Advertisement putting method and system based on multi-source data analysis

Abstract

The invention belongs to the field of advertisement putting, and provides an advertisement putting method and system based on multi-source data analysis, and the method comprises the steps: collecting multi-source original data related to a user; identifying a plurality of cognitive state nodes based on the click behavior data and the transaction path data; constructing a cognitive behavior causal atlas based on the plurality of cognitive state nodes, wherein nodes of the causal atlas represent user cognitive states; predicting an advertisement response probability and a conversion probability of a target user by adopting a Bayesian inference model in combination with a historical behavior sample and a path structure in the causal atlas, and estimating a state transition probability of the user from a current state node to a target state node based on different advertisement intervention contents; and based on the state transition probability and the causal atlas structure, determining an optimal advertisement intervention path of the user from the current cognitive state to the expected conversion state, and constructing a corresponding advertisement putting sequence based on the path.

Description

Technical Field

[0001] The present invention belongs to the field of advertising delivery, and in particular relates to an advertising delivery method and system based on multi-source data analysis. Background Art

[0002] With the development of mobile internet and the ubiquity of smart devices, advertising technology has evolved from traditional static display methods to intelligent recommendation models based on user profiles and behavioral predictions. Existing advertising systems generally use user click behavior data, browsing paths, and past purchase records as a modeling foundation. These systems then use methods such as deep learning and logistic regression to build user interest models, effectively matching content with target audiences.

[0003] However, existing advertising recommendation methods have the following shortcomings:

[0004] First, existing methods usually extract features based on a single dimension or surface behavior, which cannot fully characterize the user's behavioral evolution path and make it difficult to accurately grasp their true cognitive state and potential conversion intentions.

[0005] Secondly, most methods only infer the connection between advertising content and user behavior through statistical correlation or machine learning prediction, and lack causal mechanism modeling to determine whether advertising intervention actually causes cognitive change. As a result, the delivery strategy cannot effectively intervene in the user's psychological path, thereby affecting the overall conversion efficiency.

[0006] Thirdly, traditional advertising systems lack path-based strategic planning capabilities and are unable to formulate coherent and effective multi-stage intervention strategies from the overall perspective of cognitive evolution. This leads to problems such as fragmented and repetitive advertising content push and lack of stage adaptability, resulting in budget waste and reduced user experience. Summary of the Invention

[0007] In order to solve the problems in the prior art, the present invention provides an advertisement delivery method based on multi-source data analysis, comprising the following steps:

[0008] Collecting multi-source raw data related to users, including at least user click behavior data and transaction path data, to characterize user behavior trajectory and purchase intention;

[0009] Identifying a plurality of cognitive state nodes based on the click behavior data and the transaction path data, wherein the cognitive state nodes are used to represent the user's psychological state or behavioral stage during the advertising interaction process;

[0010] Constructing a cognitive behavior causal graph based on the multiple cognitive state nodes, wherein the nodes of the causal graph represent the user's cognitive state, and the directed edges of the causal graph represent the causal transfer relationship between different advertising contents and the user's cognitive state;

[0011] Combining historical behavior samples with the path structure in the causal graph, a Bayesian inference model is used to predict the target user's advertising response probability and conversion probability, and the state transition probability of the user from the current state node to the target state node is estimated based on different advertising intervention content;

[0012] Based on the state transition probability and causal graph structure, the optimal advertising intervention path for the user to move from the current cognitive state to the expected conversion state is determined, and a corresponding advertising delivery sequence is constructed based on the path.

[0013] Furthermore, the step of identifying multiple cognitive state nodes includes:

[0014] During the system initialization phase, a label system consisting of seven user cognitive states is predefined: initial exposure, potential interest, clear interest, wait-and-see comparison, conversion preparation, actual conversion, and post-conversion diffusion.

[0015] Constructing a user's behavior sequence based on click behavior data and transaction path data, and including behavior event type, behavior content label, entry location, and operation duration characteristics;

[0016] Input the behavior sequence into a multi-layer Transformer encoder with a self-attention mechanism for modeling, and output a contextual aggregation representation of each behavior node;

[0017] The aggregated representation is classified using a Softmax classifier to generate a corresponding cognitive state label sequence, and consistency constraints and post-processing are performed on the label sequence to obtain the user's cognitive state path.

[0018] Furthermore, the steps of constructing a cognitive behavioral causal map include:

[0019] Based on the user cognitive state label sequence, construct a state transition sample triple of the user under different advertising delivery, wherein the triple includes a state pair, an advertisement delivery, and a transition result;

[0020] Constructing a multi-type directed graph containing state nodes and advertising intervention edges, wherein the edges carry conditional probabilities of state transitions;

[0021] Use Bayesian causal inference to calculate the causal effect of state transition under advertising intervention. If the causal effect exceeds the set threshold, the edge is retained.

[0022] The training samples and causal edge weights are updated regularly based on a sliding time window. If the advertisement content changes, the subgraph is reconstructed and the intervention ability of the new advertisement is estimated.

[0023] Furthermore, the prediction steps of state transition probability include:

[0024] Build a Bayesian network that jointly models the user's current status, ad content, behavioral characteristics, and predicted status;

[0025] The graph structure of the Bayesian network is determined by a structural learning algorithm, and the conditional probability table is established by maximum likelihood estimation or Bayesian estimation;

[0026] Input the current state, the ad to be delivered, and user behavior characteristics, and use variational inference or forward sampling methods to calculate the conditional probability distribution of transitioning from the current state to each candidate state;

[0027] For multiple candidate ads, the corresponding ad response probability, state transition probability and path utility score are output respectively and used for subsequent path planning.

[0028] Furthermore, the steps of constructing the ad delivery sequence include:

[0029] Constructing a weighted directed graph of state nodes and advertising intervention edges, where the weight of the edge is a weighted function of the state transition probability and the advertising budget cost;

[0030] Use the maximum product path algorithm or the shortest path algorithm to search for the optimal path from the current state to the expected transformation state;

[0031] Determine the corresponding advertising delivery sequence according to the optimal path, and specify the advertising content, display timing, channel location and budget control parameters for each stage;

[0032] The generated ad delivery sequence is stored in the user's personalized ad plan cache, and the delivery execution engine triggers the delivery in real time based on the user's status.

