Advertisement operation data analysis system and method based on multi-dimensional data fusion analysis

Abstract

The invention relates to the technical field of advertisement data analysis, and discloses an advertisement operation data analysis system and method based on multi-dimensional data fusion analysis, and the system comprises a data fusion module which is used for carrying out the collection, privacy protection preprocessing and standardization of multi-platform heterogeneous data; the space-time atlas construction module is used for constructing a dynamic heterogeneous graph which takes the user as a node and takes the cross-platform behavior path as an edge, and training to obtain a user behavior space-time feature association atlas and an advertisement effect prediction model; the strategy optimization module is used for dynamically distributing a multi-target weight based on the prediction result, generating an adjustment strategy and synchronizing the adjustment strategy to each advertisement platform; and the federal enhancement module is used for generating encrypted enhanced data through a generative adversarial network, and mixing the encrypted enhanced data with real data in a self-adaptive proportion to solve the problem of cold start data sparseness. The method corresponds to the system. According to the method, safe fusion, accurate prediction, real-time optimization and efficient cold start of the advertisement data are realized.

Description

Technical Field

[0001] This application relates to the field of advertising data analysis technology, specifically an advertising operation data analysis system and method based on multi-dimensional data fusion analysis. Background Technology

[0002] In advertising operations, advertisers need to manage campaigns on multiple heterogeneous advertising platforms, such as search engines, social media, and e-commerce platforms. The significant differences in data protocols and independent user identification systems across these platforms make it difficult to securely link and integrate cross-platform user behavior data, resulting in data silos. Existing technical solutions, when dealing with such scenarios, lack the ability to effectively model the spatiotemporal dynamic dependencies of user behavior across platforms and cannot address the sparsity of training data during the cold start phase while protecting user privacy. This leads to insufficient accuracy in advertising performance prediction models, delayed response due to reliance on offline analysis for strategy adjustments, and difficulty in achieving technical synergy in advertising operation data analysis regarding data fusion completeness, model prediction accuracy, and real-time strategy execution. This has become a core technical bottleneck restricting the improvement of the automation level of advertising technology systems. Summary of the Invention

[0003] Therefore, it is necessary to provide an advertising operation data analysis system and method based on multi-dimensional data fusion analysis that can improve the accuracy, reliability, and timeliness of strategy execution of advertising operation data analysis, addressing the aforementioned technical issues.

[0004] On the one hand, this application provides an advertising operation data analysis system based on multi-dimensional data fusion analysis, including: The data fusion module is configured to collect user behavior data and ad consumption data from multiple heterogeneous advertising platforms, perform storage operations and privacy protection preprocessing on the user behavior data and ad consumption data, and generate standardized advertising operation data. The spatiotemporal graph construction module is configured to construct a dynamic heterogeneous graph based on the standardized advertising operation data, using user device identifiers as nodes and cross-platform user behavior jump paths as directed edges. The user spatiotemporal coordinates, advertising exposure sequences, and conversion event identifiers are used as temporal features of nodes and edges and input into the spatiotemporal graph neural network for training. The module outputs a spatiotemporal feature association graph of user behavior and constructs an advertising effect prediction model based on the spatiotemporal feature association graph of user behavior and multi-objective advertising effect indicators. The strategy optimization module is configured to dynamically allocate the weights of click-through rate, conversion rate and return on investment based on the output of the advertising effect prediction model, generate an advertising placement adjustment strategy, and synchronize the advertising placement adjustment strategy to the multiple heterogeneous advertising platforms through an event-driven mechanism based on the Raft consensus algorithm. The federated enhancement module is configured to enhance local user behavior data through a generative adversarial network under a federated learning framework, encrypt the enhanced data to generate encrypted enhanced data, and mix the encrypted enhanced data with local real data in an adaptive ratio for federated training of the advertising effect prediction model under the condition of sparse training data.

[0005] In one embodiment, the data fusion module includes: The API adaptation unit is configured to perform protocol conversion and routing of API requests from different advertising platforms based on the rule engine, and output standardized advertising data streams. A hybrid storage unit is configured to write streaming data from the standardized advertising data stream to a time-series database and archive historical data to a distributed file system; The privacy protection unit is configured to use a private record association scheme based on elliptic curve cryptography to encrypt and associate cross-platform user identifiers, and to use a differential privacy federated averaging algorithm to add noise to the local model gradient.

[0006] In one embodiment, writing streaming data from the standardized advertising data stream to a time-series database and archiving historical data to a distributed file system includes: High-throughput writes are performed on user behavior event streams using a time-series database; Historical data from the time-series database is sharded and archived to a distributed file system according to a preset period. A unified query interface is used to perform correlation queries between real-time data in the time-series database and historical data in the distributed file system.

[0007] In one embodiment, the spatiotemporal map construction module includes: The graph neural network unit is configured to construct a dynamic heterogeneous graph with user device identifiers as nodes and cross-platform user behavior jump paths as directed edges, and encode user spatiotemporal coordinates, ad exposure sequences and conversion event identifiers as temporal features and input them into the spatiotemporal graph neural network. The multi-objective optimization unit is configured with click-through rate, conversion rate and return on investment as optimization objectives, and uses a Bayesian optimization method based on joint entropy search to dynamically adjust the weights of each objective.

[0008] In one embodiment, the nodes of the dynamic heterogeneous graph are user device identifiers, the edges are cross-platform user behavior jump paths, and the temporal features include spatiotemporal coordinates composed of user geographic location and timestamps, a time-ordered sequence of advertising exposure events, and conversion event identifiers.

[0009] In one embodiment, the strategy optimization module includes: The weight calculation unit is configured to calculate the influence weight of user behavior spatiotemporal features on advertising effectiveness based on the user behavior spatiotemporal feature association graph, and generate a spatiotemporal weight matrix. The strategy generation unit is configured to use a Bayesian optimization method to search for the hyperparameters of the reinforcement learning model and generate an advertising bidding strategy based on the optimized reinforcement learning model.

[0010] In one embodiment, the strategy optimization module further includes: The event scheduling unit is configured to reach a consensus on the advertising placement adjustment strategy based on the Raft consensus algorithm, and distribute the consensus-based strategy event to each advertising platform through the stream processing platform.

[0011] In one embodiment, the step of enhancing local user behavior data using a generative adversarial network, encrypting the enhanced data to generate encrypted enhanced data, and mixing the encrypted enhanced data with local real data in an adaptive ratio includes: Conditional generative adversarial networks are used to enhance local user behavior data. The conditional inputs of the conditional generative adversarial network include advertising creative feature vectors and user spatiotemporal context encoding. The enhanced data is encrypted using homomorphic encryption technology, and the secure exchange of the encrypted enhanced data is achieved based on a secure multi-party computation protocol. The encrypted enhanced data obtained from the exchange is mixed with the local real data in a proportion calculated based on the sparsity of the local data.

[0012] In one embodiment, a decision support module is also included, configured as follows: The click-through rate, conversion rate, and return on investment output by the advertising effectiveness prediction model are visualized. Based on the visualization results and the analysis results of the strategy optimization module, parameter adjustment suggestions are generated, including the time period interval, user targeting feature tag set, or budget allocation ratio.

[0013] On the other hand, this application provides an advertising operation data analysis method based on multi-dimensional data fusion analysis, applied to the advertising operation data analysis system based on multi-dimensional data fusion analysis as described above, including the following steps: Collect user behavior data and ad consumption data from multiple heterogeneous advertising platforms, store and preprocess the user behavior data and ad consumption data for privacy protection, and generate standardized advertising operation data; Under the federated learning framework, local user behavior data is enhanced by generative adversarial networks, the enhanced data is encrypted to generate encrypted enhanced data, and the encrypted enhanced data is mixed with local real data in an adaptive ratio to generate hybrid training data. The hybrid training data is used to perform federated training on the advertising effect prediction model under the condition of sparse training data. Based on the standardized advertising operation data, a dynamic heterogeneous graph is constructed with user device identifiers as nodes and cross-platform user behavior jump paths as directed edges. User spatiotemporal coordinates, advertising exposure sequences, and conversion event identifiers are used as temporal features of nodes and edges and input into the spatiotemporal graph neural network for training. The output is a user behavior spatiotemporal feature association graph. Based on the user behavior spatiotemporal feature association graph and multi-objective advertising effect indicators, the advertising effect prediction model is constructed using the mixed training data. Based on the output of the advertising performance prediction model, the weights of click-through rate, conversion rate, and return on investment are dynamically allocated to generate an advertising placement adjustment strategy. The advertising placement adjustment strategy is then synchronized to the multiple heterogeneous advertising platforms through an event-driven mechanism based on the Raft consensus algorithm.

