An advertisement click-through rate prediction system

Abstract

The application discloses an advertisement click rate prediction system, comprising a data acquisition module for real-time acquisition of user behavior data, attribute data, context data and biological characteristic signal data; a multi-modal feature extraction module for extracting multiple feature vectors; a dynamic feature fusion module for dynamically weighting and fusing multi-modal features through an adaptive attention fusion network; a graph neural network coding module for coding a heterogeneous graph structure to generate a graph embedding feature vector; an adaptive routing prediction module for dynamically selecting an expert subnetwork through a router network for prediction; a real-time feedback calibration module for online parameter updating based on real click feedback and generating a calibration signal; and a multi-objective optimization decision module for outputting an advertisement recommendation list based on multi-objective optimization. The application realizes accurate prediction and intelligent decision through adaptive routing and multi-objective optimization, and significantly improves the accuracy and scene adaptability of advertisement click rate prediction.

Description

Technical Field

[0001] This invention relates to the fields of internet advertising and artificial intelligence technology, and in particular to an advertising click-through rate prediction system. Specifically, it relates to an advertising click-through rate prediction system based on multimodal feature fusion and adaptive routing mechanism. Background Technology

[0002] With the rapid development of internet technology and the continued growth of the digital advertising market, click-through rate (CTR) prediction has become one of the most crucial technical aspects of advertising delivery systems. Accurate CTR prediction helps advertising platforms achieve optimal traffic allocation within limited ad display opportunities, improving both the advertiser's ROI and the user experience and platform revenue.

[0003] In existing technologies, ad click-through rate (CTR) prediction methods have primarily evolved from traditional machine learning methods to deep learning methods. Early CTR prediction methods mainly employed models such as Logistic Regression (LR) and Factorization Machine (FM), utilizing user historical click records, ad features, and contextual information to estimate click probability. With the development of deep learning technology, CTR prediction models based on Deep Neural Networks (DNNs) have been widely used, such as DeepFM, xDeepFM, and DCN (Deep & Cross Network), which improve prediction accuracy by explicitly or implicitly modeling feature interactions.

[0004] However, existing technologies still have many technical shortcomings and deficiencies in practical applications. First, current CTR prediction systems mainly rely on users' historical behavioral data (such as click records and browsing duration) for modeling, lacking the ability to perceive users' real-time cognitive state and emotional responses. Users' click decisions are not only influenced by historical behavioral patterns, but also to a large extent by instantaneous physiological and psychological factors such as the degree of attention concentration, emotional arousal level, and visual focus in the current state. Existing technologies cannot capture this dynamic and deep user state information that is difficult to reflect directly through behavioral data, resulting in a significant ceiling effect on prediction accuracy.

[0005] Secondly, existing feature fusion methods suffer from information redundancy and noise interference when handling multi-source heterogeneous features. When a system integrates multiple types of features (such as behavioral features, attribute features, contextual features, and even biological features), complex nonlinear interactions may exist between different features. Simply relying on shared embedding layers or simple concatenation operations cannot effectively distinguish the importance of each modality feature, and may even introduce noisy features that negatively impact the prediction results. Although there has been considerable research on feature interaction modeling, there is still no effective solution to the technical problem of dynamically adjusting the fusion weights of each modality feature based on the user's real-time state.

[0006] Furthermore, existing single-model CTR prediction methods struggle to adapt to diverse advertising scenarios. In real-world advertising environments, different user groups exhibit significantly different behavioral patterns, and different ad types display varying feature distributions. A single model architecture often fails to maintain consistent superior performance across all scenarios. Current multi-model ensemble methods typically employ fixed-weight fusion or simple voting mechanisms, lacking the ability to dynamically select the optimal sub-model based on input features, leading to wasted computational resources and suboptimal prediction results.

[0007] Furthermore, existing CTR prediction systems suffer from update latency and model confusion issues in their feedback learning mechanisms. In real-world advertising environments, user click feedback is often delayed and sparse. Existing methods typically use batch offline training to update the model, failing to utilize the latest user feedback information for real-time prediction calibration. While online learning methods can alleviate this problem to some extent, existing online learning methods exhibit shortcomings such as oscillating model parameter updates and slow convergence speed when handling delayed feedback and label flipping. Summary of the Invention

[0008] In view of the shortcomings of existing technologies, the purpose of this invention is to provide an advertising click-through rate prediction system, which has the following effects:

[0009] The above-mentioned objective of this invention is achieved through the following technical solutions:

[0010] An ad click-through rate prediction system includes:

[0011] The data acquisition module is used to collect user data of at least one user in real time. The user data includes user behavior data, user attribute data, contextual data, and user biometric signal data. The user biometric signal data includes at least one type of biometric signal selected from the following: eye tracking signal, facial micro-expression signal, skin conductance response signal, and electroencephalogram (EEG) signal.

[0012] A multimodal feature extraction module, connected to the data acquisition module, is used to extract features from the user behavior data, user attribute data, and contextual data respectively to obtain behavior feature vectors, attribute feature vectors, and contextual feature vectors, and to perform time-frequency domain joint analysis on the user biometric signal data to extract at least one biometric feature vector.

[0013] A dynamic feature fusion module, connected to the multimodal feature extraction module, is used to input the behavioral feature vector, the attribute feature vector, the context feature vector, and the at least one biometric feature vector into an adaptive attention fusion network. The module calculates the cross-attention weights between the multiple feature vectors through a multi-head self-attention mechanism and generates a fused feature vector based on the cross-attention weights. The adaptive attention fusion network includes a learnable feature weight allocation layer, which dynamically adjusts the fusion weights of the multiple feature vectors according to the real-time biometric signals of different users at the prediction time.

[0014] The graph neural network encoding module, connected to the dynamic feature fusion module, is used to construct a heterogeneous graph structure using the fused feature vector as node features, based on at least one of the following relationships: historical interaction relationship between users and advertisements, social relationship between users, and content association relationship between advertisements. The heterogeneous graph structure is then encoded by a multi-layer graph neural network to obtain a graph embedding feature vector containing structural information.

[0015] An adaptive route prediction module, connected to the graph neural network encoding module, is used to input the graph embedding feature vector into an adaptive route prediction network. The adaptive route prediction network includes multiple parallel expert prediction subnetworks and at least one router network. The router network dynamically selects and activates at least one expert prediction subnetwork based on the real-time changes of the graph embedding feature vector and the user's biometric signal data, and weights and fuses the outputs of the activated expert prediction subnetworks to generate a preliminary click-through rate prediction value.

[0016] The real-time feedback calibration module is connected to the data acquisition module and the adaptive route prediction module, respectively. It is used to generate feedback tags based on the user's subsequent real click behavior data after the current advertisement exposure, and to perform online incremental updates on the decision parameters of the router network of the adaptive route prediction module based on the feedback tags. At the same time, it generates calibration signals to feed back to the dynamic feature fusion module to adjust the learnable feature weight allocation layer in the adaptive attention fusion network.

[0017] A multi-objective optimization decision module, connected to the real-time feedback calibration module, is used to acquire at least two ad candidate sets and, based on the preliminary click-through rate prediction value and the calibration signal output by the real-time feedback calibration module, calculate a comprehensive placement value score for each ad in the at least two ad candidate sets. The comprehensive placement value score is calculated using a multi-objective optimization function based on at least two indicators among predicted click-through rate, expected revenue, and user negative feedback risk. The module then sorts the ads in the at least two ad candidate sets according to the comprehensive placement value score and outputs an ad recommendation list.

[0018] Furthermore, the multimodal feature extraction module includes: a biometric encoding unit, used to preprocess and extract time-frequency domain features from the user's biometric signal data, wherein it extracts gaze duration, gaze frequency, pupil diameter changes, and saccade path features from the eye-tracking signal; extracts micro-expression temporal features from the facial micro-expression signal using a three-dimensional convolutional neural network; extracts time-domain statistical features and frequency-domain power spectrum features of the skin conductance level change curve from the skin conductance response signal; and extracts power spectral density and functional connectivity strength features of multiple brain region frequency bands from the electroencephalogram signal; and a feature alignment unit, connected to the biometric encoding unit, used to align the biometric vector with other feature vectors in the temporal dimension and output the time-aligned multimodal feature tensor.

[0019] Further, the adaptive attention fusion network includes: a feature embedding layer, used to map each input feature vector to an embedding space of the same dimension to obtain an embedded feature matrix; a multi-head self-attention layer, connected to the feature embedding layer, used to perform multi-head self-attention calculation on the embedded feature matrix to generate an attention-weighted feature representation, wherein the multi-head self-attention layer uses a dimensionality reduction calculation mechanism to reduce computational complexity; a feature weight allocation layer, connected to the multi-head self-attention layer, used to dynamically calculate the fusion weight vector of each modality feature based on the biometric state recognition result generated by the user's biometric signal data, wherein the biometric state recognition result is used to characterize the user's current level of attention concentration and emotional arousal; and a feature fusion layer, connected to the feature weight allocation layer, used to perform weighted fusion of the attention-weighted feature representation according to the fusion weight vector to generate a fusion feature vector.

[0020] Furthermore, the graph neural network encoding module includes: a heterogeneous graph construction unit, used to construct a heterogeneous graph structure based on user nodes, ad nodes, and context nodes, wherein the edge types in the heterogeneous graph include multiple types such as user history click edges, user-ad exposure edges, ad content similarity edges, and user social relationship edges; a node feature initialization unit, connected to the heterogeneous graph construction unit, used to use the fused feature vector output by the dynamic feature fusion module as the initial feature vector of the user nodes in the heterogeneous graph; a neighbor aggregation unit, connected to the node feature initialization unit, used to perform type-aware neighbor aggregation operations according to the edge type through each layer of the multi-layer graph neural network, wherein different types of edges use different aggregation functions and attention weights; and a pooling encoding unit, connected to the neighbor aggregation unit, used to output the user node embedding vector output by the last layer of the multi-layer graph neural network as the graph embedding feature vector.

[0021] Further, the adaptive route prediction network includes: a state encoder, used to jointly encode the temporal changes of the graph embedding feature vector and the user biometric signal data to generate a route state vector; a router network, connected to the state encoder, including a gating mechanism network, used to calculate the selection probability of each of the multiple parallel expert prediction sub-networks based on the route state vector, and activate at least one expert prediction sub-network based on the selection probability, wherein a single expert prediction sub-network is activated when the maximum selection probability is greater than a preset threshold, otherwise multiple expert prediction sub-networks are activated; multiple parallel expert prediction sub-networks, each with different network structures and parameter configurations, used to handle different types of prediction scenarios; and an output fusion layer, connected to the router network and the multiple parallel expert prediction sub-networks, used to perform a weighted summation of the outputs of the activated expert prediction sub-networks according to the selection probability to generate a preliminary click-through rate prediction value.