[0033] Another aspect of the present invention provides an advertising delivery system based on multi-source data analysis, comprising the following modules:

[0034] A data collection module is used to collect multi-source raw data related to users. The raw data includes at least user click behavior data and transaction path data, which are used to characterize user behavior trajectory and purchase intention;

[0035] a state identification module, configured to identify a plurality of cognitive state nodes based on the click behavior data and the transaction path data, wherein the cognitive state nodes are used to represent the user's psychological state or behavioral stage during the advertising interaction process;

[0036] a causal modeling module for constructing a cognitive behavior causal graph based on the plurality of cognitive state nodes, wherein the nodes of the causal graph represent the user's cognitive state, and the directed edges of the causal graph represent the causal transfer relationship between different advertising contents and the user's cognitive state;

[0037] A probabilistic reasoning module is used to combine historical behavior samples with the path structure in the causal graph, use a Bayesian reasoning model to predict the target user's advertising response probability and conversion probability, and estimate the state transition probability of the user from the current state node to the target state node based on different advertising intervention content;

[0038] The path planning module is used to determine the optimal advertising intervention path for the user from the current cognitive state to the expected conversion state based on the state transition probability and the causal graph structure, and to construct a corresponding advertising delivery sequence based on the path.

[0039] Furthermore, the state recognition module includes:

[0040] The label system initialization unit is used to predefine a label system containing seven user cognitive states during the system initialization phase: initial exposure, potential interest, clear interest, wait-and-see comparison, conversion preparation, actual conversion, and post-conversion diffusion;

[0041] A behavior sequence construction unit is used to construct a user behavior sequence based on click behavior data and transaction path data. The behavior sequence includes behavior event type, behavior content label, entry location and operation duration characteristics;

[0042] A sequence modeling unit, configured to input the behavior sequence into a multi-layer Transformer encoder including a self-attention mechanism for modeling, and output a contextual aggregation representation of each behavior node;

[0043] The state classification unit is used to classify the context aggregation representation to generate a corresponding cognitive state label sequence, and perform consistency constraints and post-processing on the label sequence to obtain the user's cognitive state path.

[0044] Furthermore, the causal modeling module includes:

[0045] A triplet generation unit is used to construct a state transition sample triplet containing a state pair, an advertisement, and a transition result based on a user cognitive state label sequence;

[0046] a causal graph structure construction unit, configured to construct a multi-type directed graph comprising state nodes and advertising intervention edges based on the triples, wherein the edges carry conditional probabilities of state transitions;

[0047] A causal effect calculation unit is used to calculate the causal effect of state transition under advertising intervention using the Bayesian causal inference method, and retain the corresponding edge when the causal effect exceeds a set threshold;

[0048] The graph maintenance unit is used to regularly update training samples and causal edge weights based on a sliding time window, reconstruct the causal subgraph when the advertising content changes, and estimate the intervention capacity of the new advertisement.

[0049] Furthermore, the probabilistic reasoning module includes:

[0050] A Bayesian network construction unit, used to construct a Bayesian network that jointly models the user's current status, advertising content, behavioral characteristics, and predicted status. The graph structure of the network is determined by a structural learning algorithm, and the conditional probability table is established by maximum likelihood estimation or Bayesian estimation.

[0051] The state probability inference unit is used to calculate the conditional probability distribution of transitioning from the current state to each candidate state using variational inference or forward sampling methods under the conditions of inputting the current state, the advertisement to be delivered, and user behavior characteristics;

[0052] The response scoring unit is used to output the corresponding advertising response probability, state transition probability and path utility score for multiple candidate advertisements, and is used for the optimal advertisement selection in the path planning module.

[0053] Furthermore, the path planning module includes:

[0054] A graph construction and weighting unit, configured to construct a weighted directed graph of state nodes and advertising intervention edges, wherein the weight of the edge is a weighted function of the state transition probability and the advertising budget cost;

[0055] A path search unit, configured to search for an optimal path from a current state to an expected transformation state based on a maximum product path algorithm or a shortest path algorithm;

[0056] A delivery sequence generation unit is used to determine the corresponding advertisement delivery sequence according to the optimal path, and set the advertisement content, display timing, channel location and budget control parameters for each stage;

[0057] The cache and scheduling unit is used to store the advertising delivery sequence into the user's personalized advertising plan cache, and trigger the presentation of advertising content in real time by the delivery execution engine according to changes in user status.

[0058] The present invention provides an advertising delivery method and system based on multi-source data analysis, which has the following significant beneficial effects:

[0059] By collecting users' click behavior data and transaction path data, and combining it with the Transformer model for behavioral sequence modeling, we can identify the evolution trajectory of users' cognitive states during advertising interaction and depict the psychological path from exposure to conversion, effectively breaking through the limitations of traditional user profiling methods in time dimension and psychological modeling.

[0060] By constructing a cognitive behavioral causal map, we can identify the causal effects of different advertising content on the transition of users' cognitive states. Different from traditional relevance ranking, we can achieve accurate judgment of advertising intervention behavior, thereby avoiding ineffective intervention and waste of resources.

[0061] With the help of the Bayesian network structure, user characteristics, advertising attributes and state variables are integrated to achieve joint modeling of advertising response probability and state transition probability, and support online updates and real-time predictions to improve the accuracy and robustness of strategic decision-making.

[0062] Based on causal graphs and probabilistic reasoning, we further introduce graph search algorithms and Markov path planning methods to achieve the optimal advertising delivery path design from the current state to the target state, thereby improving the strategic nature and stage adaptability of the advertising system. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 is a flow chart of the method of the present invention;

[0065] Figure 2 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0066] Below, the invention is preferably described with reference to the accompanying drawings and specific embodiments.

[0067] This embodiment solves the above problem through the following steps:

[0068] In one embodiment, reference Figure 1 The present invention provides an advertising delivery method based on multi-source data analysis. By collecting and integrating user-related data from multiple heterogeneous sources, combining the user's behavioral trajectories and cognitive responses in different business systems, a unified data expression model is constructed, and cognitive behavioral analysis and causal reasoning algorithms are further introduced to identify the user's psychological state changes and behavioral conversion paths during the advertising exposure process, and ultimately realize personalized and dynamic advertising content generation and delivery strategy planning, thereby significantly improving the advertising reach efficiency and conversion effect.

[0069] Step S1: Collect multi-source original data related to the user, where the original data at least includes the user's click behavior data and transaction path data, which are used to characterize the user's behavior trajectory and purchase intention.