[0014] The aforementioned advertising operation data analysis system and method based on multi-dimensional data fusion analysis eliminates cross-platform data silos and ensures secure association of user identifiers by standardizing and preprocessing multi-source heterogeneous advertising data to protect privacy. A dynamic heterogeneous graph is constructed using user device identifiers as nodes and cross-platform behavior jump paths as directed edges. User spatiotemporal coordinates, advertising exposure sequences, and conversion events are encoded as temporal features and input into a spatiotemporal graph neural network to accurately depict the spatiotemporal dynamic evolution of user behavior, improving the training quality and generalization ability of the advertising effect prediction model. An advertising placement adjustment strategy is generated based on multi-objective dynamic weight allocation, and the strategy is synchronized in real time and executed consistently across multiple platforms using the Raft consensus algorithm and event-driven mechanism. Encrypted and enhanced data is generated using a conditional generative adversarial network within a federated learning framework and mixed with real data according to an adaptive computation ratio based on local data sparsity, effectively improving the convergence stability of model training during the cold start phase. The aforementioned technical features form an organically synergistic technical closed loop, which, while ensuring user privacy and data security, systematically solves key technical challenges such as cross-platform advertising data fusion, accurate spatiotemporal modeling of user behavior, real-time strategy optimization, and cold start data sparsity. This improves the overall technical performance of the advertising operation data analysis system in terms of data processing integrity, model prediction accuracy, strategy execution timeliness, and system operational robustness. Attached Figure Description

[0015] Figure 1 A structural block diagram of an advertising operation data analysis system based on multi-dimensional data fusion analysis provided in this application embodiment; Figure 2 A flowchart illustrating the advertising operation data analysis method based on multi-dimensional data fusion analysis provided in this application embodiment. Detailed Implementation

[0016] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0017] Current advertising operations have entered a stage of full-domain collaboration. Advertisers generally need to manage campaigns on heterogeneous advertising platforms such as search engines, social media, and e-commerce platforms simultaneously. Each platform uses independent API protocol standards (e.g., Google Ads uses RESTful JSON, Meta Ads uses GraphQL, and Taobao Union uses a private binary protocol), user identification systems (e.g., device IDs, cookies, and platform UIDs are incompatible), and data storage architectures (time-series databases and relational databases are used interchangeably). This results in highly fragmented user behavior data across platforms. The behavioral trajectory of the same user in social media exposure, search platform clicks, and e-commerce platform conversions is stored separately on different platforms. Moreover, due to the constraints of laws such as the Personal Information Protection Law and the Data Security Law, the original user identifiers cannot be exchanged in plaintext. Traditional data fusion solutions based on rule mapping or third-party ID matching not only generally have low correlation accuracy but also pose a risk of user privacy leakage. Given this weak data foundation, existing advertising performance prediction models mostly employ static feature modeling methods such as logistic regression and XGBoost, simplifying user behavior into isolated event vectors. This makes it difficult to depict the spatiotemporal sequence dependencies and dynamic evolution of cross-platform behavior transitions, such as "user exposure on short video platforms → keyword search the next day → conversion on e-commerce platforms the following day," leading to significant biases in the model's identification of users' true intentions. Especially in cold-start scenarios such as launching new advertising campaigns or expanding new user groups, the available sample size for a single advertising platform client is often less than 1000 (a typical industry threshold). While existing federated learning frameworks can protect gradient privacy, they lack high-quality synthetic data generation and secure exchange mechanisms, failing to effectively alleviate the local data sparsity problem. Model training is prone to overfitting or convergence stagnation. Furthermore, strategy optimization often relies on T+1 offline analysis to generate adjustment plans, and strategy distribution depends on independent API calls from each platform, lacking a distributed consistency guarantee mechanism. When multi-platform strategies need coordinated adjustment (such as budget reallocation triggering cross-platform bidding linkage), network latency or platform response differences often lead to strategy execution conflicts, budget overruns, or traffic mismatches.

[0018] While the industry has attempted to introduce technologies such as graph neural networks to model user relationships, Bayesian optimization to adjust multi-objective weights, and generative adversarial networks to enhance data, these solutions are often applied in isolation to single stages: graph neural networks fail to integrate spatiotemporal dynamic features with advertising business sequences; federated learning and data augmentation lack privacy-secure cross-client exchange mechanisms; and policy generation and distribution do not form a closed loop of "analysis-decision-execution." This technological fragmentation makes it difficult for advertising operation systems to coordinate and optimize data fusion completeness, behavioral modeling accuracy, cold start adaptability, and policy execution real-time performance. Ultimately, this restricts the substantial improvement of the automation level and operational efficiency of advertising decisions in complex multi-platform environments. There is an urgent need for a technical solution that can connect the entire chain of data fusion, spatiotemporal modeling, privacy enhancement, and real-time optimization to systematically solve the coupled technical problems of data silos, inaccurate modeling, cold start difficulties, and policy lag in cross-platform advertising operations.

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Furthermore, it should be emphasized that this system complies with relevant laws and regulations such as the Personal Information Protection Law and the Data Security Law throughout the entire process of data collection, processing, and exchange. All user behavior data is acquired based on the user's explicit authorization and the developer agreements of each advertising platform. Privacy-preserving preprocessing technology aims to further enhance the security of data use and protect user privacy rights.

[0020] Firstly, this embodiment provides an advertising operation data analysis system based on multi-dimensional data fusion analysis, such as... Figure 1 As shown, the system includes: The data fusion module is configured to collect user behavior data and ad consumption data from multiple heterogeneous advertising platforms, perform storage operations and privacy protection preprocessing on the user behavior data and ad consumption data, and generate standardized advertising operation data. The spatiotemporal graph construction module is configured to construct a dynamic heterogeneous graph based on standardized advertising operation data, using user device identifiers as nodes and cross-platform user behavior jump paths as directed edges. User spatiotemporal coordinates, advertising exposure sequences, and conversion event identifiers are used as temporal features of nodes and edges to be input into the spatiotemporal graph neural network for training. The output is a spatiotemporal feature association graph of user behavior, and an advertising effect prediction model is constructed based on the spatiotemporal feature association graph of user behavior and multi-objective advertising effect indicators. The strategy optimization module is configured to dynamically allocate the weights of click-through rate, conversion rate, and return on investment based on the output of the advertising performance prediction model, generate an advertising placement adjustment strategy, and synchronize the advertising placement adjustment strategy to multiple heterogeneous advertising platforms through an event-driven mechanism based on the Raft consensus algorithm. The federated enhancement module is configured to enhance local user behavior data through generative adversarial networks within a federated learning framework, encrypt the enhanced data to generate encrypted enhanced data, and mix the encrypted enhanced data with local real data in an adaptive ratio for federated training of the advertising effect prediction model under sparse training data conditions.

[0021] The data fusion module uses a scheduling engine to call the open API interfaces of various advertising platforms at preset time intervals to collect raw data streams. After protocol parsing, the raw data streams are stored in a memory buffer. The storage operation adopts a layered architecture, with real-time streaming data written to a time-series database and historical archived data stored in a distributed file system. In the privacy protection preprocessing, cross-platform user identifier association uses a private record association scheme based on elliptic curve cryptography to achieve encrypted matching, and the model training gradient processing uses a differential privacy federated averaging algorithm to add Laplace noise.

[0022] The data fusion module is implemented as follows: Deployed on a cloud server cluster, it runs through a set of horizontally scalable data collection agent processes. Each agent process is configured to use an open application programming interface (API) with one or more specific advertising platforms (such as Google Ads, Meta Ads, and ByteDance's Pangle platform). It initiates HTTPS requests according to a preset collection period (e.g., every 5 minutes) or listens for real-time callbacks from the platform via Webhook. The raw data formats collected include, but are not limited to, JSON (JavaScript Object Notation, RESTful API), Protocol Buffers (gRPC interface), or platform-specific binary formats. The module includes a data buffer queue (specifically implemented using Apache Kafka 3.5.0, with three times the number of topic partitions as the platform). All collection agents serialize the raw log data (user behavior data: impressions, clicks, dwell time; ad consumption data: cost per thousand impressions, cost per thousand impressions) into Avro format and push it to a designated Kafka topic, such as a topic named standardized-ad-logs. Storage operations refer to the following: downstream stream processing jobs (using Apache Flink 1.17.1) consume data from the Kafka topic in real time, performing data cleaning (removing illegal characters and filling null values), formatting (e.g., standardizing timestamps to Unix millisecond timestamps, currency to USD, and mapping event types to internal enumeration values), and parsing business fields. Privacy-preserving preprocessing is completed within the stream processing job by calling a separate privacy computing service. Finally, the processed, standardized advertising operation data is written to the real-time data area (hosted by Taosi Data 3.0.4.1 time-series database, categorized by <advertiser ID, date> tags) and the historical archive area (hosted by Apache Hadoop HDFS 3.3.4, with data stored as Parquet columnar files in the / ads_data / daily / {yyyy-MM-dd} / directory structure).