[0022] Furthermore, the real-time feedback calibration module includes: a feedback data acquisition unit, used to acquire real click behavior data of users within a preset time window after the current ad exposure, the real click behavior data including whether a click occurred and the time of the click; an online gradient update unit, connected to the feedback data acquisition unit, used to calculate the loss gradient based on the difference between the real click behavior data and the preliminary click-through rate prediction value, and use the loss gradient to perform online gradient descent update on the parameters of the router network, wherein the online gradient update is performed in real time after each feedback arrives without retraining the entire prediction network; a calibration signal generation unit, connected to the online gradient update unit, used to generate a calibration signal according to the sign and magnitude of the loss gradient, the calibration signal including adjustment direction and adjustment step size indication information for the feature weight allocation layer in the dynamic feature fusion module; and a delayed feedback processing unit, used to cache delayed click feedback data arriving after the preset time window, and perform batch parameter updates after meeting the cache quantity threshold.

[0023] Furthermore, the multi-objective optimization decision module includes: a multi-dimensional indicator calculation unit, used to calculate a predicted click-through rate correction value based on the preliminary click-through rate prediction value and the calibration signal for each candidate advertisement, calculate an expected revenue value based on the advertiser's bidding information and advertisement type, and calculate a user negative feedback risk value based on historical user negative feedback data; a multi-objective optimization unit, connected to the multi-dimensional indicator calculation unit, used to comprehensively calculate the predicted click-through rate correction value, the expected revenue value, and the user negative feedback risk value through a preset multi-objective optimization function to obtain a comprehensive placement value score for each candidate advertisement, wherein the multi-objective optimization function adopts a Pareto optimization strategy, allowing the advertiser to dynamically adjust the weight coefficients of each indicator according to different business objectives; and a sorting output unit, connected to the multi-objective optimization unit, used to sort the candidate advertisements from high to low according to the comprehensive placement value score, select at least one advertisement with the highest score to generate an advertisement recommendation list, and output it to the advertisement placement end.

[0024] Furthermore, the system also includes: a distributed deployment architecture comprising at least one central server and multiple edge computing nodes, wherein the data acquisition module and the multimodal feature extraction module are deployed on the edge computing node side to complete the acquisition and preliminary feature extraction processing of user biometric signals near the user terminal; the dynamic feature fusion module, the graph neural network encoding module, the adaptive route prediction module, the real-time feedback calibration module, and the multi-objective optimization decision module are deployed on the central server side; and a privacy protection unit, deployed on the edge computing node side, for performing localized differential privacy processing on the acquired user biometric signal data, deleting the original biometric signal data after feature extraction, and only encrypting and transmitting the feature vectors to the central server.

[0025] Furthermore, the data acquisition module includes: a behavioral data acquisition unit, used to collect in real time the user's historical click records, browsing duration, search query terms, purchase records, and interaction operation sequences generated on at least one Internet platform; an attribute data acquisition unit, used to obtain the user's age, gender, region, device type, and long-term interest tags; a context data acquisition unit, used to collect in real time the time information, location information, weather information, network status, and page context information of the current advertising display scenario; and a biometric acquisition unit, used to collect the user's biometric signal data through at least one sensing device configured on the user's terminal device, wherein the sensing device includes one or more of a camera, an eye-tracking sensor, a skin conductance sensor, and an electroencephalogram (EEG) signal acquisition device.

[0026] Furthermore, the multiple parallel expert prediction subnetworks include: a first expert subnetwork employing a deep factorization machine architecture for handling click-through rate (CTR) prediction in sparse discrete feature scenarios; a second expert subnetwork employing a deep cross-network architecture for handling CTR prediction in high-order feature interaction scenarios; a third expert subnetwork employing a sequence transformer architecture for handling CTR prediction in user behavior sequence scenarios; and a fourth expert subnetwork employing a multi-gated hybrid expert network architecture for handling CTR prediction in multi-task and multi-scenario scenarios. The router network dynamically selects and activates one or more of the first to fourth expert subnetworks based on the routing state vector, and weights and fuses the outputs of the activated expert subnetworks.

[0027] In summary, compared with the prior art, the present invention has at least one of the following beneficial technical effects:

[0028] This invention introduces user biometric signal data (including eye-tracking signals, facial micro-expression signals, skin conductance signals, and electroencephalogram signals) as input features for the prediction model, achieving quantitative perception of users' instantaneous cognitive state and emotional responses. Compared to existing technologies that rely solely on historical behavioral data, this invention can capture deeper user state information (such as attention concentration, emotional arousal level, and visual focus) that is difficult to reflect directly through behavioral data, significantly improving the accuracy and robustness of click-through rate prediction.

[0029] The dynamic feature fusion module designed in this invention employs an adaptive attention fusion network. It calculates the cross-attention weights between multimodal features through a multi-head self-attention mechanism and dynamically adjusts the fusion weights of each modality based on the user's real-time biometric signals. This technical solution solves the problems of information redundancy and noise interference in multi-source heterogeneous feature fusion, achieving adaptive feature fusion based on the user's real-time state and avoiding the performance loss caused by simple splicing or fixed-weight fusion in existing technologies.

[0030] The adaptive route prediction network designed in this invention comprises multiple expert prediction subnetworks with different network structures and a dynamic router network. Unlike existing single-model methods and fixed-weight ensemble methods, the router network dynamically selects and activates the expert subnetwork most suitable for the current input features based on real-time changes in graph embedding feature vectors and user biometric signals. This achieves dynamic adaptation of the model architecture to the input distribution, resulting in superior prediction performance across different user groups and advertising scenarios, while avoiding waste of computational resources.

[0031] This invention achieves online incremental updates of the prediction model through a real-time feedback calibration module. This module updates the decision parameters of the router network in real-time using gradients based on real click feedback and generates calibration signals that are fed back to the dynamic feature fusion module to adjust the feature fusion weights, thus constructing a closed-loop optimization mechanism of "prediction-feedback-calibration". Compared to existing batch offline update methods, this invention can capture changes in user behavior patterns more promptly, effectively mitigating the problems of delayed feedback and model confusion.

[0032] This invention uses a graph neural network encoding module to model the historical interaction relationship between users and advertisements, the social relationship between users, and the content association relationship between advertisements in a heterogeneous graph structure. It also uses a multi-layer graph neural network for encoding, which makes full use of the structural information in the user-advertisement-context relationship network and makes up for the shortcomings of traditional feature vector representation in expressing the association structure.

[0033] The multi-objective optimization decision module of this invention not only considers click-through rate prediction itself, but also incorporates expected revenue and user negative feedback risk into the comprehensive decision-making. It achieves a dynamic balance between multiple business objectives through Pareto optimization strategy, avoiding suboptimal decisions that may result from optimizing a single indicator.

[0034] This invention, through the design of a distributed deployment architecture and a privacy protection unit, deploys the collection and feature extraction of user biometric signals on the edge computing node side, completing local data processing and privacy protection at the user terminal. Only the de-identified feature vectors are encrypted and transmitted to the central server, which not only ensures the system's real-time processing capability of biometric signals, but also fully protects the user's privacy data security.

[0035] In summary, this invention provides an advertising click-through rate prediction system that integrates multimodal biometric perception, dynamic adaptive feature fusion, graph structure relationship encoding, adaptive route prediction, online feedback calibration, and multi-objective decision optimization. It significantly improves prediction accuracy, scene adaptability, response speed, and privacy protection. Attached Figure Description

[0036] Figure 1 This is a block diagram of the overall structure of the advertising click-through rate prediction system in Embodiment 1 of the present invention.

[0037] Figure 2 This is a block diagram of the internal structure of the multimodal feature extraction module in Embodiment 1 of the present invention.

[0038] Figure 3 This is a block diagram of the internal structure of the adaptive attention fusion network in Embodiment 1 of the present invention.

[0039] Figure 4 This is a block diagram of the internal structure of the neural network encoding module in Embodiment 1 of the present invention.

[0040] Figure 5 This is a block diagram of the internal structure of the adaptive route prediction network in Embodiment 1 of the present invention.

[0041] Figure 6 This is a block diagram of the internal structure of the real-time feedback calibration module in Embodiment 1 of the present invention.

[0042] Figure 7 This is a block diagram of the internal structure of the multi-objective optimization decision module in Embodiment 1 of the present invention.

[0043] Figure 8 This is a schematic diagram of the distributed deployment architecture of the system in Embodiment 1 of the present invention.

[0044] Figure labels: 100, Data acquisition module; 200, Multimodal feature extraction module; 300, Dynamic feature fusion module; 400, Graph neural network encoding module; 500, Adaptive route prediction module; 600, Real-time feedback calibration module; 700, Multi-objective optimization decision-making module; 201, Behavioral feature extraction unit; 202, Attribute feature extraction unit; 203, Contextual feature extraction unit; 204, Biometric feature encoding unit; 205, Feature alignment unit; 301, Feature embedding layer; 302, Multi-head self-attention layer; 30 3. Feature weight allocation layer; 304. Feature fusion layer; 401. Heterogeneous graph construction unit; 402. Node feature initialization unit; 403. Neighbor aggregation unit; 404. Pooling encoding unit; 501. State encoder; 502. Router network; 504. Output fusion layer; 601. Feedback data acquisition unit; 602. Online gradient update unit; 603. Calibration signal generation unit; 604. Delay feedback processing unit; 701. Multidimensional index calculation unit; 702. Multi-objective optimization unit; 703. Sorting output unit. Detailed Implementation

[0045] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0046] It should be noted that the terms "comprising" and "including" and any variations thereof in the specification and claims of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those steps or modules that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0047] Before describing the specific embodiments of the present invention in detail, the core technical concept involved in the present invention will first be generally explained. The core technical concept of the present invention lies in: quantifying the user's real-time cognitive state and emotional response by introducing multimodal user biometric signals (eye tracking, facial micro-expressions, skin conductance response, electroencephalogram signals, etc.), and on this basis, constructing a complete technical closed loop of "multimodal feature perception → dynamic adaptive fusion → graph structure relationship encoding → adaptive route prediction → online feedback calibration → multi-objective optimization decision-making". This technical closed loop covers all aspects from data acquisition to decision output, with each module cooperating and reinforcing each other to jointly achieve high-precision prediction of ad click-through rates and intelligent ad delivery decisions.

[0048] The technical solution of the present invention will be described in detail below through specific embodiments. It should be understood that the following embodiments are only a part of the many possible implementations of the present invention and are not intended to limit the scope of protection of the present invention.

[0049] Example 1

[0050] This embodiment provides an advertising click-through rate prediction system. Please refer to... Figure 1 , Figure 1 The overall structural block diagram of the ad click-through rate prediction system in this embodiment is shown.

[0051] like Figure 1 As shown, the ad click-through rate prediction system provided in this embodiment includes: a data acquisition module 100, a multimodal feature extraction module 200, a dynamic feature fusion module 300, a graph neural network encoding module 400, an adaptive route prediction module 500, a real-time feedback calibration module 600, and a multi-objective optimization decision module 700.