[0070] In advertising strategies, users' true interests and conversion intentions are often hidden in their multidimensional behavioral data. A single data dimension cannot fully reflect users' interest drivers and decision-making processes. Therefore, effectively integrating user behavioral data from multiple sources is a prerequisite for building an accurate advertising push mechanism. The core of this step is to dynamically characterize users' behavioral trajectories by fusing click behavior data and transaction path data from multiple data sources, thereby inferring their potential reactions to advertising content. Click behavior reveals users' explicit preferences, while transaction paths reveal the entire process users go through to complete their conversion intentions. Combining the two can more completely characterize users' cognitive evolution and true purchase intentions.

[0071] Select data systems from at least the following sources as the objects of information collection:

[0072] User behavior log system for the web or app;

[0073] Sales management system (such as PIC system, smart work badge);

[0074] Contract and commission management systems (such as channel distribution systems and universal marketing systems);

[0075] User reporting and transaction confirmation system.

[0076] Record every click event of users in digital advertising touchpoints (such as mini-programs, promotional links, and in-app ads); each click event should include timestamp, click location, ad content ID, ad type, user ID, page, and other information; if there is page dwell time, sliding path, and jump behavior after clicking, they should also be collected.

[0077] The transaction path is defined as the complete process from the user's first contact to the final order (subscription, contract or successful recommendation); key behavioral nodes are extracted, such as registering an account, filling in customer information, reporting customers, viewing property details, collecting commissions, signing contracts, etc.; a behavioral sequence is constructed for each user, and whether it is finally converted (that is, whether a certain transaction closed-loop behavior is completed); if there are multiple channel sources, data fusion is achieved through a unified user identification (mobile phone number, device ID, account system).

[0078] By collecting and integrating click behavior data and transaction path data through the above steps, we can not only refine the entire process from user interest to action, but also capture the fluctuations in intent implicit in their behavior. Unlike traditional static profiling, this method can depict the user's cognitive dynamics and path selection throughout the entire advertising exposure process, providing a solid data foundation for subsequent cognitive state modeling, causal reasoning, and the generation of advertising intervention paths.

[0079] For example, a user is browsing property listings on a real estate app and clicks on a "Recommend and Win" ad card. The system records the following sequence of actions:

[0080] 2025-03-02 09:15, clicked on the ad "Recommend a friend and earn commission";

[0081] On March 2, 2025, at 9:18 AM, I jumped to the project details page and browsed the "XX Garden" property for 3 minutes;

[0082] 2025-03-02 09:25, registered as a free agent and bound mobile phone number;

[0083] 2025-03-03 11:02, submit the first recommended customer information and click Submit;

[0084] At 10:00 AM on March 10, 2025, the customer signed the subscription agreement and a commission confirmation form was generated.

[0085] The above process represents a complete transaction path, and each behavioral node will be input into the subsequent modeling process as a cognitive state node. Based on this, the system can identify Xiao Li as sensitive to "commission incentives" and appreciate the "value for money" of the property, identifying him as a high-intent customer. In the future, it can deliver "more commission strategy ads" to him to further activate his communication behavior.

[0086] Step S2: identifying a plurality of cognitive state nodes based on the click behavior data and the transaction path data, wherein the cognitive state nodes are used to represent the user's psychological state or behavioral stage during the advertisement interaction process.

[0087] Throughout the advertising process, users don't exist as static "interested" or "uninterested" labels. Instead, they exhibit a dynamically evolving cognitive process at different stages, such as from "first attention" to "hesitation and comparison," to "expressing purchase intent" and even "actual conversion." While these psychological changes cannot be directly observed, they can be indirectly modeled through user behavior sequences. The purpose of identifying these cognitive state nodes is to abstract user behavior into a quantifiable cognitive path, allowing subsequent advertising interventions to accurately identify the user's "current psychological stage" and, accordingly, match appropriate advertising content and delivery timing, thereby effectively improving the effectiveness of advertising interventions and conversion success rates.

[0088] Traditional advertising systems often simply aggregate user behavior, such as "clicked the ad three times," ignoring the underlying behavioral dynamics. However, by building a cognitive state node system, each action can be assigned a specific meaning, transforming the advertising system from a "reactive" model to an "understanding" model.

[0089] The specific implementation process of step S2 includes:

[0090] Step S21: Define cognitive state label system

[0091] First, during the system initialization phase, a fixed sequence of user cognitive status labels is predefined. This sequence forms a complete link from the first ad impression to the final conversion, specifically including the following seven status labels:

[0092] State S1: Initial exposure

[0093] State S2: Potential Interest

[0094] State S3: Clear interest

[0095] State S4: Wait and see

[0096] State S5: Conversion Preparation

[0097] State S6: Actual conversion

[0098] State S7: Diffusion after transformation

[0099] This state label system is used to represent the user's psychological evolution stage during the advertising interaction process and is the classification target of the subsequent state recognition model.

[0100] Step S2: Constructing user behavior sequence

[0101] The user click behavior data and transaction path data collected in step S1 must first be constructed into a unified ordered time series input. The specific steps are as follows:

[0102] All behavioral events of each user are arranged in ascending order by timestamp to form an event sequence;

[0103] Each event is represented as a five-tuple:

[0104] e i =(t i ,a i ,c i ,p i ,s i )in:

[0105] t i Indicates the time when the event occurred;

[0106] a i Indicates the type of behavior (clicking on an ad, staying on a page, registering, submitting a recommendation, signing a contract, etc., mapped to discrete codes);

[0107] c i Indicates the content label of the behavior (such as property name, advertisement ID, commission policy ID, etc.);

[0108] pi Indicates the location or entry point of the action (Mini Program, App homepage, activity page);

[0109] s i Indicates the dwell time or operation duration (unit: seconds).

[0110] One-hot encoding is used for all discrete attributes (such as behavior type and entry location); the behavior content label is encoded using the pre-trained BERT model, and the behavior content text is converted into a vector of dimension T; the residence time and behavior time interval are normalized; the above vectors are concatenated to form a behavior feature vector sequence X = {x1, x2, ..., x T}

[0111] Step S23: Use Transformer to model user behavior sequence

[0112] Use the standard Transformer Encoder architecture to model the user's behavioral feature sequence; each input vector x i Add position encoding PE i , as input embedding; Transformer contains 4 layers of Encoder, each layer has self-attention and feedforward network, using LayerNorm and residual connection; the output is a set of context-enhanced sequence representations {h1,h2,…,h T}, each h t Represents the user's behavioral context aggregation state at time t.

