A cross-platform intelligent analysis and delivery method fusing car networking data and internet marketing data

Abstract

The application belongs to the technical field of intelligent marketing data of Internet of Vehicles, and discloses a cross-platform intelligent analysis and delivery method fusing Internet of Vehicles data and Internet marketing data. Cross-platform data is unified into a semantic label-feature value-time stamp triple format through a data standardization alignment stage, and data correlation is established in combination with a cross-modal semantic mapping algorithm. In a federal causal fusion stage, horizontal federal learning is taken as a framework, time sequence dynamic causal graphs are fused to eliminate false correlations, homomorphic encryption and differential privacy double protection mechanisms are simultaneously adopted, and noise is adjusted according to data sensitivity classification. The multi-source data fusion problem is solved, and a unified data resource pool is formed. A four-level system of basic label-scene semantic label-causal correlation label-demand intention label is established: the basic label extracts attribute data, the scene semantic label is generated based on a four-dimensional scene semantic graph, the causal correlation label is derived from causal inference results, and the demand intention label integrates the first three features.

Description

Technical Field

[0001] This invention belongs to the field of intelligent technology for vehicle network marketing data, specifically a cross-platform intelligent analysis and delivery method that integrates vehicle network data and internet marketing data. Background Technology

[0002] The rapid popularization of vehicle-to-everything (V2X) technology, coupled with the widespread application of smart terminals, has enabled internet marketing to gradually break away from traditional models and move towards scenario-based and precise approaches. The synergistic interaction between V2X and internet marketing can break down data barriers between different platforms and deeply explore the value behind users' entire behavioral journey. However, existing technologies still face the following technical challenges:

[0003] Existing technologies can only splice together the surface features of vehicle network data and marketing data, such as spatiotemporal matching and user identification association. They cannot remove false correlations in the data, such as the coincidental association between a user's accidental passage through a business district and an ad click. This leads to misjudgments and biases in demand prediction and advertising strategies based on fused data, making it difficult to accurately locate the driving factors of users' real needs.

[0004] Vehicle-to-everything (V2X) scenarios have complex semantic attributes, such as the need for relaxation during commuting congestion and the need for replenishment after long-distance driving. Existing technologies can only extract shallow features such as driving trajectory and vehicle speed, and cannot analyze the user intent behind the scenario. At the same time, the volume of cross-platform data is growing exponentially, and traditional centralized computing architectures are unable to support end-to-end real-time response of data fusion, demand prediction and delivery decision-making, resulting in a time lag between delivery strategies and dynamic user scenarios, and missing the best time to reach users. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-platform intelligent analysis and delivery method that integrates vehicle network data and internet marketing data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a cross-platform intelligent analysis and delivery method that integrates vehicle network data and internet marketing data, the specific steps of which are as follows:

[0007] Data collection and preprocessing stage: Collect multimodal data from the Internet of Vehicles and multimodal data from Internet marketing, preprocess the collected data, and output the preprocessed data;

[0008] Data standardization and alignment stage: The preprocessed data is standardized in format and encoding, cross-modal associations are established, conflicting data is removed and missing data is filled, and finally standardized semantically aligned data is output.

[0009] Federated Causal Fusion Stage: Based on the standardized semantic alignment data, the initial data association is completed; a temporal dynamic causal graph and a dual privacy protection mechanism are integrated into the horizontal federated learning framework to eliminate false correlations and establish a data resource pool;

[0010] User tag construction phase: Based on the data resource pool, establish a four-level tag system, including basic tags, scenario semantic tags, causal association tags, and demand intent tags; dynamically adjust tag weights and output a dynamic user tag system;

[0011] Demand forecasting modeling phase: A hybrid model is established to extract shallow features and enhance user understanding; federated distillation and model compression techniques are used, reinforcement learning algorithms are introduced to optimize model parameters, and demand forecasting results are output;

[0012] Real-time delivery execution phase: Based on demand forecasting and scenario semantics, delivery decisions are broken down into edge execution and cloud optimization; relying on a three-tier computing power pool architecture, delivery is carried out simultaneously across multiple channels, communication latency is monitored, and delivery performance and behavioral data are collected at the same time;

[0013] Causal closed-loop iterative stage: Based on the campaign effect and behavioral data, calculate the causal effect and establish a global attribution model; use the attribution results to empower the preceding steps.