[0023] In the spatiotemporal graph construction module, the dynamic heterogeneous graph construction and training process is as follows: (1) Graph Construction: Traverse the user behavior event sequence in the standardized advertising operation data. When the same virtual user identifier (a globally unique virtual user identifier (VUID) generated by the privacy protection preprocessing module) is detected on the platform... Generate behavioral events (e.g., after exposure) within a preset time window Within 1800 seconds (the time limit is 1800 seconds), on the platform Generate behavioral events (For example, when clicking or converting), a line will be created in the graph based on the platform. The corresponding node is the starting point and the platform. Directed edges with the corresponding node as the endpoint. Nodes are created based on VUIDs, and are reused if they already exist. The weight of the edge is initialized to 1 and accumulates with the frequency of events on the same jump path.

[0024] (2) Node feature encoding: Each node Temporal feature vectors It consists of three parts: .in, The user equipment identifier embedding function is an embedding layer that will... (i.e., the hash value of the VUID) is mapped to a 64-dimensional dense vector; This is a spatiotemporal coordinate normalization function; ,in The latitude and longitude of the geographical location at the time the behavior occurred (obtained by IP resolution or GPS). Uses a Unix timestamp. This function performs operations on latitude and longitude based on preset geographic boundaries. Min-Max normalization for timestamps Periodic features such as hour and day of the week are extracted and sine and cosine encoded respectively, and finally a 31-dimensional normalized vector is output. Encoding function for ad exposure sequence. It is a sequence of ad impression events for this user within the last 7 days, arranged in reverse chronological order. This function encodes the variable-length sequence into a 128-dimensional fixed-length vector using a single-layer GRU network (128 hidden units). Therefore, .

[0025] (3) Edge feature encoding: Each directed edge (from node) To the node Temporal feature vectors Defined as: .in, This is a time interval normalization function; The unit is seconds. This function will... A linear mapping to the interval [0,1] is given by the following formula: ,in Seconds (1 day). Encode functions for platform types. For binary pairs This function concatenates the source and target platform identifiers and then performs One-Hot encoding. Assuming the system supports 10 platforms, the output dimension is 20 (considering directionality). This is a function for converting event identifiers. This is a binary variable. It is 1 if a conversion event (such as a purchase) occurs on the target platform during the redirection process, and 0 otherwise. The function directly outputs this value.

[0026] (4) Spatiotemporal graph neural network structure: A two-layer stacked structure of "spatial graph convolution-temporal convolution" is adopted. The input is the node feature matrix. ( (Number of nodes) and edge feature matrix ( Let be the number of sides. (For edge feature dimensions). The first spatial graph convolution uses the GraphSAGE mean aggregator to aggregate neighbor node information and connected edge features: ,in For learnable weight matrix, The ReLU activation function is used. Subsequently, a temporal convolutional layer applies a one-dimensional causal convolutional kernel (size 3, stride 1) to the feature sequences of each node across consecutive time slices to capture temporal evolution patterns. The second layer has the same structure, but the output dimension is adjusted to 128 dimensions. The network ultimately outputs a 128-dimensional embedding vector for all nodes. These vectors together constitute a spatiotemporal feature correlation map of user behavior.

[0027] (5) Model Training: The network uses the prediction of the platform the user is most likely to visit next (multi-class classification task) and the probability of conversion during that jump (binary classification task) as joint training objectives. The loss function is defined as follows: ,in For cross-entropy loss, The loss is a binary cross-entropy loss. The adaptive moment estimation optimizer (Adam) is trained for 200 epochs on the constructed training graph dataset using a learning rate of 0.001.

[0028] (6) Construction of advertising effect prediction model: using the node embedding vectors output by the spatiotemporal graph neural network trained above. As the core input feature, it is concatenated with current real-time context features (such as a 15-dimensional vector encoded from ad placement type, current hour, and ad creative category) to form a 143-dimensional fused feature vector. This vector is input into a three-layer fully connected network (dimensions 128, 64, and 32 respectively), and finally mapped to three scalars through an output layer. The first two layers use the ReLU activation function. In the output layer, the predicted values ​​of click-through rate (CTR) and conversion rate (CVR) are constrained to (0,1) using the Sigmoid activation function, while the predicted value of return on investment (ROI) uses a linear activation function. This prediction model is trained in a supervised manner using historical ad placement records and their corresponding actual performance metrics (CTR, CVR, ROI), with the mean squared error as the loss function.

[0029] The strategy optimization module is implemented through dynamic weight allocation, strategy generation, and synchronous execution. Dynamic weight allocation is achieved using a lightweight neural network (called the weight network): this network uses the output vector of the advertising effectiveness prediction model for the current round. Current advertising campaign budget burn rate (Consumed / Total Budget), and a market competition index obtained from an external API. The input is passed through two fully connected layers (dimension 3→16→16), and then through a Softmax layer, which outputs three normalized real-time weight coefficients. ,satisfy The generation of advertising adjustment strategies is handled by a reinforcement learning agent based on the Deep Deterministic Policy Gradient (DDPG) algorithm. Its state... This includes the aforementioned spatiotemporal weight matrix (weight distribution after aggregation by region-time period), real-time account status (budget, balance, consumption rate of each channel), and market competition characteristics. Actions Defined as a bid adjustment factor for different preset audience packages or ad placements, typically ranging from [0.5, 2.0]. Rewards Calculate based on the actual advertising effect within a certain period of time (e.g., 1 hour) after the strategy is implemented: ,in This indicates an improvement compared to the baseline strategy. This represents the cost penalty coefficient. The event-driven mechanism based on the Raft consensus algorithm refers to the following: The system deploys a Raft consensus cluster consisting of 3 nodes, each running a policy scheduling service. When a reinforcement learning agent generates a new policy (e.g., {"action":"bid_adjustment","target":"audience_123","multiplier":1.2}), it submits it as a log entry to the current Raft leader node. The leader replicates this entry to all followers, and after receiving successful confirmation from more than half of the nodes, the entry is committed. Subsequently, the leader packages the committed policy entry into an event message in a specific format (such as JSON) and publishes it to a highly available Apache Kafka topic (e.g., a topic named ad-strategy-commands). The executor gateway services of each advertising platform subscribe to this topic as consumers, receive the event within milliseconds of latency, parse it, and convert it into a call to the corresponding platform API (e.g., calling the AdjustBid service of the Google Ads API), completing the final distribution and execution of the policy.

[0030] The workflow of the federated enhancement module within the federated learning framework is as follows: Each advertising platform client participating in federated learning (referred to as a participant) runs this module independently. First, Generative Adversarial Network (GAN) enhancement processing: Each participant trains a conditional generative adversarial network locally. Generator The input is a random noise vector Advertising creative feature vector (From a pre-trained creative embedding model), and the user's current spatiotemporal context encoding. (Taken from the encoded vector of the user's most recent behavior). The generator attempts to generate realistic user behavior feature vectors. Discriminator Then it attempts to distinguish real data With generated data Secondly, data encryption: Participants use the Paillier homomorphic encryption algorithm to generate their own public-private key pairs. For each piece of locally generated augmented data Use your own public key Encrypt the data to obtain the encrypted data. Then, secure exchange and aggregation: the participants use a secure multi-party computation (MPC) protocol (specifically employing a combination of offline Beaver Triple preparation and online secret-sharing computation) to aggregate the feature vectors of all encrypted enhanced data in ciphertext state using a weighted average. The aggregator (which can be a coordinating server or a rotating client) only operates on the ciphertext to obtain the aggregated ciphertext result. And broadcast it to all participants. Each participant uses their own private key. After decryption, what is obtained is the aggregated, enhanced data features, not the original data from other parties. Finally, adaptive proportional mixing: each participant... Calculate an adaptive mixing ratio based on local data. The formula for calculating the adaptive mixing ratio is: in, As a participant The number of local real samples. This represents the average real sample size for all participants in the current federal training round. As a participant Local data distribution (Estimated by the mean and covariance of its eigenvectors) and the global data distribution The Kullback-Leibler divergence between the parties (estimated by the coordination server based on statistics reported by each party). This represents the current number of federal training rounds. The coefficient is an adjustable positive value, and in this embodiment it is empirically set to... , , . For a minimal constant (e.g.) ), to prevent division by zero. For a sigmoid function, the result is mapped to the (0,1) interval as the mixing ratio. (Participants) Random sampling from the decrypted aggregated enhanced data One sample, and its local The real samples are mixed to form the final training dataset, which is used to update the local advertising effectiveness prediction model.