[0113] Represent each moment h t Enter the Softmax classifier: The classifier outputs a 7-dimensional probability vector corresponding to the seven aforementioned cognitive states; cross-entropy loss is used for model training, and labels are derived from expert annotations of the historical conversion process or automatically generated based on the post-transaction backtracking path.

[0114] Generate the corresponding state label sequence Y = {y1,y2,…,y T}, if there are consecutive identical labels, the segments are merged to obtain a sequence of cognitive state fragments as the state path.

[0115] Step S24: State consistency constraints and post-processing

[0116] To avoid invalid state jumps in model predictions, such as abnormal paths like “S1→S5→S2”, the following constraint strategy is adopted:

[0117] The state labels have a partial order relationship, and the number of the subsequent state must not be lower than that of the previous state. The maximum subsegment method is used to correct short-term fluctuating states (such as an S2 inserted in the middle). The final output is the state transition path.

[0118] This step addresses the difficulty of interpreting and modeling user behavior by building a behavioral sequence recognition system based on the Transformer model, transforming the user behavior process into a cognitive state path with clear semantics. Compared to rule-based or static classification approaches, this method offers the following advantages: it can capture the temporal relationship between behaviors and distinguish the impact of behavioral sequence on cognitive evolution; it integrates the semantics of previous and subsequent behaviors through a self-attention mechanism, enhancing state recognition accuracy; and the generated cognitive state labels provide precise anchor points for advertising intervention and personalized content generation. Finally, the state labeling system can be continuously updated based on business needs, and the behavioral feature vector structure can be seamlessly expanded to other input dimensions.

[0119] For example, a user's recent behavior record is as follows:

[0120] 2025-04-10 10:01: Click on the "XX Garden New Project Online" ad card, browse for 2 minutes and then return;

[0121] 2025-04-10 10:05: Jump to the floor plan page and stay for 3 minutes;

[0122] 2025-04-10 10:12: View the "Recommendation Commission Earning Guide" and bind the broker identity;

[0123] 2025-04-10 11:30: Submit the first recommended customer;

[0124] 2025-04-11 09:00: Visit the project details page again;

[0125] 2025-04-12 17:30: Click on the "Limited-time Double Rebate" event advertisement to enter the subscription process.

[0126] Through the above behavioral sequence, the system automatically identifies Xiao Wang's cognitive state evolution as follows:

[0127] Initial exposure → Increased interest → Participation and interaction → Increased intention → High conversion intention

[0128] Finally, the system will mark them as users in the "high-intent stage". The system can then deliver "commission redemption speed comparison" ads to them to further guide them to sign contracts. At the same time, they will be marked as "seed spreaders" and recommended to deliver "invite friends to earn double commissions" campaign ads.

[0129] The identification of this cognitive state path enables the advertising content to be highly matched with the user's state, achieving the dual goals of precise reach and psychological intervention.

[0130] Step S3: construct a cognitive behavior causal graph based on the multiple cognitive state nodes, wherein the nodes of the causal graph represent the user's cognitive state, and the directed edges of the causal graph represent the causal transfer relationship between different advertising contents and the user's cognitive state.

[0131] In advertising delivery systems, understanding how user behavior changes in response to advertising is more critical than simply knowing what users do. Traditional recommendation systems only model behavioral correlations, ignoring the actual causal impact of advertising behavior as an intervening variable on the user's cognitive evolution path. If we cannot clearly define the causal relationship between advertising and user state transitions, advertising interventions will be blind and difficult to achieve strategic guidance. The purpose of constructing a cognitive-behavioral causal graph is to capture the causal impact of advertising on users' psychological states and behavioral transition paths, thereby enabling intelligent intervention reasoning about what type of advertising will enable users to successfully transition from one cognitive state to the next.

[0132] The nodes of the causal graph represent cognitive states, and the edges represent possible paths of state changes after an advertisement is launched. This structure can not only explain historical behavioral paths, but also be used to predict what advertisements should be launched in the current state to promote conversions, providing a solid foundation for subsequent advertising path optimization and content recommendations.

[0133] In order to achieve the above purpose, in step 3, the following steps are implemented

[0134] Step S31, prepare input data set

[0135] Using the user cognitive state sequence identified in step S2:

[0136] S u ={(t1,s1),(t2,s2),...,(t n ,s n )}

[0137] Among them, each s i Represents a cognitive state label, timestamp t i Indicates the time when the status changes.

[0138] For every two adjacent state pairs (s i ,s i+1 ), check whether there is any advertising interference behavior within the time interval:

[0139] If the user is in [t i ,t i+1 ] clicked or received an ad k, it is considered as a candidate causal pair of “advertising intervention leads to state transfer”.

[0140] Construct a training sample set:

[0141] D={(s i ,a k ,s i+1 )}

[0142] In state s i Place an ad k Then the user enters state s i+1 The observation triples of .

[0143] Step S32: Constructing a causal graph structure

[0144] Define the causal graph structure: Construct a multi-type directed graph G = (V, E), where:

[0145] The node set V = S, i.e., all cognitive state labels;

[0146] Edge set E={(s i ,s j ;a k )}, indicating an ad k The state transition path induced;

[0147] Each edge carries a conditional probability weight P(s j |s i ,a k ), indicating that from state s i Advertisement k Transfer to s j The estimated probability of .

[0148] For each state transition pair and ad type, the transition strength is calculated using the Bayesian causal inference method enhanced by conditional mutual information:

[0149]

[0150] Among them, P(s j |do(a k ),s i ) represents the state transition probability after the advertising intervention;

[0151] P(s j |s i ) represents the transition probability in the natural state (without advertising intervention);

[0152] Specifically, the causal graph inference framework DoWhy+EconML can be used to estimate the conditional causal effect.

[0153] Calculate CE for all triples. If CE(si ,a k ,s j )>δ then it is considered that advertisement a k In state s i Next pair transfer to s j If there is a significant causal effect, the edge is retained; δ is the causal significance threshold (such as 0.2), which is set through cross-validation.

[0154] Step S33: Graph update and maintenance mechanism

[0155] Advertising effectiveness changes over time, so we periodically update the training sample set and transfer statistics using a rolling window. When the ad content changes, we reconstruct the subgraph and update the causal edge weights. We also use short-term A / B testing to supplement samples for new ads, quickly estimating their intervention capabilities under different conditions.

