A multi-channel marketing touch opportunity decision-making method and system based on multi-source behavior data

Abstract

This invention discloses a method and system for decision-making on multi-channel marketing outreach timing based on multi-source behavioral data, aiming to solve the problems of fragmented marketing outreach data and blind decision-making in existing systems. The method includes: collecting and aligning multi-source heterogeneous data with entities using the OneID system; constructing multimodal features including a time decay factor; identifying user dynamic intent using the Transformer architecture; jointly deciding on the optimal outreach channel, timing, and frequency threshold by combining Weibull distribution and deep reinforcement learning; and triggering model self-evolution through an asynchronous feedback mechanism. Through the above scheme, this application achieves fine-grained control of outreach timing, maximizing conversion value and reducing user fatigue under resource constraints.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method and system for decision-making on multi-channel marketing outreach timing based on multi-source behavioral data. Background Technology

[0002] With the deep integration of information technology and the continuous expansion of the mobile internet ecosystem, the interaction interface between enterprises and users has evolved from a single offline or PC-based model to a multi-dimensional, all-scenario reach matrix encompassing mobile applications (APPs), social media, SMS services, email, and various smart terminals. Against this backdrop, multi-channel collaborative marketing has become an increasingly important core means for enterprises to improve user activity, enhance brand loyalty, and optimize resource allocation. Early marketing outreach solutions primarily relied on basic Customer Relationship Management (CRM) systems, using static analysis of user profiles and historical purchase records to set pre-defined trigger rules or timed push strategies based on human experience. In the early stages of technological development, this model played a positive role in handling data environments with limited dimensions and relatively isolated channels, achieving a certain degree of transition from "broadcasting" to "preliminary segmentation."

[0003] However, with the continuous development of related technologies and the increasingly stringent requirements for reach accuracy and user experience in application scenarios, the inherent characteristics of traditional marketing decision-making mechanisms at the principle level are gradually revealing deep limitations when facing the current complex multi-source heterogeneous data environment. Specifically, existing technical solutions are often limited by the data silo effect, and behavioral data from different channels cannot achieve true spatiotemporal alignment and semantic fusion at the underlying logic level. When users' browsing, searching, collecting, and conversion behaviors on different platforms and terminals are segmented in independent storage and analysis architectures, the system struggles to construct a continuous and complete dynamic intent evolution curve. Furthermore, due to the lack of in-depth correlation analysis of multi-source behavioral data, existing triggering mechanisms mostly exhibit a "passive response" or "mechanical cycle" characteristic.

[0004] In traditional decision-making logic, in pursuit of short-term conversion metrics, systems often tend to increase reach frequency or adopt a "bombardment" strategy of simultaneous multi-channel push notifications. However, this approach, which ignores time-based adaptability, channel synergy, and user fatigue, leads to severe negative feedback effects at the technical level. On the one hand, due to the inability to scientifically define the optimal timing for outreach, a large amount of marketing resources are invested in inactive or high-load periods, resulting in a significant waste of computing resources and marketing budgets. On the other hand, the lack of cross-channel frequency control means that users receive repetitive and redundant information in a short period, easily triggering psychological resistance and leading to extremely high complaint and churn rates. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for decision-making on multi-channel marketing outreach based on multi-source behavioral data, thereby solving the aforementioned technical problems in the prior art.

[0006] The objective of this invention can be achieved through the following technical solutions: A method and system for decision-making on multi-channel marketing outreach based on multi-source behavioral data includes the following steps: S101. Real-time acquisition and standardized preprocessing of multi-source heterogeneous behavioral data: A data acquisition engine is used to acquire multi-source behavioral data, including data from mobile applications, web pages, social media mini-programs, and offline sensing devices. This multi-source behavioral data includes atomic-level behavioral event streams. Each behavioral event stream contains a unique user identifier, device fingerprint information, behavior trigger timestamp, behavior type code, page path, dwell time, and business-related parameters. In the preprocessing stage, the OneID system, based on a combination of deterministic logic and probabilistic mapping, is used to merge temporary identifiers scattered across different channels into a globally unified user entity, establishing a user identity mapping table. The acquired raw data is cleaned to remove invalid abnormal noise and crawler traffic. A sliding window-based time alignment algorithm is used to discretize the cross-channel behavioral sequences on a unified time axis, generating a normalized multi-dimensional behavioral tensor. S102. Construction of Multimodal Feature Engineering Based on Deep Intent Perception: Based on the multidimensional behavior tensor, a feature extraction procedure is executed to extract static profile features, dynamic temporal features, and contextual environment features. The static profile features cover the user's sociodemographic attributes and long-term preference index. The dynamic temporal features are achieved by calculating the user's behavior frequency, conversion path depth, and interest decay coefficient within a specific period. A behavior weight calculation formula based on time-exponential decay is introduced. ; in, Indicates user At the current time point For the action item Cumulative interest score, Let be the initial weighting coefficient for the j-th action, and λ be the preset time decay factor. This is a historical timestamp of the event. S103. Construct a user dynamic intent recognition model based on the Transformer architecture: Input the processed multimodal features into a pre-trained intent recognition network. The intent recognition network uses a multi-head self-attention mechanism to capture long-distance dependencies in the user behavior sequence. The weight distribution is calculated using the query vector Q, key vector K, and value vector V, and its mathematical expression is as follows: ; Based on this, the model outputs the predicted conversion probability value and potential intent category label of the user in the future preset step size, and integrates an auxiliary branch based on the Long Short-Term Memory Network (LSTM) to extract the user's impulsive consumption characteristics in the short term, and then weights and fuses the outputs of the two. S104. Multi-channel reach timing decision based on survival analysis and reinforcement learning: A reach timing prediction model is constructed using the Weibull distribution to find the moment with the highest user response probability. The risk function h(t) for reach timing is defined as: ; in, The shape parameter determines the trend of the response probability over time. The scaling parameter is used; simultaneously, a strategy iteration algorithm based on Deep Q-Network (DQN) is adopted, defining the state space S as the user's real-time feature vector, the action space A as the combination of different channels and intensities of outreach, and the reward function R as the weighted difference between conversion contribution, user fatigue penalty, and outreach cost. R=α Conversion-β Fatigue-δ Channe; By maximizing the cumulative expected return, the optimal reach channel, optimal reach time, and recommended reach frequency threshold are output for the current state. S105. Closed-loop feedback mechanism and model self-evolution training: Establish a real-time feedback acquisition module to capture the user's immediate behavioral feedback after receiving the touch information. The feedback data is written to the distributed storage cluster through an asynchronous message queue. Start the model parameter automatic correction program and introduce the feedback results as incremental samples into the online learning logic. When the prediction accuracy or conversion rate is lower than the preset performance baseline, trigger the automatic tuning mechanism and use the stochastic gradient descent algorithm SGD to update the parameters of the intent recognition model and reinforcement learning policy network.