[0031] Based on the above, the data fusion module addresses the core issues of cross-platform data acquisition, standardization, and privacy security. The spatiotemporal graph construction module achieves high-fidelity modeling of the spatiotemporal dynamic dependencies of user behavior across platforms through finely defined dynamic heterogeneous graphs and deep spatiotemporal graph neural networks. The policy optimization module combines prediction results with multi-objective dynamic optimization and reinforcement learning, and achieves intelligent policy generation and strong consistency with low-latency cross-platform synchronization through Raft consensus and event flow mechanisms. The federated enhancement module creatively combines conditional generative adversarial networks, homomorphic encryption, secure multi-party computation, and an innovative adaptive hybrid ratio formula, effectively mitigating cold start and data sparsity issues while strictly protecting privacy. These four modules are clearly defined and closely integrated, forming a complete, closed-loop technical solution from secure data fusion, deep spatiotemporal modeling, intelligent dynamic optimization to privacy-enhancing training.

[0032] The refined implementation of data fusion faces technical challenges, including how to efficiently access heterogeneous interfaces, how to reasonably store data with different timeliness, and how to embed privacy protection in each stage of data processing. Therefore, in one embodiment, the data fusion module includes: The API adaptation unit is configured to perform protocol conversion and routing of API requests from different advertising platforms based on the rule engine, and output standardized advertising data streams. The hybrid storage unit is configured to write streaming data from a standardized advertising data stream to a time-series database and archive historical data to a distributed file system. The privacy protection unit is configured to use a private record association scheme based on elliptic curve cryptography to encrypt and associate cross-platform user identifiers, and to use a differential privacy federated averaging algorithm to add noise to the local model gradient.

[0033] The API adaptation unit is implemented as follows: This unit maintains a centralized "Application Programming Interface (API) rule configuration library," defining rules in YAML format. Each rule targets a specific advertising platform API endpoint. The rules include the following key components: 1) Request Builder: Defines the Hypertext Transfer Protocol (HTTP) method, Uniform Resource Locator (URL) template, authentication method, and request header mapping. 2) Parameter Mapper: Maps the system's internally unified query parameters to the specific parameter names and formats required by the target API. 3) Response Parser: Defines extraction rules based on JSON paths or XPath to extract target data rows and fields from the raw response. 4) Exception Handler: Defines retry strategies, circuit breaker conditions, and failure alerts. During unit runtime, all rules are loaded into memory. When data from a specific platform needs to be retrieved, the rule engine dynamically executes the above steps based on the platform identifier and matching rules, converting the heterogeneous API response into an internally unified standardized advertising data stream format and immediately sending it to the specified message queue topic.

[0034] The hybrid storage unit is implemented as follows: It comprises two core sub-services: a real-time write service and a scheduled archiving service. The real-time write service is a resident stream processing job that subscribes to standardized advertising data stream topics. The job performs lightweight extraction, transformation, and loading of each record before writing it through the time-series database connector. In the time-series database, an independent sub-table is created for each data collection source, efficiently managed using its super-table model. The scheduled archiving service is a daily job scheduled by a workflow scheduler. Every morning, this job starts, executing a structured query language to retrieve all data from the previous day from the time-series database. It performs hash sharding based on advertiser identifiers, converts the data in each shard into a columnar storage format, and uploads it in parallel to the distributed file system path. Upon successful archiving, it triggers the time-series database to delete expired data.

[0035] The specific implementation of the privacy protection unit includes two independent algorithms: First, a private record association scheme based on elliptic curve cryptography: Assume there are two participants, A and B. Both parties pre-agree to use the same elliptic curve parameters and base points. For each user's original identifier locally, each party first calculates its secure hash. Then, each party calculates the points on the elliptic curve. The two sides exchanged the calculated point sets. When associating points, the pairing property of bilinear elliptic curves can be used to determine whether two points come from different sets. and Does it meet the requirements? If they are equal, it means That is, the original identifiers are the same, thus achieving encrypted association without exchanging identifiers in plaintext.

[0036] Second, the differential privacy federated averaging algorithm: In each round of training of federated learning, each client... After training the advertising effectiveness prediction model locally, the gradient vector of the model parameters is obtained. First, gradient clipping is performed: the L2 norm of the gradient vector is calculated. ,if Then let .in, This is a preset gradient clipping threshold used to control the maximum impact (sensitivity) of a single client gradient; for example, it can be set to... Then, add Gaussian noise: generate a signal similar to... Same-dimensional random noise vector ,in It is a noise multiplier whose value is related to the target differential privacy budget. It is the identity matrix. The gradient after adding noise... Uploaded to a central server for aggregation (i.e., the federated average algorithm).

[0037] Based on the above, the API adaptation unit, through a rules engine, enables flexible and unified access to diverse and highly dynamic advertising platform interfaces, shielding the complexity of underlying protocols and ensuring the stability and scalability of data collection. The hybrid storage unit, through a combination of time-series databases and distributed file systems, constructs a hierarchical storage architecture for hot, warm, and cold data. This meets the demands of real-time advertising performance analysis for high-throughput, low-latency writing and querying, while also satisfying the needs for long-term storage of historical data and large-scale offline analysis through low-cost archiving, thus optimizing overall storage costs and efficiency. The privacy protection unit, by combining elliptic curve cryptography-based identifier association with a differential privacy federated averaging algorithm, constructs dual privacy protection at both the data and model training levels. The former ensures the security of the cross-platform user identity association process, preventing the leakage of plaintext identifiers; the latter, by adding controllable noise to the model gradient, effectively defends against privacy attacks that infer original data from shared gradients. These two layers of protection together provide a solid and compliant data security foundation for subsequent cross-platform deep analysis and federated learning, serving as key technical support for breaking down data silos without violating privacy regulations.

[0038] To achieve efficient storage and unified access to massive amounts of advertising data, in one embodiment, streaming data from standardized advertising data streams is written to a time-series database, and historical data is archived to a distributed file system, including: High-throughput writes are performed on user behavior event streams using a time-series database; Historical data from the time-series database is sharded and archived to a distributed file system according to a preset period. A unified query interface is used to perform correlation queries between real-time data in the time-series database and historical data in the distributed file system.

[0039] Specifically, for high-throughput writes to user behavior event streams via a time-series database: this step is implemented by the real-time write service of the hybrid storage unit. After consuming data from the message queue, the job utilizes the asynchronous write interface of the time-series database to accumulate a certain number of records in memory or, after a certain time interval, assembles them into a batch for submission. The storage engine of the time-series database, optimized for time-series data, can increase the throughput of such batch writes to over 100,000 records per second while maintaining millisecond-level query latency. During writing, each record must contain a timestamp field, which serves as the first part of the primary key, ensuring the ordered storage of data in time sequence.

[0040] For archiving historical data shards from the time-series database to the distributed file system at a preset period: This step is completed by an archiving job scheduled and executed by the workflow scheduler. The preset period is daily. The main steps of the archiving job are: 1) Data extraction: Using the Java database connection connector of the time-series database, execute structured query language queries. 2) Data sharding: To improve parallel processing efficiency, the query result set is divided in memory according to the hash value of the advertiser identifier. 3) Format conversion and writing: Startup The system consists of several parallel subtasks, each responsible for converting a data shard into a columnar storage file format. After conversion, the file is uploaded to the corresponding directory via the application programming interface of the distributed file system.

[0041] For performing join queries between real-time data in the time-series database and historical data in the distributed file system via a unified query interface: the system constructs a federated query engine. Two directories are configured in the engine: one pointing to the time-series database cluster, and the other pointing to the distributed file system. When a user or upper-layer application submits a Structured Query Language (SCL) query, the engine's query optimizer parses the time conditions in the SCL. For recent time conditions, the query is pushed down to the time-series database directory for execution; for distant time conditions, the query is pushed down to the distributed file system directory for computation. The federated query engine's coordinating node is responsible for merging, sorting, or aggregating the intermediate results returned from the two data sources, ultimately returning a complete result set to the client.

[0042] Based on the above, the specific operational steps of hybrid storage achieve efficient, economical, and transparent management of massive advertising data throughout its entire lifecycle. The high-throughput write step fully leverages the storage engine features designed for time-series data in time-series databases, solving the technical challenge of extremely high instantaneous throughput of event data in advertising scenarios and ensuring the data freshness required for real-time analysis. The periodic sharding and archiving step reduces the total cost of ownership for long-term data storage by migrating cold data from high-performance but costly time-series databases to a more cost-effective distributed file system and employing an efficient columnar compression format, while preserving data analyzability. The unified relational query interface, through query federation technology, logically provides the application layer with a single view for seamless access to real-time and historical full-volume data, completely hiding the complex differences between the underlying multi-storage systems. This combined operation enables the system to balance the low-latency requirements of real-time analysis with the low-cost needs of historical data mining, simplifies the development complexity of upper-layer business logic, and improves the operational efficiency and resource utilization of the entire data stack.