[0156] In this step, constructing a cognitive behavioral causal map has the following significant advantages:

[0157] It can identify "which ads really drive changes in cognitive status" rather than just sorting by click frequency; it can design the optimal delivery sequence based on different ad combination paths in the graph; it not only knows who the user is, but also knows "at which stage he is and what content he needs to promote conversion"; and it is convenient for advertising delivery management decisions and effect tracking.

[0158] Specific examples:

[0159] In cognitive state S2: latent interest, the intervention of the advertisement "limited-time commission increase" leads to a transfer to S3: clear interest with a probability of 0.58, a natural transfer probability of 0.23, and a causal effect of 0.35;

[0160] In the cognitive state S4: waiting and comparing, the advertisement "customer praise video recommendation" can guide the transfer to S5: conversion preparation, with a CE of 0.41;

[0161] In the path from S3 to S5, the CE of the advertisement “FAQs” is negative, indicating that the advertisement may cause cognitive degradation, and the system will automatically avoid pushing the advertisement in the S3 state.

[0162] According to the graph structure, when the user is currently in the S2 state, the system will give priority to delivering advertising content with "limited-time commission increase" to maximize the probability of successful state transition, accurately implement psychological stage intervention, and optimize the advertising budget delivery returns.

[0163] In step S4, the Bayesian inference model is used to predict the target user's advertising response probability and conversion probability by combining historical behavior samples with the path structure in the causal graph, and the state transition probability of the user from the current state node to the target state node is estimated based on different advertising intervention contents.

[0164] The essence of advertising is a strategic optimization problem with interventionist intent. Different advertising content has varying intervention effects on different users in different cognitive states. In the previous step, we clarified the structure of "whether advertising is causal" by constructing a cognitive-behavioral causal graph. However, in actual advertising decisions, we need to further quantify the probability that each type of advertising content will successfully guide users through state transitions in a specific state, thereby achieving optimal path planning and delivery control. This problem is essentially a Bayesian conditional inference problem: given the user's known cognitive state and the advertising intervention behavior, we need to predict whether the user will transition to the next stage.

[0165] The advantage of using Bayesian reasoning is that it naturally integrates prior knowledge (historical behavioral samples) with the structural information in the causal graph. This not only outputs transition probabilities but also continuously updates prediction accuracy based on observed data, enabling adaptive intelligent inference. Furthermore, modeling within the Bayesian framework is highly interpretable, making it suitable for applications in high-risk decision-making environments, such as advertising budget control.

[0166] In order to achieve the above purpose, step S4 can be implemented through the following process:

[0167] Step S41: Constructing a state-advertisement-response triplet dataset

[0168] Extract all edges (s i ,a k ,s j ), indicating that in state s i Place an ad k Then transfer to s j observation samples.

[0169] For each state edge, record the following variables:

[0170] Current status: s i

[0171] Ad content ID: a k

[0172] Ad type features (e.g., incentive-based, content-based, social): encoded as a feature vector;

[0173] User feature vector: output by the behavior representation model in step S2;

[0174] Final state: s j , or whether it has been converted to a certain target state (such as converted / unconverted)

[0175] Construct training sample set: D={(x u ,si ,a k ,s j )}

[0176] Among them, x u It is the historical behavior representation of the user at the time of delivery, s i It is the state before launch, s j It is a subsequent state.

[0177] Step S42: Building a Bayesian network structure

[0178] Construct a joint probability distribution model:

[0179] P(S t+1 ,A t ,S t ,X u )

[0180] in:

[0181] S t Indicates the user's current status;

[0182] A t Indicates the content of the advertisement delivered;

[0183] X u Indicates the user's current behavior;

[0184] S t+1 Indicates the status of the forecast.

[0185] Construct a directed graph model of the following dependencies:

[0186] X u →S t+1 Indicates that user characteristics affect the final state;

[0187] S t →S t+1 This means that the original state has a decisive influence on the transition path;

[0188] A t →S t+1 indicates that advertising intervention is the core causal variable;

[0189] S t →A t This indicates that the type of advertising is influenced by cognitive status (i.e., targeted delivery).

[0190] When training the model, a structural learning algorithm (such as greedy structure search or maximum expected posterior algorithm) can be used to determine the graph structure; maximum likelihood estimation or Bayesian estimation can be used to calculate the conditional probability table (CPT); continuous variables (such as behavior vectors) can be modeled using a Gaussian mixture Bayesian network, or a variational Bayesian approximation method can be used to process high-dimensional variables. The specific training process is not the focus of the present invention, and those skilled in the art can use any relevant technology in the prior art to implement it.

[0191] Step S43: State transition probability prediction

[0192] In the prediction phase, input:

[0193] Current user status S t =s i ;

[0194] Current advertising plan A t =a k

[0195] User behavior feature vector X u .

[0196] Infer the target state probability distribution:

[0197] P(S t+1 =s j |S t =s i ,A t =a k ,X u )

[0198] Use variational inference or forward sampling methods to perform conditional probability inference on the Bayesian network to obtain the transition probability vector That is, from state s i Through advertising k Transition to any state s j The probability distribution of .

[0199] If the target state is set to s * , then the corresponding conversion probability is:

[0200] P convert =P(S t+1 =s * |S t ,A t ,X u )

[0201] Step S44: Output response probability and path score

[0202] For multiple slate ads k , respectively calculate the corresponding:

[0203] Ad response probability (such as click-through rate): P(click|A t =a k ,S t ,X u );

[0204] State transition probability: P(S t+1 =s * |S t ,A t =a k ,X u )

[0205] Path utility score: set weight w c , w t Comprehensive response rate and conversion rate:

[0206] Score(a k )=w c P(click)+w t P(convert)

[0207] The advertising content with the highest score is selected as the optimal intervention content and enters the next step of the advertising path planning module.

[0208] In this step, compared to traditional CTR prediction models, Bayesian networks integrate the three-dimensional causal relationship of "user state, ad intervention, and behavioral conversion," more closely resonating with actual advertising-driven mechanisms. Furthermore, Bayesian models naturally support incremental learning and prior updates, enabling real-time adjustments to predictions based on new behavioral data. Each state transition prediction can be interpreted as a probabilistic decision under specific conditions, making it suitable for strategy formulation and delivery optimization. This model quantifies edge transition probabilities based on the causal graph structure, achieving integrated modeling of both structural and quantitative information.