[0007] Furthermore, the OneID system construction in S101 specifically involves a graph database-based association discovery algorithm: treating device fingerprints, mobile phone numbers, login accounts, and cookie information as graph nodes, and treating their co-occurrence relationships in the same login time space as edges; By executing connected component identification and maximal clique discovery algorithms, the set of nodes with a probability correlation degree exceeding 0.95 is clustered into a single natural person ID, and a globally unique UID mapping table is generated, ensuring that massive amounts of data stored in the distributed database can complete entity alignment within milliseconds.

[0008] Furthermore, the feature engineering in S102 also includes a dimensionality reduction processing logic based on an autoencoder: for sparse behavioral features of thousands of dimensions, an encoder is constructed using a multilayer perceptron to map the high-dimensional feature space to a low-dimensional dense vector space, reducing the computational overhead of subsequent models while retaining more than 95% of the information entropy; the encoding process satisfies the following objective of minimizing reconstruction error: ; in, For encoding functions, The decoding function is trained using the backpropagation algorithm to obtain more representative hidden layer features.

[0009] Furthermore, in S103, the intent recognition model also includes a feature cross-network based on an attention mechanism. This feature cross-network explicitly captures the nonlinear interactions between features through multiple layers of cross-operations, with each layer of cross-operation satisfying the following recursive relationship: ; in, The original input features, For the first Layer cross features, and These are the learning parameters for this layer, used to mine higher-order correlation information between specific weather, specific geographical location, and specific channel preferences.

[0010] Furthermore, in S104, based on the Weibull distribution, a non-homogeneous Poisson process NHPP is introduced to model the temporal characteristics of user behavior; The instantaneous rate at which users produce specific behaviors in different time segments is characterized by estimating the intensity function λ(t) of the non-homogeneous Poisson process. The parameter estimation of the intensity function adopts the maximum likelihood estimation method. By fitting historical behavior time point data, the unique active time window of each user is calculated, thereby ensuring that the trigger command is issued within the physiological and behavioral inertia cycle when the user's attention is most concentrated.

[0011] Furthermore, the fatigue penalty term FF in S104 adopts a nonlinear cumulative model: the system maintains a scrolling reach counter for each UID, and calculates the fatigue score based on the historical reach count and time interval. The specific calculation formula is as follows: ; in, The interference intensity weight for the kth reach is... This represents the time span from the current moment to the point of arrival. For smoothing purposes, when the fatigue score exceeds the preset dynamic safety threshold, the system automatically blocks all high-interference channels from reaching the target until the score decays to a safe range.

[0012] Furthermore, S105 introduces a resource quota optimization strategy based on backpropagation: When multiple marketing campaigns compete for the same high-value channel quota during the same time period, the system ranks them based on the Expected Conversion Value (ECV) of each campaign. The formula for calculating ECV is as follows: ; in, The conversion probability predicted by the model. Converting value into units To minimize reach costs, the system prioritizes outreach tasks with higher ECV values, thereby maximizing overall business growth within the constraints of limited marketing resources.

[0013] Furthermore, the system is functionally divided into the following core hardware and software modules: Data access and integration module: includes a multi-protocol access gateway to support MQTT, HTTP, and gRPC communication protocols and receive encrypted data streams from the global acquisition terminal. The data access and integration module is equipped with a high-speed buffer and a Kafka-based distributed message bus. Full-domain feature computing center: Deployed on the computing cluster, it consists of an offline feature mining unit and an online real-time feature computing unit. The online real-time feature computing unit uses a stream computing framework to aggregate the event stream entering the bus in real time and calculate the statistical indicators within the sliding window. The offline feature mining unit periodically performs in-depth analysis on historical data and updates the user's long-term interest distribution map. Intelligent Decision Brain: As the core hub of the system, it integrates an intent recognition inference engine, a timing probability evaluator, and a reinforcement learning controller. The intelligent decision brain runs on a high-performance GPU computing unit, performs real-time inference on global features through a pre-loaded deep learning model weight file, receives business rule instructions through the northbound interface, and issues decision instruction sets through the southbound interface. Multi-channel collaborative execution module: It is responsible for receiving the instruction set generated by the decision-making brain and converting it into executable tasks for specific channels. It has a task priority scheduling algorithm that queues and issues out tasks based on the cost and quota of marketing resources and the status of the channel. It also has the functions of abnormal retry and automatic channel switching. Effect tracking and self-evolution module: Real-time monitoring of the reach effect across the entire chain, obtaining conversion closed-loop data through the tracking point feedback mechanism, with an internal online evaluation unit that dynamically evaluates the decision-making effectiveness of the current model through calculated indicators, and triggers incremental training or parameter fine-tuning of the model based on the evaluation results.

[0014] Furthermore, the system also includes a privacy protection and security computing module: this module is located between the data access and integration module and the global feature computing center. It introduces a differential privacy algorithm to inject controlled noise into the feature vectors; the injection strength of the differential privacy is controlled by the parameter ε. Control, satisfying the following constraints: ; in, and The datasets are adjacent datasets that differ by only one sample; the privacy protection and security computing module supports encrypted computation based on a hardware trusted execution environment (TEE) to ensure that the model inference process is executed in isolated secure memory.

