a lightweight network protocol; A streaming event processing module is connected to the output end of the front-end data collection unit, including a unified event access gateway, a distributed message queue, and a complex event processing engine cluster; The unified event access gateway is used to receive event logs from different front-end sources, perform format verification, basic cleaning, and standardization, and output a standardized event stream in a unified format; The distributed message queue is connected to the unified event access gateway, used to buffer, store, and distribute the standardized event stream, and ensure high throughput and sequence of event data; the complex event processing engine cluster subscribes to the event stream from the distributed message queue; the complex event processing engine cluster carries a dynamically loadable rule set, wherein each rule defines a complex pattern composed of multiple atomic events according to specific logic and time sequence; the complex event processing engine cluster is configured to slide a time window on the continuous event stream, perform pattern matching and state calculation in real time, and when a complete pattern matching any rule is detected, a structured marketing trigger instruction is generated and published to the downstream decision-making message topic; A real-time intelligent decision center is connected to the streaming event processing module, including a real-time feature library, an online prediction model, and a decision engine; The real-time feature library connects a user portrait database, a behavior log database and a real-time calculation storage, is used for inquiring or calculating a multi-dimensional feature vector related to a current user, a current scene and all candidate channels in a low delay manner when a decision request is received; the online prediction model is connected with the real-time feature library; the online prediction model is used for receiving the multi-dimensional feature vector and inferring a probability value of a current user achieving a target conversion through each candidate channel in real time ; the decision engine is configured to receive the marketing trigger instruction; The real-time feature library and the online prediction model are called to obtain all dynamic parameters required for calculation; according to the preset decision logic and scoring formula, a comprehensive utility score is calculated for each candidate channel; Based on the scoring results, an optimal selection logic is executed to generate a final channel selection decision and associated marketing content identification; A multi-channel adaptation and execution layer is connected to the output interface of the real-time intelligent decision center, including an instruction standardization and routing module and a channel adapter cluster; the instruction standardization and routing module is used to receive the decision results output by the dynamic decision engine, and encapsulate them into internal unified and channel-independent marketing execution instructions; the channel adapter cluster is composed of multiple independent channel adapter microservices, each of which encapsulates protocol logic for communicating with an external touch channel service; the channel adapter cluster subscribes to the instructions published by the instruction standardization and routing module, and according to the channel type identification in the instructions, converts the unified internal instructions into the protocol format, parameter structure required by a specific external channel, and performs identity authentication, and then calls the delivery interface of the channel to complete the actual delivery of marketing content. The data closed loop and optimization module is connected with the multi-channel adaptation and execution layer and the real-time intelligent decision center, and comprises a full-link tracking and attribution device, a real-time feedback stream processing pipeline and a model and strategy optimizer. The full-link tracking and attribution device assigns a globally unique tracking identifier to each touch action when the multi-channel adaptation and execution layer generates an execution instruction, and embeds the identifier in a traceable link or a buried point. The attribution device is used to collect subsequent interaction stream data of a user, and accurately attributes a conversion behavior to a specific marketing touch event by matching the tracking identifier. The real-time feedback stream processing pipeline is used to continuously consume a positive and negative feedback event stream after attribution. The model and strategy optimizer is connected with the real-time feedback stream processing pipeline and the real-time intelligent decision center. The model and strategy optimizer is configured to update user dynamic preference features in a real-time feature library in an incremental or small batch manner based on the feedback event stream, and periodically or triggeredly start a retraining process to update model parameters in an online prediction model by using accumulated new feedback data, thereby completing a self-evolution closed loop of the system.

10. The cross-channel marketing automation system of claim 9, wherein: The touch channel types supported by the multi-channel adaptation and execution layer include a mobile push service, an SMS gateway, an email server, a WebSocket in-site message service and an instant messaging application interface.

Images