[0052] The connections between the modules are as follows: the output of the data acquisition module 100 is connected to the input of the multimodal feature extraction module 200; the output of the multimodal feature extraction module 200 is connected to the input of the dynamic feature fusion module 300; the output of the dynamic feature fusion module 300 is connected to the input of the graph neural network encoding module 400; the output of the graph neural network encoding module 400 is connected to the input of the adaptive route prediction module 500; the output of the adaptive route prediction module 500 is connected to the first input of the multi-objective optimization decision module 700 and the first input of the real-time feedback calibration module 600, respectively; the second input of the real-time feedback calibration module 600 is connected to the output of the data acquisition module 100; the first output of the real-time feedback calibration module 600 is connected to the calibration input of the adaptive route prediction module 500; the second output of the real-time feedback calibration module 600 is connected to the calibration input of the dynamic feature fusion module 300; the second input of the multi-objective optimization decision module 700 is connected to the third output of the real-time feedback calibration module 600; and the output of the multi-objective optimization decision module 700 outputs the advertising recommendation list.

[0053] The core workflow of this system is as follows: the data acquisition module 100 collects various types of user data in real time; the multimodal feature extraction module 200 extracts features from this data to obtain behavioral feature vectors, attribute feature vectors, contextual feature vectors, and biometric feature vectors; the dynamic feature fusion module 300 fuses these feature vectors into a fused feature vector through an adaptive attention fusion network; the graph neural network encoding module 400 uses the fused feature vector as node features and encodes it on a heterogeneous graph structure to obtain a graph embedding feature vector; the adaptive route prediction module 500 generates a preliminary click-through rate prediction value through an adaptive route prediction network; the real-time feedback calibration module 600 performs online calibration and parameter updates for the system based on real click feedback; and the multi-objective optimization decision module 700 performs comprehensive scoring and ranking of candidate advertisements based on the preliminary prediction value, calibration signal, and multi-objective optimization function, and outputs the final advertisement recommendation list.

[0054] The data acquisition module 100 is used to collect user data from at least one user in real time. In this embodiment, the data acquisition module 100 specifically includes: a behavior data acquisition unit, an attribute data acquisition unit, a context data acquisition unit, and a biometric feature acquisition unit.

[0055] The behavioral data acquisition unit is used to collect in real time the user's historical click records, browsing duration, search query terms, purchase records, and interaction sequences on at least one Internet platform.

[0056] In practical implementation, the behavioral data collection unit can achieve real-time data collection through a Software Development Kit (SDK) embedded in the user's terminal device. For example, when a user browses content on an e-commerce platform or social media platform, the SDK can record the user's click on an advertisement each time, including the timestamp of the click, the identifier of the clicked advertisement, and the page location where the click occurred. Simultaneously, the SDK can also record the user's browsing duration (i.e., the length of time the user stays on a specific page), search query terms (the query text entered by the user in the search box), purchase records (the product information and transaction amount of the user's actual purchase), and interaction sequence (the complete operational trajectory of the user's interaction with page elements, including swiping, clicking, zooming, etc.).

[0057] The data collected by the behavioral data acquisition unit is stored in a local cache in the form of a time series and is periodically or in real-time transmitted to the multimodal feature extraction module 200 for subsequent processing. In actual deployment, in order to balance real-time performance and bandwidth resource utilization, a data batch compression transmission strategy can be adopted, that is, data is transmitted all at once after the data volume reaches a certain threshold. However, for scenarios with high timeliness requirements (such as real-time bidding advertising), a high-priority channel can also be set to achieve real-time data transmission.

[0058] The attribute data acquisition unit is used to obtain the user's age, gender, region, device type, and long-term interest tags.

[0059] This part of the data typically has relatively stable characteristics and does not require real-time collection. In practice, attribute data can be obtained from the basic information users fill in when registering on the platform, or it can be inferred through analysis of user behavior patterns. For example, a user's age and gender information can be obtained directly from the registration information, geographic information can be obtained through IP address location or GPS location data, and device type can be obtained by parsing the user agent string of the user's terminal device. Long-term interest tags can be obtained through cluster analysis of users' historical behavior, such as statistically analyzing the types of advertisements a user clicked over a period of time to identify the main categories the user is interested in (such as electronics, clothing, automobiles, travel, etc.), thereby generating a long-term interest tag vector for the user.

[0060] The context data acquisition unit is used to collect time information, location information, weather information, network status, and page context information of the current advertising display scenario in real time.

[0061] Contextual data is an important situational factor influencing users' click decisions. For example, users' click preferences may differ significantly at different times of day (such as weekday daytime and weekend evening); users' geographical location (such as at home, in the office, or in a shopping mall) will also affect their interest in different types of advertisements; weather information (such as sunny, rainy, and temperature) may also influence users' consumption decision tendencies.

[0062] In practice, time information can be obtained directly from the system clock, location information can be obtained through GPS or base station positioning, weather information can be obtained by calling third-party weather service APIs (such as Hefeng Weather, OpenWeatherMap, etc.) based on the user's current location, network status (WiFi, 4G, 5G, etc.) can be read through the device's network connection manager, and page context information (such as the current page's topic category, existing ad types on the page, etc.) can be obtained through text analysis and classification models of the page content.

[0063] The biometric acquisition unit is one of the core innovations of this invention, used to acquire user biometric signal data through at least one sensing device configured on the user terminal device. The user biometric signal data includes at least one type of biometric signal selected from the following: eye-tracking signals, facial micro-expression signals, skin conductance response signals, and electroencephalogram (EEG) signals.

[0064] It should be noted that the biometric data collection of this invention is conducted with the user's full knowledge and explicit authorization. Before the user uses the biometric data collection function for the first time, the system will display a detailed privacy policy statement, and the relevant sensors will only be activated after obtaining the user's explicit consent.

[0065] The following details the methods for acquiring various biometric signals:

[0066] Eye-tracking signals can be acquired using the front-facing camera of the user's terminal device. Specifically, computer vision technology is used to detect the user's facial feature points (including eye contour points, pupil center points, and corner points of the eyes) in real time. By analyzing the relative positional changes of the pupil and corneal reflection in the image sequence, the user's fixation point position (i.e., the coordinates of the position the user is looking at on the screen), fixation duration (the length of time the user's gaze remains on a certain area), fixation frequency (the number of times the user looks at a certain area per unit time), pupil diameter change (the change in pupil size with light and emotional state), and saccade path (the trajectory of jumps between fixation points) can be calculated.

[0067] In practical implementations, open-source or commercially licensed eye-tracking SDKs can be used (such as mobile eye-tracking solutions provided by Tobii, EyeTech, etc.). For ordinary mobile devices that do not have professional eye-tracking hardware, camera-based uncalibrated eye-tracking algorithms can already achieve a usable level of accuracy. Eye-tracking signals are typically sampled at a frequency of 30 to 60 frames per second, generating a data stream containing information such as timestamps, gaze coordinates, and pupil diameter.

[0068] Facial micro-expression signals are also captured using the front-facing camera of the user's terminal device. Micro-expressions refer to extremely brief (usually lasting 1 / 25 to 1 / 5 of a second) facial muscle movements that can reflect a user's true emotional state. Unlike ordinary expressions, micro-expressions are often unconsciously revealed by the user and are difficult to deliberately fake, thus possessing unique value in reflecting a user's true emotional reactions.

[0069] In its implementation, the system captures a sequence of user facial images from a camera at a frame rate of 30 frames per second or higher. It then uses facial landmark detection algorithms (such as regression tree-based or deep learning-based methods) to locate 68 or more key points on the user's face (including contour points of eyebrows, eyes, nose, mouth, etc.). Finally, it analyzes the displacement changes of these key points between consecutive frames to identify micro-expression types (such as happiness, surprise, disgust, fear, sadness, anger, etc.). The identified micro-expression types and their intensity values ​​are used as part of the biometric signal.

[0070] Electrodermal response (EDS) is a physiological signal that reflects a user's level of emotional arousal by measuring changes in the electrical conductivity of the skin surface. When a user experiences an emotional response, the autonomic nervous system regulates the activity of sweat glands, leading to changes in skin conductivity. EDS has the advantages of high sensitivity and high temporal resolution, enabling it to reflect instantaneous emotional fluctuations in a user.

[0071] In practice, skin conductance signals can be collected using skin conductance sensors on wearable devices such as smartwatches and smart bracelets, or via skin conductance sensors attached to mobile devices (such as sensors integrated into phone cases). These sensors typically consist of two electrodes that measure changes in the conductivity of the skin surface by applying a weak, constant voltage (usually below 0.5V, completely safe for the human body). Because the effective frequency range of skin conductance signals is relatively low (mainly in the 0-2Hz range), a sampling frequency of 20-50Hz is usually sufficient.

[0072] Electroencephalography (EEG) is a physiological signal that collects spontaneous, rhythmic electrical activity of neuronal populations in the cerebral cortex by placing electrodes on the scalp. EEG signals can reflect different levels of cognitive processing in users, including attention levels, cognitive load, and emotional state. Studies have shown that when users see different types of advertisements, the energy of different brain regions' EEG bands (such as theta waves, alpha waves, beta waves, and gamma waves) undergoes characteristic changes, and these changes are significantly correlated with clicking behavior.

[0073] In practice, EEG signal acquisition can be achieved using consumer-grade Brain-Computer Interface (BCI) devices, such as portable EEG acquisition headbands or headphones provided by companies like NeuroSky and Emotiv. These devices typically employ dry electrode technology, allowing users to acquire EEG signals simply by wearing the headband, eliminating the need for conductive gel and making them convenient to use. The EEG acquisition device connects to the user's terminal device via Bluetooth or USB, transmitting the acquired multi-channel EEG data to the terminal for processing in real time. Since the effective frequency band of EEG signals is 0.5-50Hz, the sampling frequency is typically set to 128Hz, 256Hz, or higher.

[0074] In this embodiment, one or more of the above-mentioned biometric acquisition methods can be selectively enabled according to the needs of different types of advertising scenarios. For example, eye tracking and facial micro-expression signals are particularly important in video advertising scenarios; in interactive advertising scenarios, electrodermal response signals can provide richer emotional information; and in high-end user research scenarios, electroencephalogram (EEG) signals can provide the deepest information. Of course, all available signal sources can also be enabled simultaneously to obtain the most complete biometric information.

[0075] The multimodal feature extraction module 200 is connected to the data acquisition module 100 and is used to extract features from user behavior data, user attribute data, and contextual data respectively to obtain behavior feature vectors, attribute feature vectors, and contextual feature vectors. It also performs time-frequency domain joint analysis on the user biometric signal data to extract at least one biometric feature vector.

[0076] Please refer to Figure 2 , Figure 2 A block diagram of the internal structure of the multimodal feature extraction module 200 is shown. For example... Figure 2 As shown, the multimodal feature extraction module 200 includes: a behavior feature extraction unit 201, an attribute feature extraction unit 202, a context feature extraction unit 203, a biometric encoding unit 204, and a feature alignment unit 205.