[0015] Furthermore, the effect tracking and self-evolution module has an automatic A / B testing function: the system dynamically divides the user group into an experimental group and a control group. The experimental group applies optimal decision logic based on reinforcement learning, while the control group applies traditional logic based on fixed rules or random strategies. By comparing the performance of the two groups in terms of conversion rate, click-through rate, unsubscription rate, and computational resource consumption rate, the effect of the technology is quantified. When the experimental group performs significantly better than the control group and the confidence level p < 0.05, the system automatically promotes the optimal model parameters to the production environment. The system also constructs a unified semantic layer for behavior based on a distributed computing engine, which maps heterogeneous behavioral instructions from different platforms to a unified behavioral classification system, and transforms heterogeneous behaviors into interactive semantic tags for marketing content, thereby eliminating strategy bias caused by data silos.

[0016] The beneficial effects of this invention are: 1. This invention introduces a non-homogeneous Poisson process to model the temporal characteristics of user behavior, based on the Weibull distribution. The system characterizes the instantaneous rate at which users generate specific behaviors in different time segments by estimating the intensity function λ(t) of the non-homogeneous Poisson process. The parameter estimation of the intensity function adopts the maximum likelihood estimation method. By fitting historical behavior time point data, a unique "active time window" for each user is calculated, thereby ensuring that the trigger command is issued within the physiological and behavioral inertia cycle when the user's attention is most concentrated.

[0017] 2. This system constructs a "unified semantic layer for behavior" based on a distributed computing engine. In this layer, behavioral instructions from different platforms are mapped to a unified behavioral classification system. Through this semantic normalization, the decision engine can measure the depth of user-brand interaction in a continuous, multi-dimensional space, completely eliminating strategy bias caused by data silos.

[0018] 3. This invention integrates advanced technologies such as OneID entity alignment, deep feature cross-networks, Transformer intent recognition, reinforcement learning timing decision-making, and closed-loop incremental learning to construct a closed-loop system from perception and recognition to intelligent decision-making, and then to collaborative execution and adaptive optimization. Compared with traditional technologies, this invention not only overcomes the limitations of human experience at the principle level, achieving fine-grained control of the timing and frequency of outreach, but also solves the technical challenge of difficult collaboration of multi-source heterogeneous data at the system level, significantly improving marketing conversion efficiency, reducing user complaint rates, and possessing extremely high engineering application value and significant non-obviousness. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0020] Figure 1 A flowchart illustrating a multi-channel marketing outreach timing decision-making method based on multi-source behavioral data, provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of a multi-channel marketing outreach timing decision system based on multi-source behavioral data, provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] like Figures 1-2 As shown, the embodiments of the present invention, based on the original specific embodiments, supplement the key technical features such as OneID system construction, feature cross-network operation, non-homogeneous Poisson process modeling, fatigue accumulation calculation, resource quota optimization, privacy protection implementation, automatic A / B testing, and behavior unified semantic layer with refined implementation details, and improve the engineering implementation parameters, algorithm execution logic, and module collaboration mechanism of each module.

[0023] All engineering implementations in this embodiment are based on a cloud-native architecture, using Kubernetes containerized deployment to support elastic scaling; data storage adopts a hybrid strategy of hot data (Redis cluster) + warm data (distributed NoSQL) + cold data (HDFS) to meet the needs of high-concurrency and high-real-time marketing scenarios; the entire link uses TLS1.3 encrypted transmission, and sensitive data undergoes irreversible hashing processing to comply with general data security regulations.

[0024] A multi-channel marketing outreach timing decision-making method based on multi-source behavioral data includes the following steps: S101. Real-time acquisition and standardized preprocessing of multi-source heterogeneous behavioral data: It enables the collection of multi-source data across the entire domain, unified OneID identity, data cleaning and denoising, time alignment and tensor generation, providing a standardized and highly available data foundation for subsequent feature engineering and model inference.

[0025] Multi-source behavioral data collection utilizes a lightweight event tracking engine deployed across all touchpoints, including mobile apps, web pages (Web / H5), social media mini-programs, and offline sensing devices (Bluetooth beacons, POS terminals, etc.). It collects atomic-level behavioral event streams, each encapsulated in a Protocol Buffers format containing seven core fields: user temporary identifier, device fingerprint (OAID / AndroidID / IMEI hash), millisecond-level timestamp, behavior type encoding (covering over 100 predefined types such as browsing / clicking / adding to cart / conversion), page path, dwell time, and business-related parameters (product SKU / promotion ID / store number). The collection engine supports 100,000 QPS concurrent collection capabilities and features a built-in traffic shaping mechanism to meet the demands of high-traffic scenarios such as e-commerce promotions.

[0026] The OneID system is built and identity mapping is based on a graph database to construct an association discovery algorithm, achieving unified user identity across channels and terminals. Nodes with a probability association degree ≥ 0.95 are clustered into a single natural person ID. The specific process is as follows: Graph node and edge initialization: Set the device fingerprint, mobile phone number (SHA256 irreversible hash), login account, and cookie information as graph nodes. The node attributes include the identifier type, generation time, and associated device. Set the co-occurrence relationship of each identifier in the same login session (within 30 minutes), the same IP segment, and the same geographical location (within 100 meters) as directed edges. The edge weight is assigned according to the co-occurrence frequency (1 time = 0.2, maximum 1.0).

[0027] Node clustering operation: First, the Union-Find algorithm is used for initial clustering, and then the Bron-Kerbosch algorithm is used for optimization. Only the set of nodes with a probability correlation greater than 0.95 is clustered into a single natural person entity. When there is a clustering conflict, the strong real-name identification priority rule is followed, and the mobile phone number and real-name login account are regarded as core nodes, and other identifications are subordinate nodes.

[0028] UID Mapping and Retrieval: Generate a globally unique 64-bit numeric UID, construct a <temporary identifier-UID> mapping table, store it in a RedisCluster cluster, set a 24-hour expiration and refresh mechanism to ensure that entity alignment is completed in milliseconds (≤5ms), and support identity queries under high concurrency.

[0029] Data cleaning and abnormal traffic filtering employ a two-layer cleaning mechanism of rule filtering + isolated forest algorithm to ensure data effectiveness ≥99%; rule filtering: removes requests from crawler IP ranges, empty requests without behavioral traces, and abnormal behaviors repeated more than 10 times within 1 minute; isolated forest algorithm: performs anomaly detection on the data after rule cleaning and removes noisy data that deviates from the normal behavior distribution (such as high-frequency clicks that are not normal user operations).