[0014] Preferably, the data acquisition and preprocessing stage is as follows:

[0015] Multimodal vehicle network data is collected through in-vehicle terminals, including driving trajectory, driving behavior, vehicle condition sensor data and scene context data. Edge nodes filter invalid data in real time and annotate the initial scene semantics.

[0016] Simultaneously, multimodal data of internet marketing is collected through marketing platform APIs and SDKs, and associated with semantic tags of marketing content; during the collection process, a semantic tag pre-annotation mechanism is embedded to perform preliminary feature extraction on unstructured data, filter out data with no semantic value at edge nodes in advance, and finally output preprocessed data.

[0017] Preferably, the data standardization and alignment stage is as follows:

[0018] Based on the preprocessed data from the data acquisition and preprocessing stage, the vehicle network JSON format data and marketing CSV format data are first uniformly converted into a semantic tag-feature value-timestamp triplet format according to the cross-platform data semantic standard protocol. At the same time, the user virtual identifier and scene semantic coding rules are standardized.

[0019] For vehicle-to-everything (V2X) sensor data, marketing text, and image data, a cross-modal semantic mapping algorithm is used to establish a connection. Then, a semantic consistency verification algorithm is used to eliminate logically conflicting data. The missing data is filled in by semantic association, and finally, standardized semantically aligned data is output.

[0020] Preferably, the federal causal fusion stage is as follows:

[0021] Based on the standardized semantic alignment data from the data standardization and alignment stage, the initial association between vehicle network data and marketing data is first completed using the user's unique virtual identifier as the core and combined with a spatiotemporal semantic matching algorithm. Then, a time-series dynamic causal graph module is integrated into the horizontal federated learning framework, and an adaptive sliding time window mechanism is adopted to automatically reconstruct the causal structure after each round of federated training, thereby capturing the dynamic correlation of time-series data in real time.

[0022] Meanwhile, the framework incorporates homomorphic encryption and differential privacy dual protection mechanisms to remove false correlations; ultimately, it uncovers the true causal relationship between scene semantics, marketing behavior, and consumption conversion, and establishes a unified data resource pool.

[0023] Preferably, the user tag construction stage is as follows:

[0024] Based on the data resource pool of the aforementioned federated causal fusion stage, a four-level system is established: basic tags, scene semantic tags, causal association tags, and demand intent tags. Basic tags are extracted from the basic attribute data of the resource pool; scene semantic tags are generated based on scene data in the resource pool and combined with a four-dimensional scene semantic graph, with each type of scene associated with a corresponding causal attribution rule library; causal association tags are directly derived from the causal inference results in the resource pool; and demand intent tags integrate the features of the first three levels of tags to characterize the user's core needs.

[0025] Subsequently, a semantic similarity algorithm is used, combined with real-time semantic changes in user scenarios and feedback from marketing behaviors. An improved version of the DQN reinforcement learning algorithm is then used to automatically adjust the tag weights and association strengths, outputting a dynamic user tag system.

[0026] Preferably, the demand forecasting modeling stage is as follows:

[0027] Based on the dynamic user tag system in the user tag construction phase, a lightweight CNN-Transformer hybrid model is established. The CNN module extracts shallow features of the scene, and the Transformer module parses the multimodal relationship between scene semantics and marketing content, thereby enhancing the ability to understand user intent. Federated distillation technology is adopted, combined with dynamic pruning and lightweight model compression technology to reduce model transmission and computation costs.

[0028] Simultaneously, a reinforcement learning algorithm with real-time conversion efficiency as the reward function is introduced to dynamically optimize model parameters and ultimately output user demand prediction results.

[0029] Preferably, the real-time delivery execution phase is as follows:

[0030] Based on the user demand prediction results in the demand prediction modeling stage, and combined with the real-time scene semantics of the Internet of Vehicles, the marketing task intelligent segmentation algorithm is first used to split the delivery decision into the edge real-time execution segment and the cloud optimization segment. Then, the delivery is executed based on the three-level computing power pool architecture of vehicle terminal, roadside edge node and regional cloud.

[0031] Edge nodes are responsible for real-time monitoring of scene semantics and judgment of delivery triggers. The cloud enables multi-channel synchronous scheduling through the marketing cloud platform API interface. At the same time, it integrates a 5G-V2X communication monitoring module to collect multi-channel delivery effect data and vehicle network subsequent behavior data in real time while executing delivery.