[0209] Specific examples:

[0210] A user is currently in cognitive state S3: clear interest. The system is faced with three alternative ads:

[0211] a1: Double the rebate for a limited time;

[0212] a2: Real estate customer word-of-mouth video;

[0213] a3: Recommended list of high commission projects.

[0214] The system infers through Bayesian networks:

[0215] P(S4|S3,a1)=0.45,P click =0.32

[0216] P(S4|S3,a2)=0.61,P click =0.21

[0217] P(S4|S3,a3)=0.37,P click =0.40

[0218] After combining the path score calculation, the system selects ad a2 for push. Although its click-through rate is not the highest, its conversion transition probability is the highest, thus reflecting the goal-driven advertising intervention strategy.

[0219] Step S5: Based on the state transition probability and the causal graph structure, determine the optimal advertising intervention path for the user to move from the current cognitive state to the expected conversion state, and construct a corresponding advertising delivery sequence based on the path.

[0220] The ultimate goal of advertising is not the "click" itself, but to guide users to complete a cognitive evolution path, continuously transitioning from the current state to the target state (such as signing, recommendation, re-conversion, etc.). However, the intervention effects of different advertisements at different cognitive stages are asymmetric, and some advertisements may be ineffective or even have negative effects at specific stages. Therefore, selecting advertisements based solely on a single conversion probability is not sufficient to guide full-process intervention. To achieve systematic and goal-oriented advertising, it is necessary to find a state transition path with the highest overall conversion probability or the lowest path cost based on the state transition probability matrix and causal graph structure obtained in the previous step, so as to formulate an advertising sequence of "what to invest first and what to invest next" and realize advertising path planning. This strategy makes advertising delivery "process-oriented" and "stage-intervention-oriented", avoiding repeated delivery, conflicting delivery and inefficient intervention.

[0221] To achieve the above purpose, the specific implementation of step S5 may include:

[0222] Step S51: define the problem modeling goal

[0223] The current state is set to S0 and the target conversion state is set to S * , build from S0 to S * A directed graph; each node in the graph is the user's cognitive state, and each edge in the graph Indicates that ad a is in state S i The state of the next trigger is transferred to S j , with transition probability

[0224] Each edge also has a cost function, specifically:

[0225] Delivery cost (such as budget consumption, display opportunity cost) or negative logarithmic probability, that is:

[0226]

[0227] The goal is to start from the current state and select a path through several advertising touchpoints to transfer the user to the target state S. * The probability of the path is the largest, or the total cost of the path is the smallest.

[0228] Step S52: Build graph structure and weight it

[0229] Traverse the causal graph and construct a directed weighted graph G = (V, E), where:

[0230] V indicates that there is a cognitive state node;

[0231] E represents all possible intervention edges in the form of triples (S i ,a,S j ).

[0232] Assign two attributes to each edge:

[0233] Transition probability weight Cost weight: such as budget consumption cost (a), or comprehensive score:

[0234]

[0235] Step S53: Path search and planning

[0236] If the goal is to maximize conversion probability: use the maximum product path algorithm (which can be converted to the shortest path through logarithmic transformation);

[0237] If the goal is to minimize the overall cost: use Dijkstra's shortest path algorithm or A* heuristic search algorithm;

[0238] Furthermore, for the case where there is uncertainty and change in state transition, the Markov decision process (MDP) is introduced and dynamic programming is used for strategy planning.

[0239] When executing path planning, input the current state S0 and the target state S * , the transition probability and cost of the advertising intervention edge; output a set of state node sequences {S0, S1, ..., S k =S *} and its corresponding advertising sequence {a0,a1,…,a k-1}.

[0240] Step S54: Generate an advertisement delivery sequence

[0241] Each state transition node corresponds to an advertising intervention action; pairing each state node in the path with an advertisement generates the following delivery sequence:

[0242] [(S0,a0),(S1,a1),…,(S k-1 ,a k-1 )]

[0243] Furthermore, for each ad:

[0244] Set the timing of delivery (such as immediately, the next day, or when user behavior meets certain conditions); specify the ad location and channel (such as mini-program recommendation, APP push, SMS notification, etc.); set the delivery budget and display frequency.

[0245] Finally, the sequence is stored in the user's personalized advertising plan cache, waiting to be triggered by the engine for execution.

[0246] This step constructs an evolutionary strategy for advertising delivery from "point" to "path". Advertisements are no longer delivered in isolation, but user behavior is guided strategically, directionally, and purposefully. Through probability-weighted path selection, the highest conversion rate can be achieved at the lowest cost. Different cognitive states are matched with different advertising types to achieve precise content matching.

[0247] Specifically and illustratively:

[0248] A user is currently in state S2: potential interest stage, and the system expects him to transition to S6: actual conversion stage.

[0249] Based on the causal graph and the state transition probability estimation of the Bayesian model, the system constructs the following transition path:

[0250]

[0251] in:

[0252] a1: "Hot property video introduction" ad, transfer probability 0.6;

[0253] a2: "Limited-time high commission referral incentive" ad, transfer probability 0.5;

[0254] a3: "Sign-up rebate payment process instructions" advertisement, transfer probability 0.8.

[0255] The total probability of path conversion is approximately:

[0256] P total =0.6×0.5×0.8=0.24

[0257] This generates the ad serving sequence:

[0258] In the first stage, a1 is pushed, and a2 is pushed 24 hours later after the user enters S3. If the user does not sign the contract within 24 hours after entering S5, the status will be re-evaluated and the dynamic planning re-investment sequence will be entered.

[0259] Through such path-based planning and intervention, advertising is no longer a single-point stimulus, but a stage-driven, conversion-oriented cognitive behavioral guidance tool.