[0030] Time alignment and multidimensional behavior tensor generation adopt a sliding window-based time alignment algorithm, setting a sliding step size of 1 minute and an observation window of 72 hours to map discrete behavior sequences across channels to a unified time axis, and quantizing and encoding the behavior features of each time step; finally, a normalized multidimensional behavior tensor is generated, with a tensor shape of [N,T,F] (N=number of users, T=number of time steps, F=feature dimension), a data type of Float32, and normalized to the [0,1] interval to eliminate the influence of dimensions.

[0031] S102. Construction of multimodal feature engineering based on deep intent perception: Based on multidimensional behavior tensors, static, dynamic, and contextual multimodal features are extracted. Time decay weights are used to characterize changes in user interests. High-dimensional feature dimensionality reduction is achieved based on autoencoders, ensuring that more than 95% of the information entropy is retained while reducing the computational cost of the model.

[0032] Multimodal feature extraction initially extracts 1500-dimensional sparse features, divided into three categories, covering all dimensions of user attributes and behaviors, including: Static profile features (200 dimensions): include sociodemographic attributes (gender, age group, city of residence, occupation type) and long-term preference index (category preference, spending power, purchase frequency, and average order value based on 180 days of historical behavior), updated weekly via offline Spark tasks.

[0033] Dynamic temporal features (1200 dimensions): Calculates the frequency of behaviors, transformation path depth, and interest decay coefficient within the observation window. The core formula for calculating the weights of behaviors with exponentially decaying execution time is as follows: ; in, λ represents the initial weight for the j-th action (conversion / payment = 1.0, add to cart / favorite = 0.6, view / click = 0.2, swipe / exit = 0.1); λ is the time decay factor, configured differently according to business category (FMCG = 0.05 / h, durable goods = 0.01 / h, services = 0.03 / h). This is the current timestamp. This serves as a timestamp for historical actions.

[0034] Contextual features (100 dimensions): Real-time collection of current time (time period / weekday / holiday), weather (sunny / rainy / temperature and humidity), device status (remaining battery power, network type), and geographical location (home / work / commuting / outdoor), updated in real time through Flink streaming tasks, with an update latency of ≤100ms.

[0035] For 1500-dimensional sparse features, a multilayer perceptron (MLP) autoencoder is constructed to map them to 128-dimensional low-dimensional dense features, minimizing the reconstruction error objective. ; in, For encoding functions, The decoding function, and its specific implementation details include: Network structure: The encoder is “input layer (1500-dimensional) → hidden layer 1 (512-dimensional, ReLU) → hidden layer 2 (256-dimensional, ReLU) → coding layer (128-dimensional, linear)”; the decoder is “coding layer (128-dimensional) → hidden layer 1 (256-dimensional, ReLU) → hidden layer 2 (512-dimensional, ReLU) → output layer (1500-dimensional)”, which is symmetrical to the encoder.

[0036] Training parameters: Adam optimizer is used, learning rate is 1e-3, batch size is 256, and number of training epochs is 50; the mean squared error of MSE is used as the loss function, and an early stopping mechanism is introduced (if the validation set loss does not decrease for 5 consecutive epochs, the training stops) to prevent overfitting.

[0037] Dimensionality reduction effect: retains ≥95% of the information entropy of the original data, and improves the efficiency of subsequent model inference by ≥60%.

[0038] S103. Construction and Inference of User Dynamic Intent Recognition Model Based on Transformer Architecture: We construct a fusion model consisting of a Transformer main model, a Bi-LSTM auxiliary branch, and a feature cross-network to achieve dynamic user intent recognition and conversion probability prediction. This model accurately captures the long-distance dependence and impulsive consumption characteristics of user behavior and uncovers high-order nonlinear correlations between features.

[0039] The Transformer main model is constructed using a 6-layer Transformer encoder, with 128-dimensional low-dimensional dense features as input. The core execution mechanism is a multi-head self-attention mechanism. ; Model details: 8 multi-head self-attention heads, 512 hidden layer dimensions, 2048 feedforward network dimensions, GELU activation function; layer normalization + residual connections are used, and the dropout rate is 0.2 to prevent overfitting; it can capture long-distance dependencies of user behavior over a period of up to 90 days (such as the correlation between browsing at the beginning of the month and purchasing at the end of the month).

[0040] The Bi-LSTM auxiliary branch integrates a two-layer bidirectional LSTM as an auxiliary branch. The input is the intermediate features of the Transformer main model, with a hidden layer dimension of 128 and a dropout rate of 0.2. It specifically extracts instantaneous behavioral features of users within the past 15 minutes, capturing impulsive consumption signals (such as multiple purchases to the cart within a short period or rapid browsing of promotional pages). The outputs of the main model and the auxiliary branch are fused through a weighted fusion layer. The fusion weights are dynamically adjusted according to the marketing scenario: 0.4 for impulsive features during major promotional events and 0.2 for everyday scenarios.

[0041] The feature cross-network operation embeds three layers of feature cross-networks before the fusion layer to explicitly mine nonlinear high-order correlations between features, adapting to the complex feature interactions in marketing scenarios: ; in, The original input features, For the first Layer cross features, and The learning parameters for this layer are used to mine high-order correlation information between specific weather, specific geographical location, and specific channel preferences. Specific implementation includes: Parameter initialization: Initialize using Xavier. Initialize to 0, and share the learning rate with the main model (1e-3). Advanced association mining: Layer 1 mines binary associations (such as 11 PM + Wi-Fi connection), Layer 2 mines ternary associations (such as beauty preferences + 11 PM + Wi-Fi connection), and Layer 3 mines quaternary and higher associations (such as dark mode + beauty preferences + 11 PM + Wi-Fi connection). Feature selection: The cross-referenced features are selected by attention weighting, and only high-order correlation features with attention weight ≥ 0.1 are retained to reduce redundant interference, and the output dimension is kept at 128 dimensions.

[0042] The model output and inference fusion model achieves dual-task inference through two output layers, deployed on the TensorRT inference engine, supporting 100,000-level QPS concurrency and inference latency ≤500ms. Softmax classification layer: Outputs category labels for users' potential intentions (high-potential purchase / comparative research / information acquisition / churn warning / no clear intention); Sigmoid regression layer: Outputs the predicted conversion probability P for a preset step size (2 / 6 / 24 hours, configurable). conv .