[0043] To overcome the shortcomings of traditional methods that treat user behavior as isolated events and fail to capture cross-platform spatiotemporal dependencies, in one embodiment, the spatiotemporal graph construction module includes: The graph neural network unit is configured to construct a dynamic heterogeneous graph with user device identifiers as nodes and cross-platform user behavior jump paths as directed edges, and encode user spatiotemporal coordinates, ad exposure sequences and conversion event identifiers as temporal features and input them into the spatiotemporal graph neural network. The multi-objective optimization unit is configured with click-through rate, conversion rate and return on investment as optimization objectives, and uses a Bayesian optimization method based on joint entropy search to dynamically adjust the weights of each objective.

[0044] In practice, the construction and encoding details of the graph neural network units are implemented as described above.

[0045] The specific implementation steps of the Bayesian optimization method based on joint entropy search in the multi-objective optimization unit are as follows: (1) Problem definition: Let the parameters of the advertising strategy to be optimized be as follows: ,in For parameter dimensions. The optimization goal is to maximize three metrics: click-through rate. Conversion rate Return on investment This is a multi-objective optimization problem aimed at finding the Pareto optimal solution set. .

[0046] (2) Gaussian process modeling: for each objective function Establish an independent Gaussian process prior for each: .in, It is the mean function (usually set as a constant). It is the covariance function (in this embodiment, the Marton 5 / 2 kernel is used). These are hyperparameters of the kernel function, here. and Both belong to the category of optimized input space For any two points in the matrix, the kernel function Used to calculate the distance between these two points in the objective function. The covariance on the dataset represents their similarity. Historical observation datasets are used. To fit these Gaussian processes, This represents the number of observation points.

[0047] (3) Monte Carlo sampling approximates joint entropy: The core of joint entropy search is to select the option that maximizes mutual information. point Directly calculating this mutual information is difficult. This unit uses Monte Carlo sampling for approximation: a. Sampling the Pareto front: Extracting from the posterior of the current Gaussian process... Groups (e.g.) Sample of objective function values These samples cover the input space A predefined set Representative points For each group of samples An approximate Pareto optimal solution set is extracted from it using the fast non-dominated sorting algorithm. .

[0048] b. Monte Carlo estimation of conditional entropy: for a candidate point Its target value In a given And assuming the Pareto front is The entropy of the distribution under the condition is denoted as The Monte Carlo estimate of mutual information is then: in, Unobserved The prediction entropy at time can be directly calculated from the posterior covariance matrix of the Gaussian process.

[0049] (4) Optimization and iteration: Use a quasi-Newton method with boundary constraints (such as L-BFGS-B) to find the maximum acquisition function. candidate points .Will The corresponding advertising strategy was tested online using A / B testing to collect real data. The observed values ​​form new data points, which are then added to the historical observation dataset. In the middle. Then, using the updated... Refit all Gaussian process models. Repeat steps (3) and (4) until the preset number of iterations (e.g., 50 rounds) is reached. The final output is a set of Pareto-optimal policy parameters.

[0050] Based on the above, the complex user behavior modeling problem and business objective optimization problem were professionally decomposed and collaboratively solved. The graph neural network unit is specifically responsible for extracting deep, structured spatiotemporal feature associations from the fused data. Its output feature map is the cornerstone of all subsequent accurate predictions and optimization decisions, solving the problem that traditional feature engineering methods struggle to characterize complex behavioral relationships. The multi-objective optimization unit addresses the inherent challenge of simultaneously balancing multiple competitive indicators such as click-through rate, conversion rate, and return on investment in advertising. Using a Bayesian optimization method based on joint entropy search, compared to traditional weighted summation or simple Pareto search, this method guides the search by maximizing the information gain of the Pareto optimal set. This allows for a more efficient and comprehensive exploration of the Pareto front balancing multiple objectives in a high-dimensional parameter space, avoiding getting trapped in local optima, and thus systematically improving the scientific nature and overall benefit of multi-objective decision-making. The establishment of these two units endows the system with both powerful representation learning capabilities and intelligent multi-objective decision-making capabilities.

[0051] To ensure that the constructed graph accurately reflects the essence of cross-platform behavior, in one embodiment, the nodes of the dynamic heterogeneous graph are user device identifiers, the edges are cross-platform user behavior jump paths, and the temporal features include spatiotemporal coordinates composed of user geographic location and timestamp, a time-sorted sequence of advertising exposure events, and conversion event identifiers.

[0052] In practical implementation, for nodes acting as user equipment identifiers: Here, the user equipment identifier specifically refers to the virtual user identifier processed by the privacy protection unit. Its generation process is as follows: For a user, their original identifier id_a (e.g., IDFA:XYZ) on platform A and their original identifier id_b (e.g., phone number hash ABC) on platform B are securely compared using a private record association scheme based on elliptic curve cryptography. If the algorithm determines that id_a and id_b belong to the same user, the system generates or assigns a globally unique, opaque string as the user's virtual user identifier, such as vu_9f86d081. Subsequently, all behavioral data about this user in the system is associated using this virtual user identifier. This ensures the uniqueness and privacy security of the nodes.

[0053] For edges representing cross-platform user behavior navigation paths: this definition clarifies the direction and semantics of the edge. A directed edge This indicates a single transaction from the user (corresponding node). On the platform Behavioral events to the same user (corresponding node) On the platform Behavioral events The jump. The creation of the edge needs to meet strict spatiotemporal proximity conditions: 1) the virtual user identifiers of the source event and the target event are the same; 2) the two events occur on different advertising platforms ( ); 3) Timestamp of the target event Must be in the source event timestamp After that, and the time difference Less than the preset threshold (In this embodiment) (seconds). The type of the edge is determined by the quadruple. The only certainty.

[0054] Regarding the specific composition of time-series features: (1) Spatiotemporal coordinates: For a single user action, its geographic location information is expressed in latitude and longitude. Indicates. Timestamp This is a Unix millisecond timestamp. During encoding, latitude and longitude are normalized using minimum and maximum values ​​respectively. , Similarly, among them and This refers to predefined geographical boundaries, such as the latitude and longitude range covering the target countries for business operations. (Regarding timestamps...) Extract its periodic characteristics: hourly component Encoded as Day of the week components Encoded as Ultimately, the spatiotemporal coordinate feature is a 6-dimensional vector: .

[0055] (2) Ad exposure sequence: For a user node, collect its exposure time window before the current time point. (In this embodiment) All advertising exposure events that occurred within a day, arranged in descending order of occurrence time, to obtain a sequence. ,in It is the unique identifier of the advertising creative within the system. The sequence length is given. The sequence is processed through an embedding layer (mapping ad_id to a 32-dimensional vector) and a single-layer gated recurrent unit network (with 64 hidden units). The final hidden state of the gated recurrent unit is used as the encoding vector of the sequence.

[0056] (3) Conversion Event Identifier: This is a label associated with a specific behavior jump (edge). In a single instance from the platform... To the platform In the redirection behavior, if on the target platform If a predefined conversion event occurs (such as "purchase successful" or "form submitted"), the flag is 1; otherwise, it is 0. The conversion event flag is a scalar and is used directly as a dimension in edge feature encoding.

[0057] Based on the above, a clear, consistent, and information-rich input is provided for subsequent graph neural network modeling, which is a prerequisite for accurate modeling. Defining nodes as privacy-preserving virtual user identifiers ensures the uniqueness of user entities and cross-platform relevance while fundamentally eliminating the risk of leakage of users' original sensitive information. Defining directed edges as cross-platform behavioral jump paths that satisfy spatiotemporal proximity allows the constructed graph to realistically and objectively reflect the flow of user interests and intentions within the digital advertising ecosystem, capturing key causal and sequential relationships. The explicit definition of temporal features (spatiotemporal coordinates, exposure sequences, conversion identifiers) ensures that the features input into the model comprehensively cover multiple key dimensions of user behavior, including space, time, historical contact, and final feedback. This clear and complete graph structure definition enables spatiotemporal graph neural networks to perform effective message passing and feature learning, thereby outputting high-quality feature association graphs and laying a solid and reliable data foundation for final performance prediction.

[0058] To achieve the transformation from prediction to decision-making, in one embodiment, the strategy optimization module includes: The weight calculation unit is configured to calculate the influence weight of user behavior spatiotemporal features on advertising effectiveness based on the spatiotemporal feature association graph of user behavior, and generate a spatiotemporal weight matrix. The strategy generation unit is configured to use a Bayesian optimization method to search for the hyperparameters of the reinforcement learning model and generate an advertising bidding strategy based on the optimized reinforcement learning model.