[0032] Preferably, the causal closed-loop iteration stage is as follows:

[0033] Using the data on the delivery effect collected during the real-time delivery execution phase and the subsequent behavior data of the Internet of Vehicles as input, a hierarchical counterfactual inference algorithm is introduced. The edge end calculates the difference in causal effect between delivery intervention and natural conversion in real time, and the cloud establishes a global attribution model based on data from multiple edge nodes to determine the causal relationship between the delivery strategy and the conversion result and eliminate external interference factors.

[0034] The causal attribution results are then used to power the preceding steps: optimizing the federal causal inference parameters in the federal causal fusion stage, adjusting the semantic understanding weights of the demand forecasting model in the demand forecasting modeling stage, and updating the user causal association tags in the user tag building stage, ultimately establishing a full-link causal process of fusion, prediction, delivery, attribution, and iteration.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. This invention unifies cross-platform data into a semantic label-feature value-timestamp triplet format through a data standardization and alignment stage, and establishes data associations by combining cross-modal semantic mapping algorithms. In the federated causal fusion stage, it uses a horizontal federated learning framework, incorporates a time-series dynamic causal graph to eliminate spurious correlations, and adopts a dual protection mechanism of homomorphic encryption and differential privacy. The passive party only transmits the encrypted gradient, and the noise is adjusted according to the data sensitivity level. This solves the problem of multi-source data fusion, forms a unified data resource pool, avoids privacy leakage, and meets data security requirements.

[0037] 2. This invention establishes a four-level system: basic tags, scene semantic tags, causal association tags, and demand intent tags. Basic tags extract attribute data, scene semantic tags are generated based on a four-dimensional scene semantic graph, causal association tags originate from causal inference results, and demand intent tags integrate the features of the first three levels. Then, through an improved version of the DQN reinforcement learning algorithm, the tag weights are dynamically adjusted in combination with real-time user scene and behavioral feedback. Combined with a lightweight CNN-Transformer hybrid model and federated distillation technology, the accuracy of demand prediction is improved, enabling precise matching of delivery to the scene and reducing ineffective delivery.

[0038] 3. In the causal closed-loop iteration stage, this invention uses the campaign effect and behavioral data as input, and uses a hierarchical counterfactual inference algorithm to enable the edge to calculate the causal effect difference in real time. The cloud establishes a global attribution model to eliminate interference and clarify the true causal relationship between campaign and conversion. Then, the attribution results are used to empower the preceding steps to optimize the federated causal inference parameters, adjust the semantic weight of the demand prediction model, and update the user causal association tags. The entire closed-loop of integration-prediction-campaign-attribution-iteration solves the pain point of only campaigning without optimization, continuously improves marketing conversion rate and ROI, and adapts to the dynamic changes in the Internet of Vehicles scenario. Attached Figure Description

[0039] Figure 1 This is a flowchart of the cross-platform intelligent analysis and delivery method that integrates vehicle network data and internet marketing data, as described in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] like Figure 1 As shown in the figure, this embodiment of the invention provides a cross-platform intelligent analysis and delivery method that integrates vehicle network data and internet marketing data. The specific steps of the method are as follows:

[0042] The data acquisition and preprocessing stage involves collecting multimodal vehicle network data through on-board units (OBU, T-Box), including driving trajectory, driving behavior, vehicle condition sensor data, and scene context data. Edge nodes filter invalid data in real time and label the initial scene semantics. The vehicle condition sensor data includes energy consumption and fault codes, and the scene context data includes road condition level, time period type, and geographical location semantic labels.

[0043] The OBU adopts the 5G-V2X communication protocol and supports an adjustable data acquisition frequency of 1Hz-10Hz; the T-Box supports CAN bus data reading, with a default acquisition frequency of 5Hz, and the acquisition interval of vehicle condition sensor data (energy consumption, fault codes) does not exceed 1 second.

[0044] Simultaneously, multimodal data on internet marketing is collected through marketing platform APIs and SDKs, including user browsing history, click behavior, conversion results, and marketing content data, and semantic tags are associated with marketing content such as leisure and entertainment and automotive services. Marketing content data includes text ads, image materials, and video clips. During the collection process, a semantic tag pre-annotation mechanism is embedded to perform preliminary feature extraction on unstructured data such as marketing video frames and vehicle fault voice prompts, and invalid data is filtered out in advance at edge nodes, and finally pre-processed data is output.