[0043] S104. Multi-channel reach timing decision-making based on survival analysis and reinforcement learning: As the core decision-making layer of the system, it uses a combination of Weibull distribution and non-homogeneous Poisson process (NHPP) to predict the timing of delivery, and uses deep Q network (DQN) to make joint channel / frequency decisions. It also introduces a nonlinear fatigue model to avoid user interference.

[0044] By jointly modeling user response time distribution and behavioral temporal characteristics based on Weibull distribution and NHPP, the timing prediction of reach can accurately locate the active time window where user attention is most concentrated, with a timing deviation of ≤12 minutes. Weibull distribution parameter fitting: The maximum likelihood estimation (MLE) method is used to fit the user's historical response time data, and the core follows the risk function formula: ; Wherein, γ is the shape parameter (response probability increases when γ>1, decreases when γ<1), and η is the scale parameter; initialize γ=1.2 and η=12 hours, and fine-tune once per hour through online feedback to locate the peak moment of response probability.

[0045] NHPP time-domain modeling: Introducing a power-law intensity function: λ(t)=λ0 t β-1 (λ0 is the initial intensity, β is the trend parameter), by fitting the user's historical behavior time point data through MLE, calculate the personalized active time window for each user (such as 7:00-8:00 and 19:00-22:00 for office workers) to ensure that the trigger command is issued within the user's physiological / behavioral inertia cycle.

[0046] The DQN-based joint channel and frequency decision-making employs a Deep Q-Network (DQN) strategy iterative algorithm, defining a state space, action space, and reward function. By maximizing the cumulative expected return, it outputs the optimal reach channel, time, and frequency thresholds. Specific implementation details are as follows: State space definition: State space S = 128-dimensional real-time user features + conversion probability P conv + Fatigue score + Remaining channel quota; Action space A = 8 types of channels (SMS / APP Push / WeChat / Internal message / Email / Douyin / Kuaishou / Offline SMS) + 3 types of intensity (single / two intervals / no contact), for a total of 24 discrete combinations.

[0047] Reward function configuration: R=α Conversion-β Fatigue-δ Channelt; Wherein, Conversion is the conversion tag (conversion=1, no conversion=0), Fatigue is the fatigue score, and Channel is the cost of reaching. The weighting coefficients are configured according to marketing objectives: conversion priority (α=0.6, β=0.2, δ=0.2), user experience priority (α=0.4, β=0.4, δ=0.2).

[0048] Training parameters: DQN uses an empirical replay pool (size = 1e6) + a target network, a learning rate of 1e-4, a discount factor of 0.95, and the target network is updated every 5000 steps; - Greedy strategy ( (It decays linearly from 0.9 to 0.1), and then continues to learn online after 1e5 rounds of pre-training.

[0049] User fatigue calculation and dynamic threshold control: Maintain a scrolling reach counter for each UID, calculate a non-linear fatigue score in real time, and avoid excessive reach that may cause user aversion. ; in: The interference intensity weight for the k-th contact is (SMS / phone = 1.0, APP Push / WeChat = 0.5, in-app message / email = 0.2). To reach the current time span (hours); =0.1 is a smoothing term to avoid a denominator of 0.

[0050] Dynamic threshold control: Set the basic security threshold to 5.0, and adjust according to user value (high value = 6.0, normal = 5.0, low value = 4.0); when the score exceeds the threshold, all scores will be automatically hidden. High interference channels with a value ≥0.5 will naturally decay to a safe range at a rate of 0.1 / hour.

[0051] The decision output combines the timing prediction results of Weibull distribution + NHPP with the channel / frequency decision results of DQN, generating an optimal decision instruction set in JSON format, including (user UID, optimal reach channel, precise reach timestamp, maximum reach frequency, marketing content ID), where the timestamp is accurate to the minute, and is sent through the southbound interface with a delay of ≤200ms.

[0052] S105, Closed-loop feedback mechanism and model self-evolution training: It enables real-time feedback collection, incremental model updates, and resource quota optimization, building a self-evolving closed loop to ensure that the model dynamically optimizes itself in response to changes in user behavior and the market environment.

[0053] Real-time feedback data collection and storage establishes a full-link tracking feedback mechanism to capture all user behavior feedback after being reached (view / click / add to cart / conversion / unsubscribe / complaint / block). Feedback data includes <decision instruction ID, user UID, reach time, feedback behavior, feedback time>, and is written to the Cassandra database through a Kafka asynchronous message queue (number of partitions = 32, number of replicas = 3). The collection latency is ≤500ms, ensuring no data loss and high reliability of transmission.

[0054] The online learning program automatically corrects and evolves model parameters, using feedback data as incremental samples for model training. It employs stochastic gradient descent (SGD) to update the intent recognition model and reinforcement learning policy network parameters. Specific triggering and training rules are as follows: Triggering conditions: When the model prediction accuracy (AUC) is less than 0.85 for 2 consecutive hours, or the actual conversion rate is less than 30% of the predicted value for 4 consecutive hours, the optimization mechanism will be automatically triggered. Incremental training: SGD learning rate 5e-5, batch size 128, number of samples for incremental training 1e4; only the top-level weights and biases of the model are updated, without retraining the entire dataset, training time ≤10 minutes, the updated model is seamlessly deployed online without service interruption.

[0055] Resource quota optimization based on ECV addresses the quota competition problem of high-value channels (SMS / Douyin / Kuaishou). It prioritizes high-ECV tasks based on bidding ranking according to Expected Conversion Value (ECV), maximizing overall business growth with limited resources. The ECV calculation formula is as follows: ECV=P conv ×V unit -C touch ; Among them, V unit Value per unit of conversion (high-value user = 100, average = 50, low-value = 20), C touch The cost per reach per channel is RMB 0.1 / message, RMB 0.01 / app push, and RMB 0.02 / WeChat. Quota allocation rules: tasks are sorted by ECV from highest to lowest, with a minimum quota granularity of 100 tasks; tasks with an ECV difference ≤ 5 are allocated using a round-robin system; when quotas are exhausted, tasks are automatically downgraded to the next best channel to ensure channel quota utilization ≥ 94%.