[0059] In specific implementation, the weight calculation unit is implemented as follows: the input of this unit is the "user behavior spatiotemporal feature association map" output by the spatiotemporal map construction module, which is specifically represented as a 128-dimensional embedding vector set of all active user nodes. ,in This is the set of currently active users. The calculation steps are as follows: (1) Node importance score: Define a learnable query vector This vector is optimized along with the entire system during model training. For each user node... Calculate the dot product of the node with the query vector and normalize it over the entire user set using the Softmax function to obtain the initial importance score for that node: , .

[0060] (2) Spatiotemporal grouping and aggregation: The system predefines a grouping and aggregation mechanism. Geographical area (such as city business district code) and A spatiotemporal grid consisting of time periods (e.g., dividing a day into six 4-hour periods). Each user... Mapped to the corresponding spatiotemporal grid based on the normalized spatiotemporal coordinates of its most recent behavior record. Among them , Then, the aggregate weight value of the grid is calculated. The arithmetic mean of the importance scores of all user nodes within the group: .

[0061] (3) Matrix generation: weight values ​​of all spatiotemporal grids By region index and time period index Arranged in order to form a Two-dimensional matrix This is the spatiotemporal weight matrix. This matrix visually demonstrates which combinations of regions and time periods currently have a higher concentration of potentially high-value users.

[0062] The implementation of the strategy generation unit is divided into two stages: First, the hyperparameter search stage: This stage uses Bayesian optimization to find the optimal combination of hyperparameters for the subsequent reinforcement learning model. The hyperparameter space is defined. It includes key parameters of the near-end policy optimization algorithm: learning rate. (Logarithmic scale), discount factor Coefficients used for policy entropy regularization The Bayesian optimizer maintains a Gaussian process model to fit the hyperparameter configuration. Performance of the corresponding model The black-box functional relationship between (the average cumulative reward evaluated in the offline simulation environment) and (the other two) is described. In each iteration, the Bayesian optimizer selects a promising new configuration based on the Gaussian process model and the desired improvement of the acquisition function. System usage Initialize a proximal policy optimization agent, train it for a fixed number of steps (e.g., 10,000 steps) in an advertising simulation environment built based on historical log replay, and evaluate its performance. .Will Add the observation set and update the Gaussian process model. Repeat this process. Wheel (such as) After (round), select performance from historical observations. Top configuration As the final result.

[0063] Second, the strategy execution phase: using the optimal hyperparameters obtained through the search. Initialize a local policy optimization agent. This agent interacts with the real-time ad delivery environment. Its state... It is a high-dimensional vector, including: the current spatiotemporal weight matrix Flattened Dimensional vectors, real-time budget consumption rates for each advertising campaign, market average cost per thousand impressions for each ad placement, and unique-hot encoding for the current time period (e.g., hour). Actions It is A continuous vector of dimensions, where each dimension corresponds to a controllable advertising leverage, such as segmenting different users. Bid adjustment multiplier .award Calculated at the end of each decision cycle (e.g., 15 minutes). The agent continuously updates its policy network through ongoing interaction with the environment. Ultimately, it learns to generate optimal ad bidding strategies (i.e., actions) based on real-time status. ).

[0064] Based on the above, the weight calculation unit creatively transforms the abstract graph neural network embedding vector into a concrete and interpretable importance distribution (spatiotemporal weight matrix) in the spatiotemporal dimension. This process white-boxes the black-box output of the deep learning model, providing clear geographical and temporal guidance for decision-making, making the allocation of advertising budgets and attention resources more targeted. The policy generation unit, through a meticulously designed two-stage process, solves the common problems of hyperparameter tuning difficulties and unstable convergence in reinforcement learning applications. First, Bayesian optimization is used to efficiently search for the optimal hyperparameter combination, providing the reinforcement learning agent with a high-performance innate configuration; then, the configured agent performs online learning and decision-making in a real environment. This combination of offline parameter optimization and online policy learning improves the feasibility of applying reinforcement learning in complex advertising bidding environments and the effectiveness of the final strategy, thereby generating more refined and adaptive intelligent bidding strategies that adapt to market dynamics.

[0065] To ensure that the optimization strategy can be executed consistently and reliably in a distributed, multi-platform environment, in one embodiment, the strategy optimization module further includes: The event scheduling unit is configured to reach a consensus on the advertising placement adjustment strategy based on the Raft consensus algorithm, and distribute the consensus-based strategy events to various advertising platforms through the stream processing platform.

[0066] The specific implementation of the event scheduling unit includes a consensus layer and a distribution layer: The consensus layer based on the Raft consensus algorithm: The unit is deployed as an independent microservice cluster containing multiple nodes. Each node runs the same service process, which embeds a Raft consensus algorithm library (e.g., using the Raft library of etcd). Nodes communicate with each other through the internal network and elect a leader, with the rest being followers. When the policy generation unit generates a new advertising adjustment policy, it sends this policy as the payload of a log entry to the leader node of the current Raft cluster via a remote procedure call. Upon receiving it, the leader first appends this log entry to its local log, and then replicates the entry in parallel to all follower nodes via the AppendEntries remote procedure call. As long as more than half of the nodes (including the leader itself) successfully persist the log entry to stable storage, the leader marks the log entry as committed. Once committed, it means that the policy has achieved strong consistency consensus throughout the distributed system. The leader then notifies all clients that the log has been committed.

[0067] The distribution layer of the stream processing platform is decoupled from the consensus layer and listens for Raft log commit events. After a log entry is committed, the leader node immediately publishes its content (i.e., the new ad delivery adjustment strategy) as a message event to a designated topic in a highly available message queue cluster. This topic is configured with multiple partitions and high replicas to ensure high throughput and disaster recovery. Each ad platform's executor gateway, acting as an independent consumer group, subscribes to this public topic. Each gateway service establishes a long-lived connection to the message queue cluster upon startup. When a new event arrives in the message queue topic, all subscribed gateways can pull the event almost simultaneously. Each gateway parses the strategy S in the event, translates it into an application programming interface (API) call format understandable by the ad platform it is responsible for, and executes it. This publish-subscribe pattern and event-stream-based architecture ensures that strategy changes can be broadcast atomically and executed reliably in parallel.

[0068] Based on the above, in complex multi-platform, multi-node deployment environments, simple command issuance is prone to policy loss, duplicate execution, or inconsistent execution states across different platforms due to network latency, node failures, and other reasons, leading to budget overruns or deployment conflicts. Using the Raft consensus algorithm, the event scheduling unit transforms each policy issuance into a distributed consistent transaction. Only after more than half of the nodes confirm persistence is the policy considered valid and committed. This ensures that even with partial node failures, the entire system maintains a consistent and immutable understanding of the historical state of executed policies, providing strong consistency guarantees. Subsequently, event broadcasting via a highly reliable message queue enables asynchronous, decoupled, and high-throughput distribution of policy commands. Gateways on each platform consume events in parallel, achieving rapid synchronous execution of policies. This collaborative design, where consensus ensures consistency and message queues ensure reliable distribution, provides a solid and reliable execution foundation for the implementation of upper-layer intelligent optimization strategies.

[0069] The core issue of the cold start problem in federated learning lies in how to securely augment the local data of each participant while protecting privacy. Therefore, in one embodiment, a generative adversarial network (GAN) is used to enhance local user behavior data, the enhanced data is then encrypted to generate encrypted enhanced data, and the encrypted enhanced data is mixed with local real data in an adaptive ratio, including: Conditional generative adversarial networks (CGANs) are used to enhance local user behavior data. The conditional inputs of the CGANs include ad creative feature vectors and user spatiotemporal context encodings. The enhanced data is encrypted using homomorphic encryption technology, and the secure exchange of the encrypted enhanced data is achieved based on a secure multi-party computation protocol. The encrypted enhanced data obtained from the exchange is mixed with the local real data in a proportion calculated based on the sparsity of the local data.

[0070] In practice, as described above, each participating party The adaptive mixing ratio is calculated based on the formula. Then, from the shared aggregated augmentation dataset... In, random sampling One sample. Finally, this... Encryption enhancement data and local The real data points are merged to form a hybrid dataset.