[0077] The behavior feature extraction unit 201 is used to extract meaningful feature vectors from the raw behavior data collected by the behavior data acquisition unit.

[0078] Specifically, for historical click records, a temporal attention-based encoding method is employed. First, the user's most recent N clicks (e.g., N=100) are arranged chronologically. Each click record contains the ad identifier (ad_id), the timestamp of the click, and the context type of the click. An embedding layer maps the discrete ad identifier and context type into a dense vector. Then, a Long Short-Term Memory (LSTM) network or Transformer model processes this click sequence to extract temporal dependencies, ultimately yielding a behavioral temporal vector representing the user's recent click behavior patterns. The numerical range of each dimension of this vector is typically controlled within the interval [-1,1] or [0,1] to facilitate the numerical stability of subsequent models.

[0079] For browsing duration characteristics, a statistical coding method is employed. For a given ad type or ad location, the average browsing duration, maximum browsing duration, standard deviation of browsing duration, and trend of browsing duration (e.g., slope parameters obtained through linear fitting) are calculated over a past period. These statistical characteristics reflect the degree of user attention invested in different types of ads. For example, a large mean browsing duration indicates that users are generally more interested in this type of ad; a small standard deviation of browsing duration indicates that user attention investment is relatively stable; and an upward trend in browsing duration indicates that user interest in this type of ad is gradually increasing.

[0080] For search query features, a text encoding method is employed. The sequence of recently entered search queries is concatenated into a text paragraph, which is then converted into a fixed-dimensional vector representation using a pre-trained word embedding model (such as Word2Vec, GloVe, or BERT). This vector captures the user's real-time intent. For example, if a user searches for "latest 2025 laptop," the feature vector will contain semantic information related to technology products, and the prediction system will tend to recommend advertisements for electronic products to that user.

[0081] For purchase record features, an attention-based weighted encoding method is employed. Since different purchase records reflect the user's current interests to varying degrees, more recent purchase records typically have higher informational value. Specifically, for each product category in the user's historical purchases, a decay weight is assigned based on the proximity of the purchase time to the current time (e.g., using an exponential decay function with a decay coefficient of 0.9). Then, the weighted product category embedding vectors are summed to obtain the purchase behavior feature vector.

[0082] For the interactive operation sequence features, a graph-based encoding method is employed. User interactions with ads (such as clicking, liking, commenting, saving, sharing, closing, etc.) are modeled as a directed graph structure, where nodes represent user behavior states and edges represent behavior transitions. A Graph Convolutional Network (GCN) is used to learn the behavioral pattern features of users during ad interactions. This method can capture the sequence and dependencies of user behaviors; for example, the chain "browse → click → purchase" will have a specific pattern corresponding to it in the feature vector.

[0083] The aforementioned behavioral feature extraction unit 201 ultimately outputs one or more behavioral feature vectors. The dimensions of these vectors can be set to 128, 256, 512, or other suitable sizes according to the needs of the actual application scenario.

[0084] The attribute feature extraction unit 202 is used to extract attribute feature vectors from the user attribute data obtained by the attribute data acquisition unit.

[0085] For categorical features (such as gender, region, device type, etc.), one-hot encoding or entity embedding is used. One-hot encoding is suitable for features with a small number of categories (such as gender, which only has three categories: male, female, and unwilling to disclose). For features with a large number of categories (such as the names of hundreds of prefecture-level cities in China), entity embedding is used to map each category to a dense vector. The mapping matrix itself is also a parameter that needs to be learned during model training.

[0086] For numerical features (such as age), normalization can be performed as needed. Age usually does not need to be normalized, but age values ​​can be mapped to the [0,1] interval (e.g., divided by 100) to facilitate the training of neural networks.

[0087] For long-term interest tag features, user interest tags are represented as multi-hot vectors. Each dimension of the vector corresponds to a preset interest category (such as technology, sports, fashion, food, etc.). If the user is interested in the category, the value at the corresponding position is 1, otherwise it is 0. Then, a fully connected layer compresses this sparse multi-hot vector into a low-dimensional dense attribute feature vector.

[0088] The results of encoding the various attribute features mentioned above are concatenated or summed to obtain the final user attribute feature vector.

[0089] The context feature extraction unit 203 is used to extract context feature vectors from the context context data obtained by the context data acquisition unit.

[0090] For time information, a periodic encoding method is employed. Because time has a periodicity (for example, 23:00 and 00:00 are numerically far apart, but are actually adjacent in terms of their time periods), directly using the raw time values ​​would result in information loss. Therefore, the following encoding method is used: time t is mapped to (sin(2πt / T), cos(2πt / T)), where T is a period length. For a 24-hour day, T is 24 × 3600 = 86400 seconds; for a 7-day week, T is 604800 seconds. This encoding method preserves the periodicity of time.

[0091] For location information, a geohash encoding method is employed. Latitude and longitude coordinates are converted into a Geohash string, with each Geohash string corresponding to a rectangular geographic region. Then, the Geohash strings are mapped to location feature vectors through an embedding layer. This method preserves the spatial proximity of locations (geohash strings of nearby points share the same prefix) while also handling large amounts of location data.

[0092] For weather information, the weather type (sunny, cloudy, overcast, rainy, snowy, etc.) is encoded as a category feature vector, and the numerical features such as temperature and humidity are normalized and treated as continuous value features. These feature vectors are then concatenated to obtain the weather feature vector.

[0093] For network status information, the network type (WiFi, 4G, 5G, others) is encoded as a category feature vector, and the signal strength is normalized and used as a numerical feature.

[0094] For page context information, a text understanding-based method is used to extract features. The title, description text, keywords, and other information of the current page are input into a lightweight text encoding model (such as DistilBERT), which outputs a fixed-dimensional page semantic feature vector. This vector reflects the topic type of the content the user is currently browsing, which helps predict the user's acceptance of advertisements in the current context.

[0095] The biometric encoding unit 204 is one of the core innovations of this invention. It is used to preprocess and perform joint time-frequency domain analysis on the user's biometric signal data to extract at least one biometric vector.

[0096] The encoding methods for eye-tracking signals, facial micro-expression signals, skin conductance response signals, and electroencephalogram (EEG) signals are described below.

[0097] For eye-tracking signals, the following features are extracted:

[0098] Fixation duration characteristics: This involves calculating the total and average fixation duration of users on each advertising area or screen area within a preset time window. A longer fixation duration generally indicates a higher level of user interest in that area.

[0099] Gaze frequency characteristics: This counts the number of times a user gazes at a specific area within a unit of time. High-frequency gazing may indicate that the user is repeatedly confirming information, or it may be a signal of confusion or hesitation.

[0100] Pupil diameter variation characteristics: Time series of pupil diameter were calculated, and statistics such as mean, variance, skewness, and kurtosis were extracted, as well as the rate of change (first derivative) and acceleration (second derivative) of diameter. Studies have shown a significant correlation between pupil diameter and cognitive load and emotional arousal level.

[0101] Sagling path features: The sequence of gaze points is converted into a set of vectors, recording the start and end positions, saccade amplitude (distance between two gaze points), saccade direction angle (directional angle from the start point to the end point), and saccade speed of each saccade. Sagling patterns can reflect the user's visual search strategy. For example, rapid short-distance saccades usually correspond to local information processing, while long-distance saccades correspond to global scene switching.

[0102] After normalization, the above features are concatenated into an eye-tracking feature vector. The dimension of the eye-tracking feature vector can be selected between 32 and 256 dimensions, depending on the specific application requirements.

[0103] For facial micro-expression signals, a 3D Convolutional Neural Network (3D CNN) is used to extract the temporal features of micro-expressions. Compared with traditional 2D CNNs, 3D CNNs can capture feature changes in both spatial and temporal dimensions simultaneously, making them particularly suitable for processing dynamic signals such as micro-expressions that change drastically in a short period of time.

[0104] In practice, the continuous video frame sequence is first constructed as a three-dimensional tensor (width × height × time). This three-dimensional tensor is then input into a pre-trained 3D CNN model (such as C3D or I3D), which has been pre-trained on a large facial expression dataset and is capable of recognizing basic micro-expression types. The output of the 3D CNN is a feature vector, with each dimension corresponding to the activation intensity of a specific micro-expression type.

[0105] Besides using 3D CNNs, a complementary approach can be employed: encoding using Facial Action Units (AUs). AUs are the basic building blocks of facial expressions. According to the Facial Action Coding System (FACS) standard, facial expressions can be decomposed into combinations of several AUs (e.g., AU1 represents the inner corner of the eyebrow rising, AU12 represents the corners of the mouth pulling upwards, etc.). By detecting the activation intensity and temporal patterns of AUs in consecutive frames, a temporal feature matrix of AUs is formed, which is then compressed into a fixed-dimensional feature vector using a convolutional neural network or a recurrent neural network.

[0106] For the electrodermal response signal, time-domain statistical features and frequency-domain power spectrum features are extracted.

[0107] Time-domain statistical characteristics include: mean, variance, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, rise time (the time required for the signal to rise from the baseline to the peak value), and half-life (the time required for the signal to fall from the peak value to half of the peak value).

[0108] The frequency domain power spectrum characteristics were obtained by transforming the electrodermal signal (EDS) signal to the frequency domain using a Short-Time Fourier Transform (STFT), and then calculating the power spectral density (PSD) of each frequency band. The effective frequency band of the EDS signal is mainly concentrated in the 0-2Hz range. The low-frequency component (0-0.2Hz) mainly reflects the baseline skin conductance level (SCL), reflecting the user's long-term emotional tone; the mid-frequency component (0.2-0.6Hz) mainly reflects the phased skin conductance response (SCR), corresponding to the user's instantaneous emotional response; and the high-frequency component (0.6-2Hz) contains noise and subtle variations.

[0109] After concatenating the above time-domain and frequency-domain features and performing standardization, the skin conductance response feature vector is obtained.

[0110] For EEG signals, power spectral density and functional connectivity strength features of multiple frequency bands (δ wave: 0.5-4Hz, θ wave: 4-8Hz, α wave: 8-13Hz, β wave: 13-30Hz, γ wave: 30-50Hz) of multiple brain regions (such as frontal lobe, temporal lobe, parietal lobe, occipital lobe) are extracted.

[0111] Power spectral density features were estimated using the Welch method. For each channel's EEG data, it was segmented into multiple overlapping Hamming windows. After calculating the FFT for each window, the average of the squared amplitudes was taken to obtain the power spectrum estimate. Then, the average power spectral density within each frequency band was calculated to obtain the energy value of that channel in that band. Studies have shown that the degree of user attention concentration is related to theta wave activity in the frontal lobe region, while emotional valence is related to the frontal asymmetry index.