[0260] See also Figure 2 In another embodiment, the present invention further provides an advertisement delivery system based on multi-source data analysis, comprising:

[0261] A data collection module is used to collect multi-source raw data related to users. The raw data includes at least user click behavior data and transaction path data, which are used to characterize user behavior trajectory and purchase intention;

[0262] a state identification module, configured to identify a plurality of cognitive state nodes based on the click behavior data and the transaction path data, wherein the cognitive state nodes are used to represent the user's psychological state or behavioral stage during the advertising interaction process;

[0263] a causal modeling module for constructing a cognitive behavior causal graph based on the plurality of cognitive state nodes, wherein the nodes of the causal graph represent the user's cognitive state, and the directed edges of the causal graph represent the causal transfer relationship between different advertising contents and the user's cognitive state;

[0264] A probabilistic reasoning module is used to combine historical behavior samples with the path structure in the causal graph, use a Bayesian reasoning model to predict the target user's advertising response probability and conversion probability, and estimate the state transition probability of the user from the current state node to the target state node based on different advertising intervention content;

[0265] The path planning module is used to determine the optimal advertising intervention path for the user from the current cognitive state to the expected conversion state based on the state transition probability and the causal graph structure, and to construct a corresponding advertising delivery sequence based on the path.

[0266] In a further implementation, the state identification module includes:

[0267] The label system initialization unit is used to predefine a label system containing seven user cognitive states during the system initialization phase: initial exposure, potential interest, clear interest, wait-and-see comparison, conversion preparation, actual conversion, and post-conversion diffusion;

[0268] A behavior sequence construction unit is used to construct a user behavior sequence based on click behavior data and transaction path data. The behavior sequence includes behavior event type, behavior content label, entry location and operation duration characteristics;

[0269] A sequence modeling unit, configured to input the behavior sequence into a multi-layer Transformer encoder including a self-attention mechanism for modeling, and output a contextual aggregation representation of each behavior node;

[0270] The state classification unit is used to classify the context aggregation representation to generate a corresponding cognitive state label sequence, and perform consistency constraints and post-processing on the label sequence to obtain the user's cognitive state path.

[0271] In a further implementation, the causal modeling module includes:

[0272] A triplet generation unit is used to construct a state transition sample triplet containing a state pair, an advertisement, and a transition result based on a user cognitive state label sequence;

[0273] a causal graph structure construction unit, configured to construct a multi-type directed graph comprising state nodes and advertising intervention edges based on the triples, wherein the edges carry conditional probabilities of state transitions;

[0274] A causal effect calculation unit is used to calculate the causal effect of state transition under advertising intervention using the Bayesian causal inference method, and retain the corresponding edge when the causal effect exceeds a set threshold;

[0275] The graph maintenance unit is used to regularly update training samples and causal edge weights based on a sliding time window, reconstruct the causal subgraph when the advertising content changes, and estimate the intervention capacity of the new advertisement.

[0276] In a further implementation, the probabilistic reasoning module includes:

[0277] A Bayesian network construction unit, used to construct a Bayesian network that jointly models the user's current status, advertising content, behavioral characteristics, and predicted status. The graph structure of the network is determined by a structural learning algorithm, and the conditional probability table is established by maximum likelihood estimation or Bayesian estimation.

[0278] The state probability inference unit is used to calculate the conditional probability distribution of transitioning from the current state to each candidate state using variational inference or forward sampling methods under the conditions of inputting the current state, the advertisement to be delivered, and user behavior characteristics;

[0279] The response scoring unit is used to output the corresponding advertising response probability, state transition probability and path utility score for multiple candidate advertisements, and is used for the optimal advertisement selection in the path planning module.

[0280] In a further implementation, the path planning module includes:

[0281] A graph construction and weighting unit, configured to construct a weighted directed graph of state nodes and advertising intervention edges, wherein the weight of the edge is a weighted function of the state transition probability and the advertising budget cost;

[0282] A path search unit, configured to search for an optimal path from a current state to an expected transformation state based on a maximum product path algorithm or a shortest path algorithm;

[0283] A delivery sequence generation unit is used to determine the corresponding advertisement delivery sequence according to the optimal path, and set the advertisement content, display timing, channel location and budget control parameters for each stage;

[0284] The cache and scheduling unit is used to store the advertising delivery sequence into the user's personalized advertising plan cache, and trigger the presentation of advertising content in real time by the delivery execution engine according to changes in user status.

[0285] It should be noted that the explanation of the embodiment of the above-mentioned advertising delivery method based on multi-source data analysis is also applicable to the device of the embodiment of the present application and will not be repeated here.

[0286] Those skilled in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented using a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0287] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0288] In the several embodiments provided in this application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), magnetic disk or optical disk, and other media that can store program code.

[0289] The above is only a specific embodiment of the present application. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in this application, which should be included in the scope of protection of this application. For some module structures that are not particularly clear in the present invention, the content recorded in the prior art shall prevail. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be regarded as part of the present invention and is used to understand the meaning of some technical features or parameters.

Claims

1. An advertising delivery method based on multi-source data analysis, characterized in that: The method comprises the following steps: Collecting multi-source raw data related to users, including at least user click behavior data and transaction path data, to characterize user behavior trajectory and purchase intention; Identifying a plurality of cognitive state nodes based on the click behavior data and the transaction path data, wherein the cognitive state nodes are used to represent the user's psychological state or behavioral stage during the advertising interaction process; Constructing a cognitive behavior causal graph based on the multiple cognitive state nodes, wherein the nodes of the causal graph represent the user's cognitive state, and the directed edges of the causal graph represent the causal transfer relationship between different advertising contents and the user's cognitive state; Combining historical behavior samples with the path structure in the causal graph, a Bayesian inference model is used to predict the target user's advertising response probability and conversion probability, and the state transition probability of the user from the current state node to the target state node is estimated based on different advertising intervention content; Based on the state transition probability and causal graph structure, the optimal advertising intervention path for the user to move from the current cognitive state to the expected conversion state is determined, and a corresponding advertising delivery sequence is constructed based on the path.

2. The advertising delivery method based on multi-source data analysis according to claim 1, characterized in that: The step of identifying multiple cognitive state nodes includes: During the system initialization phase, a label system consisting of seven user cognitive states is predefined: initial exposure, potential interest, clear interest, wait-and-see comparison, conversion preparation, actual conversion, and post-conversion diffusion. Constructing a user's behavior sequence based on click behavior data and transaction path data, and including behavior event type, behavior content label, entry location, and operation duration characteristics; Input the behavior sequence into a multi-layer Transformer encoder with a self-attention mechanism for modeling, and output a contextual aggregation representation of each behavior node; The aggregated representation is classified using a Softmax classifier to generate a corresponding cognitive state label sequence, and consistency constraints and post-processing are performed on the label sequence to obtain the user's cognitive state path.