[0045] Invalid data includes: abnormal coordinates in the driving trajectory that exceed the city's geographical boundaries, abnormal values ​​in driving behavior data with a speed greater than 200 km / h, and repeated values ​​of the same fault code three times in a row in sensor data; data without semantic value includes: pure black / pure white frames without advertising content in marketing video frames, and meaningless noise with a duration of less than 2 seconds in vehicle fault voice prompts.

[0046] The semantic label pre-annotation mechanism is implemented based on a lightweight ResNet-18 model: for marketing text data, semantic labels are annotated by combining jieba word segmentation with an industry marketing thesaurus; for marketing image / video frames, features are extracted by ResNet-18 and matched with a pre-trained marketing material feature library.

[0047] The data standardization and alignment stage is specifically as follows:

[0048] Based on the preprocessed data from the data acquisition and preprocessing stage, the vehicle network JSON format data and marketing CSV format data are first uniformly converted into a semantic tag-feature value-timestamp triplet format according to the cross-platform data semantic standard protocol. At the same time, the user virtual identifier and scene semantic coding rules are standardized.

[0049] For vehicle network sensor data and marketing text and image data, a cross-modal semantic mapping algorithm is used to establish a correlation, such as the semantic correlation between the vehicle low fuel sensor signal and the gas station discount text advertisement, to solve the problem of semantic fragmentation of multimodal data. Then, a semantic consistency verification algorithm is used to remove logically conflicting data such as long-distance driving scenarios and short-distance shopping marketing clicks. The missing data is filled in by semantic correlation to improve the model, and finally standardized semantically aligned data is output.

[0050] The cross-modal semantic mapping algorithm is based on cosine similarity calculation: it converts vehicle network sensor data and marketing data into the same high-dimensional feature space, and establishes an association when the similarity is >0.7; for image-based marketing data, it first extracts text description vectors through the CLIP model, and then calculates the similarity with the sensor data feature vectors.

[0051] The conflict determination rules of the semantic consistency verification algorithm are as follows: when the matching degree between the semantic encoding of the vehicle network scenario and the marketing tag encoding is less than 0.3, it is determined to be a conflict; when both conditions are met, namely "the duration of the scenario is less than the effective duration of the marketing content" and "the user's behavior history does not show any preference for this type of marketing", it is also determined to be a conflict.

[0052] The federal causal fusion phase is specifically as follows:

[0053] Based on the standardized semantic alignment data in the data standardization and alignment stage, the user's unique virtual identifier is used as the core, combined with spatiotemporal semantic matching algorithms, such as semantic association matching between commuting scenarios and workplace consumer marketing data, to complete the initial association between vehicle network data and marketing data; then, a time-series dynamic causal graph module is integrated into the horizontal federated learning framework, using a sliding time window mechanism that is adaptively adjusted from 1 to 5 minutes, and automatically reconstructing the causal structure after each round of federated training to capture the dynamic association relationship of time-series data in real time.

[0054] The sliding time window is adjusted based on the following criteria: when the frequency of changes in the vehicle-to-everything (V2X) scenario is greater than 3 times per minute, the window is set to 1 minute; when the scenario is stable, the window is set to 5 minutes; the adjustment trigger condition is "data correlation change greater than 20% within 3 consecutive windows";

[0055] Meanwhile, the framework incorporates a dual protection mechanism of homomorphic encryption and differential privacy. The passive party only transmits the encrypted causal effect gradient, and the noise is dynamically adjusted according to the data sensitivity level to remove false correlations while ensuring data privacy. Ultimately, it explores the real causal relationship between scenario semantics, marketing behavior and consumption conversion, and establishes a unified data resource pool.

[0056] Homomorphic encryption uses the Paillier algorithm with a public key length of 2048 bits and an encryption efficiency of ≥1000 messages / second; differential privacy uses the Laplace noise mechanism with noise intensity ε set according to data sensitivity levels; causal gradient transmission requires "noise addition-fragmentation" processing before transmission to prevent gradient backpropagation attacks.

[0057] Formula for the difference in causal effects:

[0058] ;

[0059] In the formula: It represents the difference in causal effect of marketing campaigns for a single user or in a single scenario, and is used to quantify the real impact of marketing campaign interventions on consumer conversion results.

[0060] This indicates the intervention group results, which are the actual conversion results generated after users receive the targeted marketing campaign, such as clicking on marketing text / images, completing a purchase, using a recommended coupon, or going to a recommended location.