[0071] Based on the above, conditional generative adversarial networks (GANs) are used for augmentation to ensure that the generated synthetic data not only resembles the distribution of real data but also is associated with specific advertising creatives and user contexts. This results in high-quality, highly relevant augmented samples, improving the effectiveness of data augmentation. Homomorphic encryption is used to encrypt the augmented data, and a secure multi-party computation protocol is combined to achieve secure aggregation and exchange in encrypted form, constructing a trusted channel for data value circulation. Each participant contributes encrypted data value (augmented data), and ultimately receives the aggregated value, without disclosing any party's original data or independently generated data. This achieves data usability without visibility, strictly meeting the privacy compliance requirements for cross-organizational data collaboration. Finally, an adaptive mixing ratio mechanism based on local data sparsity, distribution differences, and training progress intelligently determines how much external augmented data each participant needs to introduce in each round of training, dynamically balancing the role of local real data and external augmented data. This avoids overfitting or underfitting, enabling the federated model to converge more robustly and faster during the cold start phase. These three interconnected steps achieve a good balance between privacy protection and data utility.

[0072] To enable advertising operators to intuitively understand the system's analysis results and take quick action, in one embodiment, the system also includes a decision support module, configured as follows: The click-through rate, conversion rate, and return on investment output by the advertising performance prediction model are visualized. Based on the visualization results and the analysis results of the strategy optimization module, parameter adjustment suggestions are generated, including the time period interval, user targeting feature tag set, or budget allocation ratio.

[0073] The decision support module provides a web-based visualization console. Visualizations are implemented using chart libraries such as ECharts or AntV, and include: 1) Spatiotemporal heatmap: Combining a geographic map with a time selector, using color depth to show the click-through rate or conversion rate of different regions at different times. 2) Multi-metric trend comparison chart: Displaying the dynamic trends of click-through rate, conversion rate, ROI, and cost of goods sold on the same time axis in line chart format, with drill-down support for detailed data on specific fluctuation periods. 3) User segmentation effect radar chart: Using radar charts to show the performance differences of different user groups (such as "high-value potential users" and "price-sensitive users") based on the strategy optimization module's segmentation, across multiple core metrics.

[0074] The parameter adjustment suggestions are generated through a rule-based reasoning process. The system has a built-in expert rule base, where each rule defines the recommendations to be given when a specific visualization pattern appears. For example, Rule 1: If "in the spatiotemporal heatmap, region A's click-through rate consistently exceeds the average by 20% during the evening hours (8:00 PM - 10:00 PM), but the conversion rate remains flat," then it is recommended to "optimize ad creatives placed in region A during the evening hours to improve their conversion-driving capabilities." Rule 2: If "multi-metric trend charts show that the ROI of an advertising campaign has been declining continuously over the past 3 days, while the cost per click has been steadily increasing," then it is recommended to "consider tightening the user targeting tag set for this campaign, or reallocate a portion of the campaign's budget (e.g., 20%) to other advertising campaigns with more stable ROI performance." Rule 3: If "the user segmentation radar chart shows that the 'highly active App users' group has a high click-through rate but a low conversion rate," then it is recommended to "examine the landing page experience for ads targeting this group, or create a separate advertising campaign for this group that places greater emphasis on conversion-driving." These suggestions will be presented in both structured text (such as JSON format) and natural language descriptions in the "Optimization Suggestions" panel of the visualization console.

[0075] Based on the above, the decision support module transforms the predicted data, weight matrices, and strategy instructions generated by complex backend algorithms into intuitive charts and concrete action suggestions on the front end. This visualization lowers the technical barrier, enabling operations personnel to quickly grasp the overall situation and key issues. Rule-based intelligent suggestion generation directly guides data analysis results to actionable operational actions, shortening the decision-making chain from insight to action and improving the operational efficiency of human-machine collaboration.

[0076] On the other hand, this embodiment provides an advertising operation data analysis method based on multi-dimensional data fusion analysis, which is applied to the advertising operation data analysis system based on multi-dimensional data fusion analysis as described above, such as... Figure 2 As shown, the method includes the following steps in sequence: Step S1: Collect user behavior data and ad consumption data from multiple heterogeneous advertising platforms, store and preprocess the user behavior data and ad consumption data for privacy protection, and generate standardized advertising operation data; Step S2: Under the federated learning framework, the local user behavior data is augmented by a generative adversarial network. The augmented data is then encrypted to generate encrypted augmented data. The encrypted augmented data is then mixed with the local real data in an adaptive ratio to generate mixed training data. The mixed training data is used to perform federated training on the advertising effect prediction model under the condition of sparse training data. Step S3: Based on standardized advertising operation data, construct a dynamic heterogeneous graph with user device identifiers as nodes and cross-platform user behavior jump paths as directed edges. Input user spatiotemporal coordinates, advertising exposure sequences, and conversion event identifiers as temporal features of nodes and edges into the spatiotemporal graph neural network for training. Output a spatiotemporal feature association graph of user behavior. Based on the spatiotemporal feature association graph of user behavior and multi-objective advertising effect indicators, construct an advertising effect prediction model using mixed training data. Step S4: Based on the output of the advertising performance prediction model, dynamically allocate the weights of click-through rate, conversion rate, and return on investment to generate an advertising placement adjustment strategy. Then, synchronize the advertising placement adjustment strategy to multiple heterogeneous advertising platforms through an event-driven mechanism based on the Raft consensus algorithm.

[0077] Specifically, step S1 aims to address the heterogeneity and privacy security issues at the data source. A data collection agent deployed on a cloud server calls the open application programming interfaces (APIs) of various advertising platforms at preset intervals to obtain raw log data. After serializing the raw data, it is pushed to a message queue to buffer high-concurrency traffic. A stream processing job consumes the data from the message queue, performing data cleaning (e.g., filtering outliers, completing missing fields) and format standardization (unifying timestamps and currency units). Privacy-preserving preprocessing is performed: a privacy computing service is called, using a private record association scheme based on elliptic curve cryptography to encrypt and match the original user identifiers from different sources, generating globally unique virtual user identifiers. For subsequent model training using federated learning, the logic of a differential privacy federated averaging algorithm is embedded in the local training step, including gradient pruning (using a preset pruning threshold). ) and adding Gaussian distribution The processed, standardized data is written to a time-series database in real time and archived to a distributed file system on a daily basis, forming a complete standardized advertising operation data warehouse.

[0078] Step S2 aims to specifically alleviate the data sparsity problem during the model's cold start phase through privacy-preserving data augmentation. On the advertising platform client (as a participant in federated learning) that needs to perform model cold start training, the user behavior data processed in step S1 is locally loaded. A conditional generative adversarial network is trained, whose generator is fed with random noise. Advertising creative feature vector and user spatiotemporal context encoding Given the conditions, generate synthetic behavior feature vectors. The Paillier homomorphic encryption algorithm is used to encrypt the enhanced data generated by each party. Each participant securely exchanges and aggregates all encrypted enhanced data in ciphertext state through a secure multi-party computation protocol, ensuring that all parties ultimately obtain the same aggregated enhanced dataset. However, they cannot obtain the original data generated by other parties. Each participating party bases its decisions on its own actual data volume. Global average data volume KL divergence between local and global data distributions and the current training round The proportion is calculated based on the aforementioned adaptive mixing ratio formula. Subsequently, from Random sampling Enhanced data, with local The real data are mixed to form a hybrid training dataset for local model updates.

[0079] Step S3 aims to achieve deep and dynamic modeling of complex cross-platform user behaviors, thereby improving prediction accuracy. Based on the standardized advertising operation data generated in step S1, a dynamic heterogeneous graph is constructed using virtual user identifiers as nodes and cross-platform behavior jump paths as directed edges. The temporal features of the nodes and edges in the graph are encoded using the aforementioned method: Node Features Edge features The constructed graph data is input into a spatiotemporal graph neural network consisting of stacked spatial graph convolutional layers (using the GraphSAGE mean aggregator) and temporal convolutional layers (one-dimensional causal convolutions) for training. Using the aforementioned joint loss function as the objective, it learns high-dimensional embedding representations of nodes and edges. These embeddings collectively constitute a spatiotemporal feature association graph of user behavior. Using the node embedding vectors of this graph as the core input features, and combining them with the hybrid training dataset generated in step S2 (aggregated and updated under a federated averaging framework), a multi-task advertising performance prediction model is trained with click-through rate, conversion rate, and return on investment as output objectives.

[0080] Step S4 aims to transform accurate predictions into executable and consistently synchronized intelligent decisions. Based on the latest output of the advertising performance prediction model, a learnable weight network is used to dynamically calculate the real-time weights of three objectives: click-through rate (CTR), conversion rate, and return on investment (ROI). The aforementioned Bayesian optimization method based on joint entropy search is used to find the optimal hyperparameter combination for the reinforcement learning model (such as the proximal strategy optimization algorithm). Using the optimized reinforcement learning model, with dynamic weights and the current system state (including the spatiotemporal weight matrix, budget consumption, market competition index, etc.) as input, a specific advertising placement adjustment strategy (such as a bid adjustment coefficient vector) is generated. This strategy is submitted as a log entry to the event scheduling unit based on the Raft consensus algorithm. After more than half of the nodes reach consensus, it is published as a message queue event. The executor gateways of each advertising platform consume this event and convert it into calls to their respective platform management application programming interfaces (APIs), achieving cross-platform synchronous updates of the strategy, thus completing the analysis-decision-execution closed loop.