[0056] The multi-channel marketing outreach timing decision system of this invention adopts a microservice architecture and is divided into six core modules. Each module communicates through a gRPC interface with an interface response time of ≤100ms and supports cross-data center disaster recovery backup.

[0057] The data access and integration module serves as the system's data entry point, enabling encrypted access, caching, and transfer of multi-source data, including: Multi-protocol access gateway: Supports MQTT, HTTP, and gRPC protocols, connects to the full-domain data collection terminal, and uses TLS 1.3 to encrypt and transmit data streams; High-speed buffer: Local memory + Redis two-level cache, cache capacity 10GB, expiration time 5 minutes, to cope with traffic peaks; Distributed message bus: Kafka topic partitions = 64, replicas = 3, message retention for 72 hours, ensuring efficient and reliable data flow.

[0058] The privacy protection and secure computing module, located between the data access and integration module and the global feature computing center, is the core of privacy and security, enabling sensitive data protection and encrypted computation. Differential privacy implementation: Gaussian noise is injected into sensitive user features (phone number / geographic location / spending power), with a noise mean of 0 and a variance σ. 2 =( 2 / 2ln(1.25 / δ))×Δ 2 ( =1.0, δ=1e-5, Δ=0.1). Differential privacy constraints: ; Non-sensitive features are free from noise, balancing privacy protection and feature effectiveness.

[0059] TEE Encryption Operation: Based on the Intel SGX Trusted Execution Environment, the core model inference and sensitive feature calculation are deployed in 16GB of isolated secure memory, accessible only to authorized processes, to prevent the theft / tampering of algorithms and data; Data anonymization: Sensitive information such as mobile phone numbers and ID card numbers are processed using SHA256 + random salt value irreversible hashing. The salt value is stored in an independent key management system (KMS) to ensure that it cannot be restored from the source.

[0060] The global feature computation center is deployed on a Flink+Spark computing cluster to realize real-time feature computation and offline mining, and includes the following units: Online feature real-time computation unit: Based on Flink event time semantics, with a sliding window of 1 minute, it calculates user dynamic time-series features and contextual features in real time, with an update latency of ≤100ms; Offline feature mining unit: Based on Spark batch processing, features are fully updated every day at midnight, and the long-term interest distribution map of users is updated weekly, supporting a data processing capacity of ≥10PB / day.

[0061] The core hub of the intelligent decision-making brain system is deployed on a high-performance GPU computing cluster. It integrates an intent recognition inference engine, a timing probability estimator, and a reinforcement learning controller. The three engines perform pipelined parallel inference, with an overall inference latency of ≤500ms. Two interfaces are used: the northbound interface receives marketing target instructions from the business side (such as improving conversion / activating dormant users) and supports dynamic parameter configuration; the southbound interface publishes decision instruction sets using a publish-subscribe model and has a built-in instruction resending mechanism (resending 3 times with an interval of 1 second) to ensure that instructions are delivered.

[0062] The multi-channel collaborative execution module is responsible for the implementation of decision-making instructions, achieving multi-channel collaborative outreach, and includes: Task scheduler: Based on ECV value and time priority sorting, it supports scheduled execution of trigger commands (accurate to the minute). Channel adapter: Transforms unified decision-making instructions into a format specific to each channel (SMS = text, APPPush = JSON, WeChat = image and text). Anomaly Handling Unit: Equipped with automatic channel switching and anomaly retry function, it automatically switches to the backup channel when a channel fails, and retryes 3 times (with a 5-second interval) to ensure a task success rate of ≥99%.

[0063] The effect tracking and self-evolution module enables end-to-end effect monitoring, model self-evolution, automated A / B testing, and the construction of a unified semantic layer for behavior, and is the core of the system's self-evolution. Online evaluation unit: Real-time calculation of core indicators such as AUC, conversion increment, ROI, and channel quota utilization rate, displayed in a visual dashboard, and triggering SMS / email alerts when indicators fall below the threshold; Automated A / B testing: Stratified random sampling is used to group users according to value (high / medium / low), and within each group, users are randomly divided into experimental and control groups (fixed rules / random strategy), with a sample size of ≥500,000 in each group; a two-sample t-test is used for significance analysis. When the confidence level p<0.05 and the conversion rate of the experimental group increases by ≥20%, the optimal model parameters are released in a phased manner (10%→50%→100%, with an interval of 1 hour between each stage). Unified Semantic Layer for Behavior: Based on FlinkSQL, a distributed semantic mapping engine is built to formulate a unified classification system for behavior across the entire domain, consisting of 5 major categories and 28 subcategories. For example, heterogeneous behaviors are mapped to unified semantic tags (such as clicking on a banner in the APP / clicking on an article in WeChat → "Marketing Content Interaction", and adding to favorites in the Web / adding to cart in the APP → "Deep Interest Interaction"). The tags are accompanied by a weight value of 0.1-1.0, enabling quantitative comparison of behaviors across channels and completely eliminating the strategy bias caused by data silos.

[0064] Table 1 shows the results of applying this system to a large-scale retail e-commerce platform's overall user activation scenario. 10 million existing active users were selected and randomly divided into experimental and control groups according to the A / B testing rules mentioned above. The experiment ran for 30 days, and the comparison results of the core indicators are as follows, fully verifying the technical effectiveness and engineering application value of this invention.

[0065] Table 1 ; Analysis of the data in Table 1 shows that after applying the technical solution of this invention, the conversion rate increased from 1.45% to 2.68%, and the click-through rate doubled. More significantly, the user unsubscription rate and fatigue score decreased substantially, demonstrating that the decision-making mechanism based on reinforcement learning and fatigue penalty can accurately identify the "non-interference zone" for reaching users. Regarding system performance, although the model complexity increased, thanks to the microservice architecture and GPU-accelerated inference, the decision latency was controlled within 500 milliseconds, fully meeting the real-time requirements of marketing scenarios.

[0066] To address the technical bottleneck of fragmented cross-channel data, this system achieves deep integration at the logical level by constructing a "unified semantic layer for behavior." In this semantic layer, the system maps atomic behaviors from different terminals to a unified weighted coordinate system. For example, for the physical operations of "adding a product to favorites on the web" and "viewing product parameters on the app," the semantic layer uniformly quantifies them as "deep interest interaction" with a value of 0.6 units. This normalization process allows the decision engine to transcend specific channel limitations and evaluate the strength of user intent within a continuous feature space, thereby eliminating strategy bias caused by data silos.