[0061] This represents the results of the control group, i.e., the natural conversion results when users did not receive the target marketing campaign, conversions not caused by campaign factors, such as users going to the gas station on their own initiative.

[0062] The user tag construction phase is specifically as follows:

[0063] Based on the data resource pool of the aforementioned federated causal fusion stage, a four-level system is established: basic tags, scene semantic tags, causal association tags, and demand intent tags. Basic tags are extracted from the basic attribute data of the resource pool; scene semantic tags are generated based on scene data in the resource pool, combined with a four-dimensional scene semantic graph, and each type of scene is associated with a corresponding causal attribution rule library; causal association tags (such as business district stay → catering consumption conversion) are directly derived from the causal inference results in the resource pool; and demand intent tags integrate the features of the first three levels of tags to characterize the user's core needs.

[0064] Basic tags include age, gender, and vehicle type; the four-dimensional scene semantic graph includes spatiotemporal environment, vehicle status, user behavior, and marketing objectives.

[0065] The four-dimensional scene semantic graph indicators for each dimension are: ① Spatiotemporal environment: longitude, latitude, time period; ② Vehicle status: fuel level, battery level, fault code; ③ User behavior: number of marketing clicks, number of consumption conversions, and scene dwell time in the past 30 days; ④ Marketing objectives: conversion priority and target audience.

[0066] Subsequently, a semantic similarity algorithm was used, combined with real-time semantic changes in user scenarios and feedback from marketing behavior, and the improved reinforcement learning algorithm of DQN was used to automatically adjust the tag weights and association strengths, outputting a dynamic user tag system;

[0067] Improvements to the DQN reinforcement learning algorithm: ① Employ a dual experience replay pool to enhance sample diversity; ② Introduce a priority sampling mechanism to double the sampling weight of high-reward samples; Core parameters: learning rate α = 0.001, discount factor γ = 0.9, and exploration rate ε linearly decays from 0.1 to 0.01.

[0068] Tag weight adjustment trigger conditions: ① Real-time semantic changes in user scenarios; ② Negative feedback from marketing behavior; ③ Conversion efficiency associated with tags decreases by more than 5% for 3 consecutive days; The adjustment range is ±0.1, ensuring that the core tag weight is not lower than 0.6.

[0069] The demand forecasting modeling stage is specifically as follows:

[0070] Based on the dynamic user tag system built during the user tag construction phase, a lightweight CNN-Transformer hybrid model is established. The CNN module extracts shallow scene features, such as driving trajectory and vehicle speed, while the Transformer module parses the multimodal relationship between scene semantics and marketing content, enhancing the ability to understand user intent. To adapt to the subsequent real-time computing needs of the edge, federated distillation technology is adopted, with sub-models trained locally on the vehicle network and marketing platform, parameters aggregated in the cloud, and lightweight model compression technology with 30% to 60% dynamic pruning to reduce model transmission and computing costs.

[0071] Dynamic pruning employs a pruning strategy based on neuron contribution: calculating the contribution of each neuron to the model's prediction accuracy and removing neurons with a contribution value <0.01; pruning ratio adjustment rules: 60% pruning at the edge and 30% pruning at the cloud; the model accuracy loss after pruning is controlled to <5%;

[0072] At the same time, a reinforcement learning algorithm with real-time conversion efficiency as the reward function is introduced to dynamically optimize the model parameters and finally output the user demand prediction results.

[0073] The rules for setting the values ​​of α, β, and γ in the reinforcement learning reward function are as follows: ① Automotive service marketing: α=0.6 (conversion rate weight), β=0.3 (CTR weight), γ=0.1 (cost weight); ② Leisure and entertainment marketing: α=0.4, β=0.5, γ=0.1; ③ The weights are updated every 24 hours based on the previous day's campaign performance, and an immediate adjustment is triggered when a certain indicator fluctuates by more than 15%.

[0074] Reinforcement learning reward function:

[0075] ;

[0076] In the formula: The real-time reward value represents the reinforcement learning algorithm and is the core objective of optimizing the parameters of the demand forecasting model. The higher the value, the better the demand forecasting results output by the current model match the business objectives of high conversion and low cost.

[0077] , , This represents a dynamic weighting coefficient that is adjusted in real time according to the marketing scenario (such as automotive service marketing or leisure and entertainment marketing) to balance the importance of different evaluation indicators.