[0112] Functional connectivity strength measures the degree of temporal synchronization between different brain regions, reflecting the integration state of the brain's functional networks. Specifically, Phase Locking Value (PLV) is used to calculate the functional connectivity strength between different channels. The calculation formula is: ,in and They represent channels respectively. and channels In time The instantaneous phase. The value ranges from 0 to 1, with a higher value indicating a higher degree of phase synchronization between the two channels. Changes in functional connection strength can reflect changes in network topology under different cognitive states of the user.

[0113] The power spectral density and functional connectivity strength features are concatenated to obtain the EEG feature vector. The dimension of this vector is determined by the number of channels and frequency bands. For example, for a 14-channel EEG device with 5 frequency bands, the power spectral density feature alone has 70 dimensions.

[0114] The feature alignment unit 205 is connected to the biometric encoding unit 204 and is used to align the biometric vector with other feature vectors (behavioral feature vector, attribute feature vector, context feature vector) in the temporal dimension and output the temporally aligned multimodal feature tensor.

[0115] Because different types of data sources have different sampling frequencies and time bases, directly mixing them can lead to timing misalignment problems. For example, behavioral data may be event-driven (generating data only when the user performs an action), with discontinuous sampling times; biometric signals are typically continuous with a fixed sampling rate (such as EEG signals at 256Hz). The function of feature alignment unit 205 is to align all these asynchronous data streams onto a unified timeline.

[0116] In its implementation, the feature alignment unit 205 employs the following alignment strategy: Using the ad exposure time t0 as the central reference point for alignment, a time window is taken both forward and backward (e.g., 10 seconds forward and 5 seconds backward). The window length can be adjusted according to the ad type and prediction requirements. For continuously sampled signals (such as eye-tracking, EEG, and electrodermatology), a sliding window method is used to extract signal data within the window on the time axis, and the data is resampled at a fixed frequency to obtain a time-series data matrix. For event-driven signals (such as click records and interactive operations), all events related to the time window are encoded into an event list, and feature representations within the window are generated through event embedding aggregation. Finally, all features are aligned along the time dimension within the window, outputting a multimodal feature tensor with dimensions of (time window length × number of modalities × feature dimension).

[0117] The dynamic feature fusion module 300 is connected to the multimodal feature extraction module 200 and is used to input behavioral feature vectors, attribute feature vectors, context feature vectors and at least one biometric feature vector into an adaptive attention fusion network. The network calculates the cross-attention weights between multiple feature vectors through a multi-head self-attention mechanism and generates a fused feature vector based on the cross-attention weights. The adaptive attention fusion network includes a learnable feature weight allocation layer, which dynamically adjusts the fusion weights of multiple feature vectors according to the real-time biometric signals of different users at the prediction time.

[0118] Please refer to Figure 3 , Figure 3 A block diagram of the internal structure of the adaptive attention fusion network is shown. Figure 3As shown, the adaptive attention fusion network includes: a feature embedding layer 301, a multi-head self-attention layer 302, a feature weight allocation layer 303, and a feature fusion layer 304.

[0119] The feature embedding layer 301 is used to map each input feature vector to an embedding space of the same dimension to obtain the embedding feature matrix.

[0120] Since feature vectors from different modalities may originally have different dimensions—for example, behavioral feature vectors might be 256-dimensional, attribute feature vectors might be 128-dimensional, contextual feature vectors might be 64-dimensional, and EEG feature vectors might be 128-dimensional—they need to be mapped to the same dimensional space D (e.g., D=256) before directly calculating attention. The feature embedding layer 301 performs a linear transformation on the feature vectors of each modality using their respective fully connected projection matrices.

[0121] ;

[0122] in It is the first Input feature vectors of each modality It is the projection matrix of this mode. It is a bias term. It is the first The embedding vectors of M modalities are processed by the embedding layer, and each modality is converted into a D-dimensional embedding vector. Stacking the embedding vectors of M modalities row by row yields an M×D embedding feature matrix.

[0123] It should be noted that fully connected projection matrices The parameters can be learned and optimized during model training using the backpropagation algorithm. This end-to-end learning approach ensures that the features of each modality are optimally mapped to the shared semantic space.

[0124] The multi-head self-attention layer 302 is connected to the feature embedding layer 301 and is used to perform multi-head self-attention calculation on the embedded feature matrix to generate attention-weighted feature representation.

[0125] Multi-head self-attention is a core component of the Transformer model. Its core idea is to allow the model to simultaneously attend to all other features while processing a single feature, assigning different attention weights based on their relevance to the current feature. This mechanism is particularly well-suited for handling multimodal fusion problems because it can capture the complex interactions between features from different modalities.

[0126] In the specific implementation, for the embedded feature matrix (M is the number of modalities, D is the embedding dimension), multi-head self-attention is calculated as follows:

[0127] First, X is compared with three learnable weight matrices. Multiply to obtain the query matrix Key matrix Sum matrix .

[0128] Then, calculate the attention score: ,in It is the dimension of the key vector, divided by This is to prevent the gradient from becoming too large and thus causing it to vanish.

[0129] The attention score is then converted into a probability distribution using the Softmax function: Attention = Softmax(Score).

[0130] Finally, calculate the weighted sum: .

[0131] Multi-head self-attention refers to the simultaneous use of multiple different sets of self-attention. The matrix (i.e., multiple attention heads) performs the above process separately, and then concatenates the outputs of each head along the feature dimension:

[0132] ;

[0133] in .

[0134] In this embodiment, to reduce computational complexity, the multi-head self-attention layer 302 employs a dimensionality reduction mechanism. The traditional self-attention computational complexity is O(M). 2 ·D), where M is the sequence length (here M is the number of modalities). Although M is relatively small (typically 5-10 modalities), further reducing complexity is still meaningful in scenarios requiring low latency. The principle of the dimensionality reduction mechanism is: before calculating the attention, the dimensions of Q and K are first reduced from D to D' (D' < D) through linear projection, then the attention score is calculated, and finally, the dimensions are increased back to D through another projection. This dimensionality reduction mechanism can significantly improve computational efficiency while sacrificing a small amount of accuracy.

[0135] The feature weight allocation layer 303 is connected to the multi-head self-attention layer 302 and is used to dynamically calculate the fusion weight vector of each modality feature based on the biometric state recognition result generated by the user's biometric signal data. The biometric state recognition result is used to characterize the user's current level of attention concentration and emotional arousal.

[0136] This is another core innovation of the present invention. Traditional multimodal fusion methods typically use fixed weights or weights that rely entirely on attention mechanisms for automatic learning. However, these weights are based solely on the relationships between features and do not consider the user's real-time physiological and psychological state. For example, a user may be in a distracted state (such as in a noisy environment), at which point their behavioral data (such as browsing history and click history) may be less reliable than usual, while eye-tracking signals (such as pupil changes and gaze patterns) may better reflect their true state. The feature weight allocation layer 303 is designed to address this problem.

[0137] In practice, the user's current state is first classified or regressed based on biometric vectors (especially eye movement and EEG signals). A pre-trained lightweight classification network is used, which takes the biometric vector as input and outputs three-dimensional state indicators: attention level (a value between 0 and 1, where 0 represents complete distraction and 1 represents complete concentration), emotional arousal level (a value between 0 and 1, where 0 represents low arousal and 1 represents high arousal), and emotional valence (a value between -1 and 1, where negative values ​​represent negative emotions and positive values ​​represent positive emotions).

[0138] Then, these three state indicators are combined with a learnable weight basis vector. The feature weight allocation layer 303 maintains a weight basis matrix. Where M is the number of modes and K is the number of state indices (K=3 here). For each mode i, its fusion weight is calculated as the dot product of the weight basis vector and the state index vector, followed by Softmax normalization:

[0139] ;

[0140] in It is the first The values ​​of each status indicator, It is modal The fusion weight.

[0141] The rationale for this design lies in: the weight basis matrix This can be understood as "the influence factor of each state index on the credibility of different modal features." These influence factors are learned through a large amount of labeled data during model training. For example, if the training data shows that the degree of attention concentration is positively correlated with the credibility of behavioral features (i.e., the more concentrated the attention, the more reliable the behavioral data), then the attention factor in the weight base corresponding to the behavioral features will be a large positive value; if the degree of attention concentration is negatively correlated with the credibility of eye movement features (i.e., when attention is scattered, eye movement features can better reflect the true attention state), then the attention factor in the weight base corresponding to the eye movement features may be a negative value.

[0142] The feature fusion layer 304 is connected to the feature weight allocation layer 303, and is used to perform weighted fusion of the attention-weighted feature representations according to the fusion weight vector to generate a fusion feature vector.

[0143] The specific calculation is as follows:

[0144] ;

[0145] in It is the first Feature representation of each modality after processing by a multi-head self-attention layer The fusion weights are calculated by the feature weight allocation layer. It is the final fused feature vector.

[0146] Fusion feature vectors It includes information enhanced by attention weighting and multimodal interaction, and adaptively adjusts the weights based on the user's real-time biometric state, so that the model can obtain information from the most reliable features in different user states.

[0147] The fused feature vector output by the dynamic feature fusion module 300 will be used as the initial feature input for the user nodes in the graph neural network encoding module 400.

[0148] The graph neural network encoding module 400 is connected to the dynamic feature fusion module 300. It is used to construct a heterogeneous graph structure by using the fused feature vector as node features and at least one of the following relationships: historical interaction relationship between users and advertisements, social relationship between users, and content association relationship between advertisements. The heterogeneous graph structure is then encoded by a multi-layer graph neural network to obtain a graph embedding feature vector containing structural information.

[0149] Please refer to Figure 4 , Figure 4 A block diagram of the internal structure of the graph neural network encoding module 400 is shown. Figure 4 As shown, the graph neural network encoding module 400 includes: a heterogeneous graph construction unit 401, a node feature initialization unit 402, a neighbor aggregation unit 403, and a pooling encoding unit 404.

[0150] The heterogeneous graph construction unit 401 is used to construct heterogeneous graph structures based on user nodes, ad nodes, and context nodes. The edge types in the heterogeneous graph include various types such as user history click edges, user-ad exposure edges, ad content similarity edges, and user social relationship edges.

[0151] Unlike ordinary isomorphic graphs, heterogeneous graphs can have a variety of node and edge types. In this embodiment, the node set of the heterogeneous graph includes:

[0152] User node (U): Each user corresponds to one user node;

[0153] Ad Node (A): Each ad corresponds to one ad node;

[0154] Context Node (C): Each context context mode corresponds to a context node (optional, whether to enable it depends on the actual application scenario).

[0155] The edge sets of heterogeneous graphs include the following types:

[0156] (1) User history click edge (UA edge): If a user has clicked on an ad within a certain period of time, a click edge is established between the corresponding user node and ad node. The edge weight can be set according to factors such as the number of clicks, the click time, and the conversion behavior after the click. For example, the more clicks, the higher the weight; the more recent the click time, the higher the weight; and the edge weight that generates a conversion behavior after the click is doubled.