[0067] Regarding model self-evolution, the closed-loop feedback mechanism employed in this invention ensures the system possesses strong environmental adaptability. The online evaluation unit continuously calculates the conversion increment and utilizes an A / B testing framework to perform canary releases of model updates. When a newly trained reinforcement learning strategy demonstrates significant superiority in the experimental group (confidence p < 0.05), the system automatically performs a one-click switch of model weights. This automated iterative mechanism greatly reduces the workload of manual tuning, enabling the system to maintain optimal decision output even when the market environment and promotional pace undergo drastic changes.

[0068] In terms of underlying architecture, this system fully adopts a cloud-native architecture. The data computing layer is deployed in containers using Kubernetes, supporting automatic scaling up and down based on traffic peaks. Data storage employs a hybrid storage strategy: hot data (such as real-time user behavior tensors) is stored in a Redis cluster consisting of 64 nodes, supporting sub-millisecond reads; warm data (such as feature vectors from the past 7 days) is stored in a distributed NoSQL database (such as Cassandra); and cold data (such as historical conversion feedback) is stored in a big data platform based on HDFS to support parameter reestimation for complex offline association analysis and survival analysis models.

[0069] In summary, the multi-channel marketing outreach timing decision-making method and system based on multi-source behavioral data provided by this invention systematically solves the core pain points of blind outreach timing, channel conflict, and poor user experience in the marketing field by introducing OneID identity alignment, multi-head self-attention intent recognition, non-homogeneous Poisson process (NHPP) temporal modeling, and a deep reinforcement learning strategy with fatigue penalty. This invention not only has rigorous mathematical modeling support but also ensures the feasibility and robustness of the solution through high-performance computing components, adaptive closed-loop mechanisms, and privacy and security protection. Quantitative comparisons using experimental data demonstrate that this invention represents a significant technological advancement in improving marketing conversion efficiency, reducing operating costs, and optimizing customer lifetime value (LTV).

[0070] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method and system for decision-making on multi-channel marketing outreach based on multi-source behavioral data, characterized in that, Includes the following steps: S101. Real-time acquisition and standardized preprocessing of multi-source heterogeneous behavioral data: The multi-source behavioral data, including mobile applications, web pages, social platform mini-programs and offline sensing devices, is acquired through the tracking engine. The multi-source behavioral data includes atomic-level behavioral event streams. Each behavioral event stream contains a unique user identifier, device fingerprint information, behavior trigger timestamp, behavior type code, page path, dwell time and business-related parameters. In the preprocessing stage, the OneID system, which combines deterministic logic and probabilistic mapping, is used to merge temporary identifiers scattered across different channels into a globally unified user entity and establish a user identity mapping relationship table. The collected raw data is cleaned to remove invalid abnormal noise and crawler traffic. A time alignment algorithm based on a sliding window is used to discretize the cross-channel behavior sequence on a unified time axis to generate a normalized multidimensional behavior tensor. S102. Construction of multimodal feature engineering based on deep intent perception: Based on the multidimensional behavior tensor, execute the feature extraction program to extract static portrait features, dynamic temporal features and contextual environment features; The static profile features encompass the user's sociodemographic attributes and long-term preference index. The dynamic temporal features are achieved by calculating the frequency of user behavior, conversion path depth, and interest decay coefficient within a specific period, and a behavior weight calculation formula based on time exponential decay is introduced: ； in, Indicates user At the current time point For the action item Cumulative interest score, Let be the initial weighting coefficient for the j-th action, and λ be the preset time decay factor. This is the current timestamp. This is a historical timestamp of the event. S103. Construct a user dynamic intent recognition model based on the Transformer architecture: Input the processed multimodal features into a pre-trained intent recognition network. The intent recognition network uses a multi-head self-attention mechanism to capture long-distance dependencies in the user behavior sequence. The weight distribution is calculated using the query vector Q, key vector K, and value vector V, and its mathematical expression is as follows: ； Based on this, the model outputs the predicted conversion probability value and potential intent category label of the user in the future preset step size, and integrates an auxiliary branch based on the Long Short-Term Memory Network (LSTM) to extract the user's impulsive consumption characteristics in the short term, and then weights and fuses the outputs of the two. S104. Multi-channel reach timing decision based on survival analysis and reinforcement learning: A reach timing prediction model is constructed using the Weibull distribution to find the moment with the highest user response probability. The risk function h(t) for reach timing is defined as: ； in, The shape parameter determines the trend of the response probability over time. The scaling parameter is used; simultaneously, a strategy iteration algorithm based on Deep Q-Network (DQN) is adopted, defining the state space S as the user's real-time feature vector, the action space A as the combination of different channels and intensities of outreach, and the reward function R as the weighted difference between conversion contribution, user fatigue penalty, and outreach cost. R=a Conversion-b Fatigue-d Channel; By maximizing the cumulative expected return, the optimal reach channel, optimal reach time, and recommended reach frequency threshold are output for the current state. S105. Closed-loop feedback mechanism and model self-evolution training: Establish a real-time feedback acquisition module to capture the user's immediate behavioral feedback after receiving the touch information. The feedback data is written to the distributed storage cluster through an asynchronous message queue. Start the model parameter automatic correction program and introduce the feedback results as incremental samples into the online learning logic. When the prediction accuracy or conversion rate is lower than the preset performance baseline, trigger the automatic tuning mechanism and use the stochastic gradient descent algorithm SGD to update the parameters of the intent recognition model and reinforcement learning policy network.

2. The multi-channel marketing outreach timing decision-making method based on multi-source behavioral data according to claim 1, characterized in that, The One ID system construction in S101 specifically involves a graph database-based association discovery algorithm: treating device fingerprints, mobile phone numbers, login accounts, and cookie information as graph nodes, and their co-occurrence relationships in the same login time space as edges; By executing connected component identification and maximal clique discovery algorithms, the set of nodes with a probability correlation degree exceeding 0.95 is clustered into a single natural person ID, and a globally unique UID mapping table is generated, ensuring that massive amounts of data stored in the distributed database can complete entity alignment within milliseconds.