[0078] This refers to the conversion rate of marketing campaigns, which is the percentage of users who complete a target action (such as making a purchase or booking a car service) after receiving marketing content.

[0079] CTR This indicates the click-through rate, which is the percentage of users who click on marketing content (text ads, image assets, video clips, etc.).

[0080] It represents the overall cost of marketing campaigns, including the computing power consumption of the three-tier computing power pool (vehicle terminals, roadside edge nodes, and regional cloud), the interface costs of multi-channel (vehicle-mounted systems, mobile apps, and offline terminals), etc., calculated by combining computing power load data and channel cost data during the real-time campaign execution phase.

[0081] The real-time delivery execution phase is specifically as follows:

[0082] Based on the user demand prediction results from the demand forecasting modeling stage, and combined with the real-time semantics of the vehicle network scenario, such as congested commuting → audio entertainment ads, low fuel level → gas station discounts, the marketing task intelligent segmentation algorithm first divides the delivery decision into an edge real-time execution segment and a cloud optimization segment. Then, it relies on a three-level computing power pool architecture of vehicle terminals, roadside edge nodes, and regional cloud to execute the delivery. When the load exceeds the 70% threshold, it calls on the computing power of surrounding vehicles to form a temporary cluster. The edge real-time execution segment includes scene matching and initial delivery, while the cloud optimization segment includes global strategy iteration and effect attribution.

[0083] The load calculation index is the average of the CPU utilization and memory usage of the three-level computing power pool: ① Vehicle terminal load = (CPU utilization + memory usage) / 2, and a load exceeding 70% is considered high load; ② The load calculation of roadside edge nodes must include the number of connected vehicles at the same time; ③ Temporary cluster calling rules: prioritize vehicles with load < 30% and distance < 1km, and the maximum size of a single cluster is 10 vehicles.

[0084] Edge nodes are responsible for real-time monitoring of scene semantics and judgment of delivery triggers. The cloud uses the marketing cloud platform API interface to realize synchronous scheduling of multiple channels (vehicle system, mobile APP, offline terminal). At the same time, it integrates a 5G-V2X communication monitoring module. When the latency exceeds 20ms, the edge fallback solution is triggered. After the network is restored, the cloud is synchronously verified to ensure that the delivery response latency does not exceed 30ms.

[0085] While executing the campaign, we collect real-time data on the campaign performance across multiple channels and subsequent vehicle-to-everything (V2X) behavioral data. Performance data includes clicks, conversion rates, and user feedback, while behavioral data includes whether users visit recommended locations and whether they use coupons.

[0086] The specific details of the causal closed-loop iteration stage are as follows:

[0087] Using the data on the delivery effect collected during the real-time delivery execution phase and the subsequent behavior data of the Internet of Vehicles as input, a hierarchical counterfactual inference algorithm is introduced. The edge end calculates the difference in causal effect between delivery intervention and natural conversion in real time, and the cloud establishes a global attribution model based on data from multiple edge nodes to determine the causal relationship between the delivery strategy and the conversion result and eliminate external interference factors.

[0088] The hierarchical counterfactual inference algorithm has the following hierarchical dimensions: ① User hierarchy; ② Scenario hierarchy; Calculation logic: Calculate the causal effect difference τ for each layer of data, and use a weighted average to obtain the final τ value;

[0089] The global attribution model uses the Gradient Boosting Tree (XGBoost) model: input features include causal effect difference τ, scene relevance score S, delivery channel, and user tag weight; training data requirements: each round must contain ≥1000 valid delivery records, and the data time span must be ≥7 days; the model evaluation metric is MAE (mean absolute error), and MAE < 0.05 is required;

[0090] The causal attribution results are then used to power the preceding steps: optimizing the federated causal inference parameters in the federated causal fusion stage, adjusting the semantic understanding weights of the demand forecasting model in the demand forecasting modeling stage, and updating the user causal association tags in the user tag building stage. Finally, a full-link causal process of fusion, prediction, delivery, attribution, and iteration is established to solve the problems of misjudgment bias and time difference, and to achieve dynamic upgrade of the entire technical solution.

[0091] Global attribution weight formula:

[0092] ;

[0093] In the formula: Indicates the first The global attribution weight of each marketing campaign strategy is used to quantify the contribution of that strategy to the final consumer conversion.