[0157] (2) User-Ad Exposure Edge (UA Exposure Edge): If a user has been shown an ad in the past (regardless of whether they clicked it), an exposure edge is established between the user node and the ad node. This edge is of great value for learning negative samples and predicting cold start scenarios. The weight of the exposure edge can be set according to factors such as the number of exposures and the duration of exposure.

[0158] (3) Ad content similarity edge (AA edge): If two ads are similar in content features, an ad-ad similarity edge is established. The content similarity of ads can be obtained by calculating the cosine similarity between the word embedding vectors of the ad title and description text, or by calculating the feature distance between the ad image embeddings. An edge is established between ad pairs whose similarity exceeds a set threshold, and the edge weight is the similarity value.

[0159] (4) User-to-User (UU) Social Relationship Edge: If two users have a social relationship (such as friendship, following, or belonging to the same interest group), a social edge is established between the user nodes. The weight of the social relationship can be set according to the relationship type and intimacy. The introduction of this edge is based on the theory of social homogeneity: users in social networks often have similar interests and preferences, and therefore user interest information can be disseminated through social relationships.

[0160] (5) User-Context Edge (UC Edge): If a user exhibits a certain behavioral pattern in a specific context, an edge can be established between the user node and the context node. Context nodes refer to typical context category nodes formed by clustering context features, such as "leisure browsing in a WiFi environment at home at night" or "fragmented browsing on a mobile network on the subway during the day." Context dependencies of user behavior can be constructed through UC edges.

[0161] In practice, the heterogeneous graph construction unit 401 periodically performs offline graph construction and updates on the full data (e.g., building a basic graph once per hour or per day), while performing incremental graph updates on real-time interactions (e.g., when a user makes a new click), that is, only updating the affected nodes and edges without rebuilding the entire graph, in order to improve response speed.

[0162] The node feature initialization unit 402 is connected to the heterogeneous graph construction unit 401 and is used to use the fused feature vector output by the dynamic feature fusion module as the initial feature vector of the user node in the heterogeneous graph.

[0163] For the initialization features of an ad node, the static features of the ad (such as advertiser identifier, ad content type, ad format, bidding information, etc.) and dynamic features (such as recent exposure count, click-through rate, conversion rate, etc.) are concatenated and then encoded through a fully connected layer to obtain the ad node feature vector.

[0164] For the initialization features of context nodes, the context feature vector corresponding to the typical context pattern is used as the initialization feature of the context node.

[0165] All nodes are initialized with a feature vector dimension of D (e.g., D=256) to ensure numerical compatibility for subsequent graph convolution operations.

[0166] The neighbor aggregation unit 403 is connected to the node feature initialization unit 402 and is used to perform type-aware neighbor aggregation operations according to the type of the edge in each layer of the multi-layer graph neural network. Different types of edges use different aggregation functions and attention weights.

[0167] This embodiment employs a heterogeneous graph convolutional network (HGCN) architecture. Unlike homogeneous graph convolution, heterogeneous graph convolution needs to distinguish the types of neighbors (i.e., differences in the types of connecting edges) when processing the neighbors of node v, and aggregates the information of different types of neighbors separately.

[0168] Specifically, for the l-th layer graph neural network, the feature representation update formula for node v in the (l+1)-th layer is:

[0169] ;

[0170] Where T is the set of all edge types. It is the set of neighboring nodes of node v connected by edges of type t. It is the learnable transformation matrix of the l-th layer network for edge type t. σ is the transformation matrix of the node's own features, and σ is the activation function (usually ReLU).

[0171] To better capture the differences in importance among different neighboring nodes, in addition to using mean aggregation, an attention-based aggregation method is introduced. For each neighbor u in the set of neighboring nodes of type t, its attention coefficient relative to node v is calculated:

[0172] ;

[0173] ;

[0174] Then, weighted aggregation is performed using the attention coefficient:

[0175] ;

[0176] This type-aware attention aggregation mechanism can adaptively adjust the intensity of information flow according to the importance of different user-advertisement relationships, enabling graph neural networks to better extract valuable structural information from the relational network during the encoding process.

[0177] Pooling coding unit 404 is connected to neighbor aggregation unit 403 and is used to output the user node embedding vector output by the last layer of the multilayer graph neural network as the graph embedding feature vector.

[0178] Specifically, after L layers of graph convolution (L is usually 2 to 3 layers), each user node obtains a node embedding vector containing L-hop neighbor information. Pooling encoding unit 404, based on the prediction target (typically the click-through rate of the current user), extracts the final embedding vector of that user node. The embedded feature vector in this graph contains the user's personalized features (from the initial feature vector) as well as complex relationship information between the user and advertisements, other users, and context.

[0179] It should be noted that when predicting the click-through rate of a specific ad, the final embedding features of the ad node can also be obtained simultaneously. Then, the user node embedding and the ad node embedding are concatenated as input to the prediction module. However, in this embodiment, considering design simplicity and module independence, the graph neural network encoding module 400 only outputs the embedding vector of the user node as the backbone feature, and the ad features are fused in subsequent steps.

[0180] The adaptive route prediction module 500 is connected to the graph neural network encoding module 400 and is used to input the graph embedding feature vector into an adaptive route prediction network. The adaptive route prediction network includes multiple parallel expert prediction subnetworks and at least one router network. The router network dynamically selects and activates at least one expert prediction subnetwork based on the real-time changes of the graph embedding feature vector and user biometric signal data, and weights and fuses the outputs of the activated expert prediction subnetworks to generate a preliminary click-through rate prediction value.

[0181] Please refer to Figure 5 , Figure 5 A block diagram of the internal structure of an adaptive route prediction network is shown. Figure 5 As shown, the adaptive route prediction network includes: a state encoder 501, a router network 502, multiple parallel expert prediction subnetworks (four are shown in the figure, but the actual number can be set as needed), and an output fusion layer 504.

[0182] The state encoder 501 is used to jointly encode the temporal changes of graph embedding feature vectors and user biometric signal data to generate routing state vectors.

[0183] The basis for routing decisions is the specific scenario in which the prediction task takes place. Scenario information mainly includes two aspects: first, user characteristics after graph encoding (reflecting the user's personality preferences and social influences); and second, the user's real-time biometric state (reflecting the user's current attention level, emotional state, etc.). The role of the state encoder 501 is to fuse these two types of information into a compact routing state vector.

[0184] In the specific implementation, the graph embedding feature vector h_graph and the biofeedback vector h_bio (taken from the current window aggregated representation of the biofeedback vector extracted by the multimodal feature extraction module 200, for example, compressing the biofeedback temporal window into a vector through average pooling) are concatenated into a joint vector. Then, a multi-layer perceptron (MLP) is used to map the joint vector to a routing state vector. :

[0185] ;

[0186] in It is the dimension of the routing state vector, which is set according to the number of expert subnetworks, and is usually 16, 32 or 64.

[0187] Router network 502 is connected to state encoder 501 and includes a gating mechanism network for calculating the selection probability of each of multiple parallel expert prediction subnetworks based on the routing state vector, and activating at least one expert prediction subnetwork based on the selection probability. When the maximum selection probability is greater than a preset threshold, a single expert prediction subnetwork is activated; otherwise, multiple expert prediction subnetworks are activated.

[0188] Router network 502 is the core decision-making component of the adaptive route prediction network. Unlike the traditional Mixture of Experts (MoE) model, which activates all experts for each sample, the router network 502 of this invention uses a sparse activation mechanism to select only one or a few of the most relevant experts for each prediction task, thereby reducing computational overhead while ensuring prediction accuracy.

[0189] In its implementation, router network 502 includes a linear transformation layer and a softmax layer:

[0190] ;

[0191] ;

[0192] in It is a size of ( The weight matrix of ) It is the number of expert subnetworks. yes A probability distribution vector of dimension 1. Indicates selecting the first The probability of an expert subnetwork.

[0193] probability distribution This reflects the degree of matching between each expert subnetwork and the prediction task in the current routing state. For example, when a user's graph embedding feature vector indicates that the user is a new user with sparse behavior, the router may tend to select the first expert (using a deep FM architecture) that is good at handling cold start or sparse feature scenarios; when a user's biometric signals indicate that the user is currently in a highly focused attention state, the router may tend to select a sequence Transformer architecture to better handle the user's recent behavior sequence; when the user's features involve multiple patterns simultaneously, the router may also select a combination of multiple experts.

[0194] When determining which experts to activate, router network 502 uses the following strategy: set a preset threshold τ (τ is usually a value between 0.5 and 0.7). .if Then only activate The corresponding single expert subnetwork has the selected expert index as follows: ;if This indicates that no single expert has an absolute advantage, thus requiring the activation of multiple expert subnetworks. A strategy for activating multiple experts could be: selecting all... Among the experts, It can be a small threshold (e.g., 0.2); or select the top K experts with the highest probability, where K is a configurable parameter (e.g., 2 or 3).

[0195] It is worth noting that the gating mechanism of router network 502 also incorporates real-time changes in user biometric signal data. Specifically, the input routing state vector s of router network 502 includes information from the biometric vector, allowing router network 502 to dynamically adjust the expert selection strategy based on changes in the user's current level of attention. For example, when the user's attention level is low, the router may tend to select a first expert (deep FM) with a simpler structure and more robust predictions, while avoiding reliance on a third expert (sequence transformer) that is sensitive to long sequences, because user behavior may be more random and unstable at low attention levels.

[0196] Multiple parallel expert prediction subnetworks each have different network structures and parameter configurations, and are used to handle different types of prediction scenarios.

[0197] In this embodiment, four expert sub-networks are set up, namely:

[0198] (1) First expert subnetwork (expert A): adopts the Deep Factorization Machine (DeepFM) architecture.

[0199] DeepFM combines Factorization Machines (FM) for low-order feature interaction modeling and Deep Neural Networks (DNNs) for high-order feature interaction modeling, making it particularly suitable for handling the abundant sparse discrete features in user-ad features. DeepFM achieves end-to-end joint training through a shared feature embedding layer, avoiding the separation problem of two-stage training in traditional methods. In this embodiment, the input to the first expert sub-network is the concatenation of the graph embedding feature vector h_graph and the ad feature vector h_ad, and the output is a prediction score (click probability). The advantage of the DeepFM architecture lies in its strong ability to model sparse features and low-order interactions, making it suitable for scenarios with sparse user interaction data.

[0200] (2) Second expert sub-network (expert B): adopts the Deep & Cross Network (DCN-V2) architecture.

[0201] DCN-V2 explicitly models high-order feature interactions through cross layers, while retaining the implicit high-order interactions captured by the DNN. Compared to DeepFM, DCN-V2 offers finer-grained modeling of feature interactions and can automatically learn feature interactions of any order. In this embodiment, the second expert subnetwork is particularly suitable for scenarios with complex nonlinear interaction relationships between ad features, between user features, and between user-ad features. For example, when there are multiple interaction effects between user attributes, ad content, contextual information, and biometrics, DCN-V2 can effectively capture these interaction relationships.