3. The multi-channel marketing outreach timing decision-making method based on multi-source behavioral data according to claim 1, characterized in that, The feature engineering in S102 also includes a dimensionality reduction processing logic based on an autoencoder: for sparse behavioral features of thousands of dimensions, an encoder is constructed using a multilayer perceptron to map the high-dimensional feature space to a low-dimensional dense vector space, reducing the computational overhead of subsequent models while retaining more than 95% of the information entropy; the encoding process satisfies the following objective of minimizing reconstruction error: ； in, For encoding functions, The decoding function is trained using the backpropagation algorithm to obtain more representative hidden layer features.

4. The multi-channel marketing outreach timing decision-making method based on multi-source behavioral data according to claim 1, characterized in that, In S103, the intent recognition model further includes a feature cross-network based on an attention mechanism. This feature cross-network explicitly captures the nonlinear interactions between features through multiple layers of cross-operations, with each layer of cross-operation satisfying the following recursive relationship: ； in, The original input features, For the first Layer cross features, and These are the learning parameters for this layer, used to mine higher-order correlation information between specific weather, specific geographical location, and specific channel preferences.

5. The multi-channel marketing outreach timing decision-making method based on multi-source behavioral data according to claim 1, characterized in that, In S104, based on the Weibull distribution, a non-homogeneous Poisson process NHPP is also introduced to model the temporal characteristics of user behavior. The instantaneous rate at which users produce specific behaviors in different time segments is characterized by estimating the intensity function λ(t) of the non-homogeneous Poisson process. The parameter estimation of the intensity function adopts the maximum likelihood estimation method. By fitting historical behavior time point data, the unique active time window of each user is calculated, thereby ensuring that the trigger command is issued within the physiological and behavioral inertia cycle when the user's attention is most concentrated.

6. The multi-channel marketing outreach timing decision-making method based on multi-source behavioral data according to claim 1, characterized in that, The fatigue penalty term FF in S104 adopts a nonlinear cumulative model: the system maintains a scrolling reach counter for each UID, and calculates the fatigue score based on the historical reach count and time interval. The specific calculation formula is as follows: ； in, The interference intensity weight for the kth reach is... This represents the time span from the current moment to the point of arrival. For smoothing purposes, when the fatigue score exceeds the preset dynamic safety threshold, the system automatically blocks all high-interference channels from reaching the target until the score decays to a safe range.

7. The multi-channel marketing outreach timing decision-making method based on multi-source behavioral data according to claim 1, characterized in that, S105 introduces a resource quota optimization strategy based on backpropagation: When multiple marketing campaigns compete for the same high-value channel quota during the same time period, the system ranks them based on the Expected Conversion Value (ECV) of each campaign. The formula for calculating ECV is as follows: ； in, The conversion probability predicted by the model. Converting value into units To minimize reach costs, the system prioritizes outreach tasks with higher ECV values, thereby maximizing overall business growth within the constraints of limited marketing resources.

8. A multi-channel marketing outreach timing decision-making system based on multi-source behavioral data, characterized in that, The system is functionally divided into the following core hardware and software modules: Data access and integration module: includes a multi-protocol access gateway to support MQTT, HTTP, and gRPC communication protocols and receive encrypted data streams from the global acquisition terminal. The data access and integration module is equipped with a high-speed buffer and a Kafka-based distributed message bus. Full-domain feature computing center: Deployed on the computing cluster, it consists of an offline feature mining unit and an online real-time feature computing unit. The online real-time feature computing unit uses a stream computing framework to aggregate the event stream entering the bus in real time and calculate the statistical indicators within the sliding window. The offline feature mining unit periodically performs in-depth analysis on historical data and updates the user's long-term interest distribution map. Intelligent Decision Brain: As the core hub of the system, it integrates an intent recognition inference engine, a timing probability evaluator, and a reinforcement learning controller. The intelligent decision brain runs on a high-performance GPU computing unit, performs real-time inference on global features through a pre-loaded deep learning model weight file, receives business rule instructions through the northbound interface, and issues decision instruction sets through the southbound interface. Multi-channel collaborative execution module: It is responsible for receiving the instruction set generated by the decision-making brain and converting it into executable tasks for specific channels. It has a task priority scheduling algorithm that queues and issues out tasks based on the cost and quota of marketing resources and the status of the channel. It also has the functions of abnormal retry and automatic channel switching. Effect tracking and self-evolution module: Real-time monitoring of the reach effect across the entire chain, obtaining conversion closed-loop data through the tracking point feedback mechanism, with an internal online evaluation unit that dynamically evaluates the decision-making effectiveness of the current model through calculated indicators, and triggers incremental training or parameter fine-tuning of the model based on the evaluation results.

9. The multi-channel marketing outreach timing decision-making system based on multi-source behavioral data according to claim 8, characterized in that, The system also includes a privacy protection and security computing module: this module is located between the data access and integration module and the global feature computing center. It introduces a differential privacy algorithm to inject controlled noise into the feature vectors; the injection strength of the differential privacy is controlled by the parameter ε. Control, satisfying the following constraints: ； in, and The datasets are adjacent datasets that differ by only one sample; the privacy protection and security computing module supports encrypted computation based on a hardware trusted execution environment (TEE) to ensure that the model inference process is executed in isolated secure memory.

10. The multi-channel marketing outreach timing decision-making system based on multi-source behavioral data according to claim 8, characterized in that, The effect tracking and self-evolution module has an automatic A / B testing function: the system dynamically divides the user group into an experimental group and a control group. The experimental group applies the optimal decision logic based on reinforcement learning, while the control group applies the traditional logic based on fixed rules or random strategies. By comparing the performance of the two groups in terms of conversion rate, click-through rate, cancellation rate, and computational resource consumption rate, the effect of the technology is quantified. When the experimental group performs significantly better than the control group and the confidence level p < 0.05, the system automatically promotes the optimal model parameters to the production environment. The system also constructs a unified semantic layer for behavior based on a distributed computing engine, which maps heterogeneous behavioral instructions from different platforms to a unified behavioral classification system, and transforms heterogeneous behaviors into interactive semantic tags for marketing content, thereby eliminating strategy bias caused by data silos.

Images