[0094] Indicates the first The causal effect difference of each marketing strategy is calculated using a hierarchical counterfactual inference algorithm, and is related to the causal effect difference formula. Consistent definition, used to reflect the first The actual impact of individual strategies on user conversion behavior when they act independently, such as the gas station discount strategy in a low-fuel scenario. This refers to the incremental conversion brought about by the strategy.

[0095] Indicates the first The relevance score between a marketing strategy and the user's real-time contextual semantics is calculated based on a four-dimensional contextual semantic graph. For example, the relevance between a congested commuting scenario and an audio entertainment advertising strategy. Higher value, marketing strategies for long-distance driving scenarios and short-distance shopping. The value is lower, used to measure the degree of matching between strategy and scenario;

[0096] Indicates the first term in the summation term The difference in the causal effect of each marketing strategy, and The definitions are completely identical;

[0097] Indicates the first term in the summation term The relevance score between marketing strategies and real-time user context semantics, and The definitions are completely identical;

[0098] This indicates the total number of independent strategies participating in this marketing campaign, corresponding to all strategies involved in simultaneous multi-channel campaigns. For example, if a campaign includes three strategies: gas station discounts (in-car infotainment), in-car music memberships (app), and offline car wash coupons (terminals), then... =3;

[0099] This represents the summation index, used to iterate through all items in this delivery. Each marketing strategy ensures the causal effect difference of each strategy ( ) × Scene relevance score ( All of these are included in the summation calculation of the denominator, ultimately achieving... Normalization, all strategies The sum of them is .

[0100] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cross-platform intelligent analysis and delivery method that integrates vehicle network data and internet marketing data, characterized in that: The specific steps of this method are as follows: Data collection and preprocessing stage: Collect multimodal data from the Internet of Vehicles and multimodal data from Internet marketing, preprocess the collected data, and output the preprocessed data; Data standardization and alignment stage: The preprocessed data is standardized in format and encoding, cross-modal associations are established, conflicting data is removed and missing data is filled, and finally standardized semantically aligned data is output. Federated Causal Fusion Stage: Based on the standardized semantic alignment data, the initial data association is completed; a temporal dynamic causal graph and a dual privacy protection mechanism are integrated into the horizontal federated learning framework to eliminate false correlations and establish a data resource pool; User tag construction phase: Based on the data resource pool, establish a four-level tag system, including basic tags, scenario semantic tags, causal association tags, and demand intent tags; dynamically adjust tag weights and output a dynamic user tag system; Demand forecasting modeling Phase 1: Establish a hybrid model to extract shallow features and enhance user understanding; employ federated distillation and model compression techniques, introduce reinforcement learning algorithms to optimize model parameters, and output demand prediction results. Real-time delivery execution phase: Based on demand forecasting and scenario semantics, delivery decisions are broken down into edge execution and cloud optimization; relying on a three-tier computing power pool architecture, delivery is carried out simultaneously across multiple channels, communication latency is monitored, and delivery performance and behavioral data are collected at the same time; Causal closed-loop iterative stage: Based on the campaign effect and behavioral data, calculate the causal effect and establish a global attribution model; use the attribution results to empower the preceding steps.

2. The cross-platform intelligent analysis and delivery method for integrating vehicle network data and internet marketing data according to claim 1, characterized in that: The specific data acquisition and preprocessing stage is as follows: Multimodal vehicle network data is collected through in-vehicle terminals, including driving trajectory, driving behavior, vehicle condition sensor data and scene context data. Edge nodes filter invalid data in real time and annotate the initial scene semantics. Simultaneously, multimodal data of internet marketing is collected through marketing platform APIs and SDKs, and associated with semantic tags of marketing content; during the collection process, a semantic tag pre-annotation mechanism is embedded to perform preliminary feature extraction on unstructured data, filter out data with no semantic value at edge nodes in advance, and finally output preprocessed data.

3. The cross-platform intelligent analysis and delivery method for integrating vehicle network data and internet marketing data according to claim 2, characterized in that: The data standardization and alignment phase is as follows: Based on the preprocessed data from the data acquisition and preprocessing stage, the vehicle network JSON format data and marketing CSV format data are first uniformly converted into a semantic tag-feature value-time stamp triplet format according to the cross-platform data semantic standard protocol, while standardizing the user virtual identifier and scene semantic coding rules. For vehicle-to-everything (V2X) sensor data, marketing text, and image data, a cross-modal semantic mapping algorithm is used to establish a connection. Then, a semantic consistency verification algorithm is used to eliminate logically conflicting data. The missing data is filled in by semantic association, and finally, standardized semantically aligned data is output.