[0202] (3) Third expert sub-network (expert C): adopts the Sequence Transformer architecture.

[0203] Sequence transformers, based on a self-attention mechanism, excel at processing user behavior sequence data. In this embodiment, the third expert subnetwork takes the user's continuous behavior sequence before ad exposure (the most recent N clicks, browsing, searches, etc.) as input and performs sequence modeling through a multi-layer Transformer encoder to capture the temporal dependencies and pattern changes between behaviors. The advantage of the sequence transformer architecture lies in its ability to learn long-distance temporal dependencies, making it suitable for scenarios where user behavior patterns have strong periodicity and sequence.

[0204] (4) Fourth expert sub-network (expert D): adopts the architecture of multi-gate mixture-of-experts (MMoE).

[0205] MMoE was originally designed for multi-task learning, achieving adaptive representation learning for different tasks through a combination of multiple expert networks and a gating network. In this embodiment, the fourth expert sub-network treats ad click-through rate prediction as a multi-task problem (e.g., simultaneously predicting click probability and conversion probability), and achieves knowledge sharing and differentiated learning between tasks through the MMoE architecture. This architecture is particularly suitable for multi-scenario, multi-task learning environments. For example, when the same system needs to provide click-through rate prediction for e-commerce ads, brand ads, and app download ads simultaneously, MMoE can learn the common feature representations shared between different ad types and their respective specific feature representations.

[0206] It should be noted that the specific architecture choices of the four expert subnetworks mentioned above do not constitute a limitation on the present invention. In practical applications, the number of experts can be increased or decreased, or other types of network architectures (such as experts based on graph neural networks, experts based on autoencoders, etc.) can be replaced according to specific business scenarios and data types.

[0207] The output fusion layer 504 is connected to the router network 502 and multiple parallel expert prediction subnetworks to perform a weighted summation of the outputs of the activated expert prediction subnetworks according to the selection probability, generating a preliminary click-through rate prediction value.

[0208] Assume the selection probability output by router network 502 is p( (dimensional vector), the set of indices of the activated expert subnetwork is For the activated expert subnetworks, their outputs are used for weighted fusion:

[0209] ;

[0210] in It is the first The predicted value output by each expert subnetwork (a scalar between 0 and 1, representing the click probability) is calculated, with the denominator being the selection probability of the activated expert subnetwork normalized. The output of inactive expert subnetworks is ignored in this prediction and not included in the calculation.

[0211] Compared with the traditional weighted average of all expert outputs, this sparse activation fusion strategy has the following advantages: (1) It is more computationally efficient because only some experts are activated; (2) It has better discrimination for different scenarios, and the router learns to select the most suitable expert combination based on the input features; (3) The model is more interpretable and can infer the feature type of the current prediction task based on the router's selection results.

[0212] Preliminary click-through rate prediction value output by the adaptive route prediction module 500 It is a scalar between 0 and 1, representing the predicted probability that a user will click on the target ad given the current input features.

[0213] The real-time feedback calibration module 600 is connected to the data acquisition module 100 and the adaptive route prediction module 500 respectively. It is used to generate feedback labels based on the user's subsequent real click behavior data after the current advertisement exposure, and to perform online incremental updates on the decision parameters of the router network of the adaptive route prediction module based on the feedback labels. At the same time, it generates calibration signals to feed back to the dynamic feature fusion module to adjust the learnable feature weight allocation layer in the adaptive attention fusion network.

[0214] Please refer to Figure 6 , Figure 6 An internal structural block diagram of the real-time feedback calibration module 600 is shown. Figure 6 As shown, the real-time feedback calibration module 600 includes: a feedback data acquisition unit 601, an online gradient update unit 602, a calibration signal generation unit 603, and a delayed feedback processing unit 604.

[0215] The feedback data acquisition unit 601 is used to collect the user's actual click behavior data within a preset time window after the current advertisement is exposed. The actual click behavior data includes whether a click was made and the time when the click occurred.

[0216] The length of the preset time window can be configured according to different business scenarios. For Real-Time Bidding (RTB) ads, users typically make a click decision within seconds of ad exposure, so the time window can be set to 30 seconds or 1 minute. For brand ads or content recommendation ads, users' click decisions may occur later, so the time window can be set to 30 minutes, 1 hour, or even longer. In this embodiment, the default time window is set to 5 minutes, while allowing ad operators to dynamically adjust it according to the characteristics of different ad types.

[0217] When a user clicks within a preset time window after the current ad is displayed, a feedback tag is displayed. =1; If the user does not click within this time window, then = 0 (negative sample). It should be noted that negative samples may contain some noise, because users may click at a later time. However, considering the real-time requirements of ad delivery, adopting a truncation strategy with a preset time window is a reasonable and efficient engineering practice.

[0218] The online gradient update unit 602 is connected to the feedback data acquisition unit 601 and is used to calculate the loss gradient based on the difference between the actual click behavior data and the initial click rate prediction value. The loss gradient is then used to update the parameters of the router network online using gradient descent. The online gradient update is performed in real time after each feedback arrives without retraining the entire prediction network.

[0219] This is the key technology for achieving rapid model adaptation and user interest tracking in this invention. Traditional methods require accumulating sufficient new data before performing batch offline retraining, resulting in slow response times and an inability to capture real-time changes in user interests. In contrast, the online gradient update unit 602 of this invention employs online learning, immediately updating the router network parameters each time user feedback arrives, enabling the model to quickly adapt to the latest user feedback information.

[0220] In the specific implementation, the loss function uses binary cross entropy (BCE) loss:

[0221] ;

[0222] in It is a real label (0 or 1). This is the click probability predicted by the model. This applies to parameters in a router network. Calculate the loss pair gradient ( Then perform online gradient descent updates:

[0223] ;

[0224] in This is the learning rate, which is usually set to a small value (such as 0.001~0.01) to ensure the stability of parameter updates.

[0225] It should be noted that, in this embodiment, the online gradient update unit 602 mainly updates the parameters of the router network (i.e., The router's decision boundary needs to adapt to changes in user behavior in real time, while the expert subnetwork's parameters are updated in batches at a relatively low frequency. This "fast router update, slow expert update" design is based on the following considerations: the router's decision boundary needs to adapt to changes in user behavior in real time, while the expert subnetwork's core feature extraction capability is relatively stable and can be maintained for a longer period. This two-stage update strategy reduces the parameter update overhead and overfitting risk of the entire system while ensuring rapid response.

[0226] The calibration signal generation unit 603 is connected to the online gradient update unit 602 and is used to generate a calibration signal based on the sign and magnitude of the loss gradient. The calibration signal contains information indicating the adjustment direction and adjustment step size of the feature weight allocation layer in the dynamic feature fusion module.

[0227] The calibration signal generation unit 603 establishes an information bridge between the online feedback calibration module and the dynamic feature fusion module. Its design philosophy is that when the online gradient update unit detects a significant change in the router's network parameters, it indicates a substantial change in the current user's feature distribution or behavioral pattern, which may require the dynamic feature fusion module to adjust its multimodal fusion weight allocation strategy.

[0228] In the specific implementation, the calibration signal generation unit 603 monitors the update process of the online gradient update unit 602, recording the magnitude of changes in the loss function value and model parameters in each update. When the loss function value shows a continuous upward trend in multiple consecutive updates, or when the parameter change magnitude exceeds a preset threshold, a calibration signal C is generated. The calibration signal C is a vector whose dimension matches the parameter dimension of the feature weight allocation layer in the dynamic feature fusion module. Each component of the calibration signal indicates the adjustment direction and step size of the corresponding modality fusion weight.

[0229] For example, if the component corresponding to a behavioral feature in the calibration signal is positive, it means that the weight of the behavioral feature should be increased in the fusion process; if the component corresponding to a biological feature is negative, it means that the weight of the biological feature should be decreased. This two-way information feedback forms a closed-loop optimization mechanism of "prediction → feedback → calibration", enabling the entire system to continuously improve and adapt itself.

[0230] The delayed feedback processing unit 604 is used to cache delayed click feedback data that arrives after a preset time window, and to perform parameter updates in batches after the cache quantity threshold is met.

[0231] In real-world advertising environments, some users' clicks may occur after a preset time window (e.g., a user is attracted to an ad but later clicks through another channel). This "late" feedback data also contains valuable information and cannot be simply discarded. The delayed feedback processing unit 604 is specifically designed to process this type of data.

[0232] In its implementation, the delayed feedback processing unit 604 maintains a feedback buffer queue. For feedback within a preset time window... Subsequent user click feedback (i.e., the ad exposure time corresponding to the feedback is earlier than the current time minus) ), and cache them. When the number of feedback samples in the cache queue reaches a preset threshold. When there are 100 (e.g.) data points, a batch parameter update is triggered, and the model parameters are updated using batch gradient descent with these delayed feedback data.

[0233] Batch updates offer better statistical stability and convergence compared to single-sample updates, effectively mitigating noise interference that may be introduced by a single delayed feedback sample. Furthermore, the low frequency of delayed feedback batch updates (e.g., updating only once every 100 delayed feedback samples) does not significantly impact the real-time performance of online predictions.

[0234] Through the collaborative work of the above four units, the real-time feedback calibration module 600 achieves efficient online learning and continuous optimization of the prediction system.

[0235] The multi-objective optimization decision module 700 is connected to the real-time feedback calibration module 600. It acquires at least two ad candidate sets and, based on preliminary click-through rate (CTR) predictions and calibration signals output by the real-time feedback calibration module, calculates a comprehensive placement value score for each ad in the at least two candidate sets. This comprehensive placement value score is calculated using a multi-objective optimization function based on at least two of the following indicators: predicted CTR, expected revenue, and user negative feedback risk. The multi-objective optimization decision module 700 then sorts the ads in the candidate sets according to the comprehensive placement value score and outputs an ad recommendation list.

[0236] Please refer to Figure 7 , Figure 7 A block diagram of the internal structure of the multi-objective optimization decision module 700 is shown. (See diagram below.) Figure 7 As shown, the multi-objective optimization decision module 700 includes: a multi-dimensional index calculation unit 701, a multi-objective optimization unit 702, and a sorting output unit 703.

[0237] The multi-dimensional indicator calculation unit 701 is used to calculate the predicted click-through rate correction value based on the preliminary click-through rate prediction value and calibration signal for each candidate advertisement, calculate the expected revenue value based on the advertiser's bidding information and advertisement type, and calculate the user negative feedback risk value based on the user's historical negative feedback data.

[0238] The calculation methods for these three indicators are explained below:

[0239] (1) Predicted click-through rate correction value ( )

[0240] Original preliminary click-through rate prediction The calibration signal C output by the real-time feedback calibration module 600 is used for correction. A simple additive calibration method can be employed for this correction.

[0241] = clamp( + λ * C, 0, 1);

[0242] The `clamp` function restricts the correction value to the range [0,1], and λ is a configurable calibration intensity hyperparameter (typically 0.01~0.1). The physical meaning of additive calibration is: if online feedback indicates that the probability of users clicking on similar ads is increasing recently (calibration signal is positive), then the predicted click-through rate is appropriately increased; otherwise, it is decreased.