3. The advertising delivery method based on multi-source data analysis according to claim 1, characterized in that: The steps of constructing the cognitive behavior causal map include: Based on the user cognitive state label sequence, construct a state transition sample triple of the user under different advertising delivery, wherein the triple includes a state pair, an advertisement delivery, and a transition result; Constructing a multi-type directed graph containing state nodes and advertising intervention edges, wherein the edges carry conditional probabilities of state transitions; Use Bayesian causal inference to calculate the causal effect of state transition under advertising intervention. If the causal effect exceeds the set threshold, the edge is retained. The training samples and causal edge weights are updated regularly based on a sliding time window. If the advertisement content changes, the subgraph is reconstructed and the intervention ability of the new advertisement is estimated.

4. The advertising delivery method based on multi-source data analysis according to claim 1, characterized in that: The step of predicting the state transition probability includes: Build a Bayesian network that jointly models the user's current status, ad content, behavioral characteristics, and predicted status; The graph structure of the Bayesian network is determined by a structural learning algorithm, and the conditional probability table is established by maximum likelihood estimation or Bayesian estimation; Input the current state, the ad to be delivered, and user behavior characteristics, and use variational inference or forward sampling methods to calculate the conditional probability distribution of transitioning from the current state to each candidate state; For multiple candidate ads, the corresponding ad response probability, state transition probability and path utility score are output respectively and used for subsequent path planning.

5. The advertising delivery method based on multi-source data analysis according to claim 1, characterized in that: The steps of constructing the advertisement delivery sequence include: Constructing a weighted directed graph of state nodes and advertising intervention edges, where the weight of the edge is a weighted function of the state transition probability and the advertising budget cost; Use the maximum product path algorithm or the shortest path algorithm to search for the optimal path from the current state to the expected transformation state; Determine the corresponding advertising delivery sequence according to the optimal path, and specify the advertising content, display timing, channel location and budget control parameters for each stage; The generated ad delivery sequence is stored in the user's personalized ad plan cache, and the delivery execution engine triggers the delivery in real time based on the user's status.

6. An advertising delivery system based on multi-source data analysis, characterized in that: The system includes the following modules: A data collection module is used to collect multi-source raw data related to users. The raw data includes at least user click behavior data and transaction path data, which are used to characterize user behavior trajectory and purchase intention; a state identification module, configured to identify a plurality of cognitive state nodes based on the click behavior data and the transaction path data, wherein the cognitive state nodes are used to represent the user's psychological state or behavioral stage during the advertising interaction process; a causal modeling module for constructing a cognitive behavior causal graph based on the plurality of cognitive state nodes, wherein the nodes of the causal graph represent the user's cognitive state, and the directed edges of the causal graph represent the causal transfer relationship between different advertising contents and the user's cognitive state; A probabilistic reasoning module is used to combine historical behavior samples with the path structure in the causal graph, use a Bayesian reasoning model to predict the target user's advertising response probability and conversion probability, and estimate the state transition probability of the user from the current state node to the target state node based on different advertising intervention content; The path planning module is used to determine the optimal advertising intervention path for the user from the current cognitive state to the expected conversion state based on the state transition probability and the causal graph structure, and to construct a corresponding advertising delivery sequence based on the path.

7. The advertising delivery system based on multi-source data analysis according to claim 6, characterized in that: The state recognition module includes: The label system initialization unit is used to predefine a label system containing seven user cognitive states during the system initialization phase: initial exposure, potential interest, clear interest, wait-and-see comparison, conversion preparation, actual conversion, and post-conversion diffusion; A behavior sequence construction unit is used to construct a user behavior sequence based on click behavior data and transaction path data. The behavior sequence includes behavior event type, behavior content label, entry location and operation duration characteristics; A sequence modeling unit, configured to input the behavior sequence into a multi-layer Transformer encoder including a self-attention mechanism for modeling, and output a contextual aggregation representation of each behavior node; The state classification unit is used to classify the context aggregation representation to generate a corresponding cognitive state label sequence, and perform consistency constraints and post-processing on the label sequence to obtain the user's cognitive state path.

8. The advertising delivery system based on multi-source data analysis according to claim 6, characterized in that: The causal modeling module includes: A triplet generation unit is used to construct a state transition sample triplet containing a state pair, an advertisement, and a transition result based on a user cognitive state label sequence; a causal graph structure construction unit, configured to construct a multi-type directed graph comprising state nodes and advertising intervention edges based on the triples, wherein the edges carry conditional probabilities of state transitions; A causal effect calculation unit is used to calculate the causal effect of state transition under advertising intervention using the Bayesian causal inference method, and retain the corresponding edge when the causal effect exceeds a set threshold; The graph maintenance unit is used to regularly update training samples and causal edge weights based on a sliding time window, reconstruct the causal subgraph when the advertising content changes, and estimate the intervention capacity of the new advertisement.

9. The advertising delivery system based on multi-source data analysis according to claim 6, characterized in that: The probabilistic reasoning module includes: A Bayesian network construction unit, used to construct a Bayesian network that jointly models the user's current status, advertising content, behavioral characteristics, and predicted status. The graph structure of the network is determined by a structural learning algorithm, and the conditional probability table is established by maximum likelihood estimation or Bayesian estimation. The state probability inference unit is used to calculate the conditional probability distribution of transitioning from the current state to each candidate state using variational inference or forward sampling methods under the conditions of inputting the current state, the advertisement to be delivered, and user behavior characteristics; The response scoring unit is used to output the corresponding advertising response probability, state transition probability and path utility score for multiple candidate advertisements, and is used for the optimal advertisement selection in the path planning module.

10. The advertising delivery system based on multi-source data analysis according to claim 6, characterized in that: The path planning module includes: A graph construction and weighting unit, configured to construct a weighted directed graph of state nodes and advertising intervention edges, wherein the weight of the edge is a weighted function of the state transition probability and the advertising budget cost; A path search unit, configured to search for an optimal path from a current state to an expected transformation state based on a maximum product path algorithm or a shortest path algorithm; A delivery sequence generation unit is used to determine the corresponding advertisement delivery sequence according to the optimal path, and set the advertisement content, display timing, channel location and budget control parameters for each stage; The cache and scheduling unit is used to store the advertising delivery sequence into the user's personalized advertising plan cache, and trigger the presentation of advertising content in real time by the delivery execution engine according to changes in user status.

Images