4. The cross-platform intelligent analysis and delivery method for integrating vehicle network data and internet marketing data according to claim 3, characterized in that: The federal causal fusion phase is as follows: Based on the standardized semantic alignment data from the data standardization and alignment stage, the initial association between vehicle network data and marketing data is first completed using the user's unique virtual identifier as the core and combined with a spatiotemporal semantic matching algorithm. Then, a time-series dynamic causal graph module is integrated into the horizontal federated learning framework, and an adaptive sliding time window mechanism is adopted to automatically reconstruct the causal structure after each round of federated training, thereby capturing the dynamic correlation of time-series data in real time. Meanwhile, the framework incorporates homomorphic encryption and differential privacy dual protection mechanisms to remove spurious correlations; Ultimately, we will uncover the true causal relationships between scene semantics, marketing behavior, and consumption conversion, and establish a unified data resource pool.

5. The cross-platform intelligent analysis and delivery method for integrating vehicle network data and internet marketing data according to claim 4, characterized in that: The user tag construction phase is as follows: Based on the data resource pool of the aforementioned federated causal fusion stage, a four-level system is established: basic tags, scene semantic tags, causal association tags, and demand intent tags. Basic tags are extracted from the basic attribute data of the resource pool; scene semantic tags are generated based on scene data in the resource pool and combined with a four-dimensional scene semantic graph, with each type of scene associated with a corresponding causal attribution rule library; causal association tags are directly derived from the causal inference results in the resource pool; and demand intent tags integrate the features of the first three levels of tags to characterize the user's core needs. Subsequently, a semantic similarity algorithm is used, combined with real-time semantic changes in user scenarios and feedback from marketing behaviors. An improved version of the DQN reinforcement learning algorithm is then used to automatically adjust the tag weights and association strengths, outputting a dynamic user tag system.

6. The cross-platform intelligent analysis and delivery method for integrating vehicle network data and internet marketing data according to claim 5, characterized in that: The demand forecasting modeling phase is detailed as follows: Based on the dynamic user tag system in the user tag construction phase, a lightweight CNN-Transformer hybrid model is established. The CNN module extracts shallow features of the scene, and the Transformer module parses the multimodal relationship between scene semantics and marketing content, thereby enhancing the ability to understand user intent. Federated distillation technology is adopted, combined with dynamic pruning and lightweight model compression technology to reduce model transmission and computation costs. Simultaneously, a reinforcement learning algorithm with real-time conversion efficiency as the reward function is introduced to dynamically optimize model parameters and ultimately output user demand prediction results.

7. The cross-platform intelligent analysis and delivery method for integrating vehicle network data and internet marketing data according to claim 6, characterized in that: The real-time delivery execution phase is as follows: Based on the user demand prediction results in the demand prediction modeling stage, and combined with the real-time scene semantics of the Internet of Vehicles, the marketing task intelligent segmentation algorithm is first used to split the delivery decision into the edge real-time execution segment and the cloud optimization segment. Then, the delivery is executed based on the three-level computing power pool architecture of vehicle terminal, roadside edge node and regional cloud. Edge nodes are responsible for real-time monitoring of scene semantics and judgment of delivery triggers. The cloud enables multi-channel synchronous scheduling through the marketing cloud platform API interface. At the same time, it integrates a 5G-V2X communication monitoring module to collect multi-channel delivery effect data and vehicle network subsequent behavior data in real time while executing delivery.

8. The cross-platform intelligent analysis and delivery method for integrating vehicle network data and internet marketing data according to claim 7, characterized in that: The specific causal closed-loop iteration stage is as follows: Using the data on the delivery effect collected during the real-time delivery execution phase and the subsequent behavior data of the Internet of Vehicles as input, a hierarchical counterfactual inference algorithm is introduced. The edge end calculates the difference in causal effect between delivery intervention and natural conversion in real time, and the cloud establishes a global attribution model based on data from multiple edge nodes to determine the causal relationship between the delivery strategy and the conversion result and eliminate external interference factors. The causal attribution results are then used to power the preceding steps: optimizing the federal causal inference parameters in the federal causal fusion stage, adjusting the semantic understanding weights of the demand forecasting model in the demand forecasting modeling stage, and updating the user causal association tags in the user tag building stage, ultimately establishing a full-link causal process of fusion, prediction, delivery, attribution, and iteration.

Images