[0081] Based on the above, the technical solution of the aforementioned system is fully reproduced from the perspective of methodological steps. Its technical effects correspond to the system claims, but the sequential logic of technical implementation and data flow are clearly defined in a process-oriented manner. This method systematically covers the entire value chain from raw data governance to intelligent decision execution through the connection of four core steps. Step S1 (data acquisition and fusion) ensures that multi-source heterogeneous data can flow into the system securely, compliantly, and with high quality, laying a reliable data foundation for subsequent analysis. Step S2 (federated data augmentation) innovatively combines federated learning, generative adversarial networks, and cryptography, providing a general method to effectively solve the sparsity of cold-start data under the premise of privacy protection, and improving the robustness of the method in data-constrained scenarios. Step S3 (spatiotemporal graph modeling and prediction) achieves machine understanding of complex user behavior patterns through a standard and in-depth analysis process of graph construction, feature encoding, spatiotemporal training, and predictive modeling, which is the core of generating accurate predictions. Step S4 (strategy optimization and synchronous execution) combines the prediction results with multi-objective optimization, reinforcement learning, and distributed consensus to ensure that the generated optimization strategy is not only intelligent but also consistently and reliably executed across platforms. These four steps are interconnected, forming a self-iteratory and continuously optimizing automated closed loop, fully realizing the technical objectives of improving the accuracy, real-time performance, security, and automation level of advertising operation data analysis.

[0082] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0083] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0084] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An advertising operation data analysis system based on multi-dimensional data fusion analysis, characterized in that, include: The data fusion module is configured to collect user behavior data and ad consumption data from multiple heterogeneous advertising platforms, perform storage operations and privacy protection preprocessing on the user behavior data and ad consumption data, and generate standardized advertising operation data. The spatiotemporal graph construction module is configured to construct a dynamic heterogeneous graph based on the standardized advertising operation data, using user device identifiers as nodes and cross-platform user behavior jump paths as directed edges. The user spatiotemporal coordinates, advertising exposure sequences, and conversion event identifiers are used as temporal features of nodes and edges and input into the spatiotemporal graph neural network for training. The module outputs a spatiotemporal feature association graph of user behavior and constructs an advertising effect prediction model based on the spatiotemporal feature association graph of user behavior and multi-objective advertising effect indicators. The strategy optimization module is configured to dynamically allocate the weights of click-through rate, conversion rate and return on investment based on the output of the advertising effect prediction model, generate an advertising placement adjustment strategy, and synchronize the advertising placement adjustment strategy to the multiple heterogeneous advertising platforms through an event-driven mechanism based on the Raft consensus algorithm. The federated enhancement module is configured to enhance local user behavior data through a generative adversarial network under a federated learning framework, encrypt the enhanced data to generate encrypted enhanced data, and mix the encrypted enhanced data with local real data in an adaptive ratio for federated training of the advertising effect prediction model under the condition of sparse training data.

2. The advertising operation data analysis system based on multi-dimensional data fusion analysis according to claim 1, characterized in that, The data fusion module includes: The API adaptation unit is configured to perform protocol conversion and routing of API requests from different advertising platforms based on the rule engine, and output standardized advertising data streams. A hybrid storage unit is configured to write streaming data from the standardized advertising data stream to a time-series database and archive historical data to a distributed file system; The privacy protection unit is configured to use a private record association scheme based on elliptic curve cryptography to encrypt and associate cross-platform user identifiers, and to use a differential privacy federated averaging algorithm to add noise to the local model gradient.

3. The advertising operation data analysis system based on multi-dimensional data fusion analysis according to claim 2, characterized in that, The step of writing streaming data from the standardized advertising data stream into a time-series database and archiving historical data to a distributed file system includes: High-throughput writes are performed on user behavior event streams using a time-series database; Historical data from the time-series database is sharded and archived to a distributed file system according to a preset period. A unified query interface is used to perform correlation queries between real-time data in the time-series database and historical data in the distributed file system.

4. The advertising operation data analysis system based on multi-dimensional data fusion analysis according to claim 1, characterized in that, The spatiotemporal map construction module includes: The graph neural network unit is configured to construct a dynamic heterogeneous graph with user device identifiers as nodes and cross-platform user behavior jump paths as directed edges, and encode user spatiotemporal coordinates, ad exposure sequences and conversion event identifiers as temporal features and input them into the spatiotemporal graph neural network. The multi-objective optimization unit is configured with click-through rate, conversion rate and return on investment as optimization objectives, and uses a Bayesian optimization method based on joint entropy search to dynamically adjust the weights of each objective.

5. The advertising operation data analysis system based on multi-dimensional data fusion analysis according to claim 4, characterized in that, The nodes of the dynamic heterogeneous graph are user device identifiers, and the edges are cross-platform user behavior jump paths. The temporal features include spatiotemporal coordinates composed of user geographic location and timestamp, a time-sorted sequence of advertising exposure events, and conversion event identifiers.

6. The advertising operation data analysis system based on multi-dimensional data fusion analysis according to claim 1, characterized in that, The strategy optimization module includes: The weight calculation unit is configured to calculate the influence weight of user behavior spatiotemporal features on advertising effectiveness based on the user behavior spatiotemporal feature association graph, and generate a spatiotemporal weight matrix. The strategy generation unit is configured to use a Bayesian optimization method to search for the hyperparameters of the reinforcement learning model and generate an advertising bidding strategy based on the optimized reinforcement learning model.

7. The advertising operation data analysis system based on multi-dimensional data fusion analysis according to claim 6, characterized in that, The strategy optimization module also includes: The event scheduling unit is configured to reach a consensus on the advertising placement adjustment strategy based on the Raft consensus algorithm, and distribute the consensus-based strategy event to each advertising platform through the stream processing platform.

8. The advertising operation data analysis system based on multi-dimensional data fusion analysis according to claim 1, characterized in that, The process of enhancing local user behavior data using a generative adversarial network, encrypting the enhanced data to generate encrypted enhanced data, and then mixing the encrypted enhanced data with local real data in an adaptive ratio includes: Conditional generative adversarial networks are used to enhance local user behavior data. The conditional inputs of the conditional generative adversarial network include advertising creative feature vectors and user spatiotemporal context encoding. The enhanced data is encrypted using homomorphic encryption technology, and the secure exchange of the encrypted enhanced data is achieved based on a secure multi-party computation protocol. The encrypted enhanced data obtained from the exchange is mixed with the local real data in a proportion calculated based on the sparsity of the local data.

9. The advertising operation data analysis system based on multi-dimensional data fusion analysis according to any one of claims 1 to 8, characterized in that, It also includes a decision support module, configured as follows: The click-through rate, conversion rate, and return on investment output by the advertising effectiveness prediction model are visualized. Based on the visualization results and the analysis results of the strategy optimization module, parameter adjustment suggestions are generated, including the time period interval, user targeting feature tag set, or budget allocation ratio.

10. A method for analyzing advertising operation data based on multi-dimensional data fusion analysis, characterized in that, Includes the following steps: Collect user behavior data and ad consumption data from multiple heterogeneous advertising platforms, store and preprocess the user behavior data and ad consumption data for privacy protection, and generate standardized advertising operation data; Under the federated learning framework, local user behavior data is enhanced by generative adversarial networks, the enhanced data is encrypted to generate encrypted enhanced data, and the encrypted enhanced data is mixed with local real data in an adaptive ratio to generate hybrid training data. The hybrid training data is used to perform federated training on the advertising effect prediction model under the condition of sparse training data. Based on the standardized advertising operation data, a dynamic heterogeneous graph is constructed with user device identifiers as nodes and cross-platform user behavior jump paths as directed edges. User spatiotemporal coordinates, advertising exposure sequences, and conversion event identifiers are used as temporal features of nodes and edges and input into the spatiotemporal graph neural network for training. The output is a user behavior spatiotemporal feature association graph. Based on the user behavior spatiotemporal feature association graph and multi-objective advertising effect indicators, the advertising effect prediction model is constructed using the mixed training data. Based on the output of the advertising performance prediction model, the weights of click-through rate, conversion rate, and return on investment are dynamically allocated to generate an advertising placement adjustment strategy. The advertising placement adjustment strategy is then synchronized to the multiple heterogeneous advertising platforms through an event-driven mechanism based on the Raft consensus algorithm.

Images