[0243] More complex multiplicative calibration or Bayesian-based calibration can also be used, with additive calibration schemes offering advantages in computational efficiency and interpretability.

[0244] (2) Expected return value ( )

[0245] For e-commerce or performance-based advertising, advertisers typically set a bid price, representing the maximum cost they are willing to pay per click. In a real-time bidding (RTB) advertising system, the platform's revenue primarily comes from the product of the advertiser's bid and the probability of a click. Furthermore, different types of ads (such as brand ads, performance ads, and app download ads) have different long-term values ​​for the platform, which can be addressed by introducing an ad type coefficient. To adjust the calculation of expected returns.

[0246] The formula for calculating expected return is:

[0247] ;

[0248] in It is the bid per click set by the advertiser. It is the advertising type coefficient (for example, γ can be 1.0 for performance advertising, 1.2 for brand advertising with brand premium, and 0.5 for public service advertising).

[0249] (3) Negative feedback risk value

[0250] User negative feedback risk is an important indicator for measuring the potential negative impact of advertising on user experience. Over-distributing irrelevant ads or frequently displaying the same ad repeatedly can lead to user aversion and negative feedback behaviors (such as closing the ad, blocking the advertiser, or uninstalling the app). While these negative feedback behaviors do not directly affect short-term revenue, they have a significant impact on the platform's long-term health and user retention.

[0251] In this embodiment, the calculation of the negative feedback risk value takes into account the following factors:

[0252] The number of times a user has historically given negative feedback to this advertiser or this type of ad in the current session (e.g., having closed similar ads 3 times, indicating a higher risk).

[0253] Users' historical click-through rate and close rate for this type of ad; ad types with high close rates and low click-through rates are considered to be of higher risk.

[0254] Users' current emotional state (identified through biometric signals) indicates that users in a negative emotional state are more likely to generate negative feedback.

[0255] After weighting and summing these factors, the result is mapped to the [0,1] interval using the Sigmoid function, yielding the negative feedback risk value risk_neg. The larger the value, the higher the risk of the advertisement triggering negative feedback from users.

[0256] The multi-objective optimization unit 702 is connected to the multi-dimensional indicator calculation unit 701. It is used to comprehensively calculate the predicted click-through rate correction value, expected revenue value and user negative feedback risk value through a preset multi-objective optimization function to obtain the comprehensive placement value score of each candidate advertisement. The multi-objective optimization function adopts the Pareto optimization strategy, which allows advertisers to dynamically adjust the weight coefficients of each indicator according to different business objectives.

[0257] Advertising placement is a typical multi-objective optimization problem, and different advertisers and platforms may have different optimization objectives at different times. For example, during the sprint period at the end of a quarter to pursue revenue targets, the weight of expected revenue may increase; in the early stages of a platform's launch, when user retention and experience are prioritized, the weight of negative feedback risk may increase. The multi-objective optimization unit 702 of this invention provides advertising operators with flexible multi-objective balancing capabilities through a weighted summation method with adjustable weight coefficients.

[0258] The formula for calculating the overall campaign value score is as follows:

[0259] ;

[0260] in:

[0261] , , These are configurable weighting coefficients that satisfy... + + = 1, the platform can dynamically adjust according to current business objectives;

[0262] It is a normalization factor for expected revenue, mapping the expected revenue values ​​of different ads to orders of magnitude comparable to p_ctr (e.g., = average( This ensures the additivity of different indicators;

[0263] The weighting sign for negative feedback risk is negative, because the higher the risk, the lower the overall score, encouraging the system to avoid high-risk deployments.

[0264] Within the theoretical framework of multi-objective optimization, the weighted summation method described above essentially selects a point on the Pareto front, the specific location of which is determined by the weighting coefficients. By adjusting the weighting coefficients, the system can make different trade-offs among the three objectives of predicting click-through rate, expected revenue, and the risk of negative user feedback.

[0265] In more advanced implementations, constrained Pareto optimization methods can be used: [The following is a list of parameters and their implications:] and Minimum acceptable threshold and The maximum acceptable threshold is determined by considering only advertisements that meet these constraints, and then selecting the optimal solution from them.

[0266] The sorting output unit 703 is connected to the multi-objective optimization unit 702 and is used to sort the candidate advertisements from high to low according to the comprehensive placement value score, select at least one advertisement with the highest score to generate an advertisement recommendation list and output it to the advertisement placement end.

[0267] In its implementation, the sorting output unit 703 first sorts all ads in the candidate ad set in descending order according to their calculated scores. Then, based on the ad slot size and the requirements of the ad placement strategy, it selects the top K ads (K is usually between 1 and 5) from the sorted results to form the final recommendation list. During the selection process, auxiliary rules such as ad diversity and frequency limits for repeated placements by the same advertiser are also considered to ensure a diverse user experience.

[0268] The ad recommendation list output by the sorting output unit 703 is sent to the ad delivery server (such as the ad display server of a media platform), where ads are displayed according to the order or weight of the recommendation list. Simultaneously, the recommendation results are also recorded in the system for subsequent feedback tracking and model training.

[0269] Please refer to Figure 8 , Figure 8 A schematic diagram of the distributed deployment architecture of the system in this embodiment is shown. Figure 8 As shown, the system in this embodiment adopts a distributed deployment architecture, which includes at least one central server and multiple edge computing nodes.

[0270] The distributed deployment architecture is designed based on the following considerations: the acquisition and feature extraction of user biometric signals require high-bandwidth, low-latency data transmission and strong real-time computing capabilities. If these processes are completed on a central server, it would result in a large amount of raw data being transmitted over the network, consuming bandwidth resources and posing serious privacy and security risks. Therefore, in this embodiment, the data acquisition module 100 and the multimodal feature extraction module 200 are deployed on the edge computing node side (i.e., close to the user terminal, which could be on the user terminal device itself or on the edge server closest to the user), completing the acquisition and preliminary feature extraction of user biometric signals near the user terminal. Only the low-dimensional feature vector after feature extraction is transmitted to the central server via encrypted network transmission.

[0271] The dynamic feature fusion module 300, graph neural network encoding module 400, adaptive route prediction module 500, real-time feedback calibration module 600, and multi-objective optimization decision-making module 700 are deployed on the central server side. These modules need to access global user data, advertising data, and relationship graph data, and deploying them on the central server can better utilize cluster computing resources and data synchronization mechanisms.

[0272] This embodiment deploys a privacy protection unit on the edge computing node side. Figure 8 (Not shown separately). The privacy protection unit is used to perform localized differential privacy processing on the collected user biometric signal data. After feature extraction is completed, the original biometric signal data is deleted, and only the feature vector is encrypted and transmitted to the central server.

[0273] Local Differential Privacy (LDP) is a technique that injects noise at the data source to protect user privacy. In its implementation, the privacy protection unit adds an appropriate amount of Laplace or Gaussian noise to the biometric vector output by the biometric encoding unit 204. The amplitude of the noise is controlled by the privacy budget parameter ε (the smaller ε is, the higher the level of privacy protection, but the greater the loss of data utility). For highly sensitive data such as biometrics, a smaller ε value (e.g., 0.1~0.5) is recommended.

[0274] After feature extraction and privacy protection processing, the original biometric signal data (such as video frames, raw EEG signals, etc.) is permanently deleted from local storage, retaining only the feature vectors with added noise and necessary data source descriptive metadata. This design complies with current mainstream privacy protection regulations (such as GDPR, Personal Information Protection Act, etc.) and effectively protects user data privacy.

[0275] The system of this invention requires model pre-training before it can be put into use. The training process consists of two stages:

[0276] Phase 1: Independent pre-training of each module: The adaptive attention fusion network of the dynamic feature fusion module, the graph neural network of the graph neural network encoding module, and each expert sub-network of the adaptive route prediction module are pre-trained independently, using historical ad click data as training samples.

[0277] Phase Two: Joint Fine-Tuning Training: All modules are assembled into a complete end-to-end network according to the connection method at the time of prediction, and the entire network is jointly trained using labeled data. In the joint training, the feature weight allocation layer of the dynamic feature fusion module and the router network of the adaptive route prediction module require additional high-quality training samples to converge. These samples need to contain complete user biometric signal records and corresponding click tags.

[0278] Once training is complete, the system can be deployed for online prediction services. During online operation, the real-time feedback calibration module 600 continuously updates the system parameters online, enabling the system to adaptively track dynamic changes in user interests and behaviors.

[0279] To better understand the overall working principle of this invention, the following explanation uses a complete advertising placement decision-making process as an example.

[0280] Suppose a user opens a news app at a certain moment. The system needs to select an ad from the candidate ad pool to display on the user's homepage. The system's workflow is as follows:

[0281] Step S1: When a user opens the App, the data acquisition module 100 immediately starts collecting data from that user. The behavior data acquisition unit acquires the user's recent clicks and browsing history (e.g., behavior records within the last 24 hours), the attribute data acquisition unit acquires the user's basic information (age, gender, etc.), the context data acquisition unit acquires the current time, location, network type, and page context information, and the biometric acquisition unit simultaneously acquires the user's eye movement trajectory, facial expressions, skin conductance response, and electroencephalogram (EEG) signals through the front-facing camera and wearable device.

[0282] Step S2: The multimodal feature extraction module 200 extracts features from the above data. The behavior feature extraction unit extracts user behavior sequence feature vectors, the attribute feature extraction unit extracts user attribute feature vectors, the context feature extraction unit extracts current context feature vectors, and the biometric encoding unit extracts various biometric vectors using techniques such as 3D CNN and power spectral analysis. The feature alignment unit aligns all features to the timeline of the advertisement exposure time.

[0283] Step S3: The dynamic feature fusion module 300 inputs the extracted multimodal features into the adaptive attention fusion network. The multi-head self-attention layer calculates the cross-attention between features of different modalities, and the feature weight allocation layer dynamically adjusts the fusion weights of each modality according to the user's biological state (e.g., low attention level, neutral emotional state). The fused feature vector integrates information from multiple dimensions, including the user's historical behavior, personal attributes, contextual information, and real-time biometrics.

[0284] Step S4: The graph neural network encoding module 400 uses the fused feature vector as the initial feature of the user node, performs multi-layer graph convolution in the pre-constructed heterogeneous graph, and incorporates information such as the user's social relationship with other users, the advertisements the user has interacted with, and the content similarity between advertisements into the feature representation, and finally outputs a graph embedding feature vector containing relational structure information.

[0285] Step S5: The state encoder of the adaptive route prediction module 500 jointly encodes the graph embedding feature vector and the biometric feature vector into a route state vector, which the router network uses to calculate the selection probability of the expert subnetwork. Assuming the user is a long-term user with rich historical behavior and a high level of current attention, the router network selects the third expert (Sequence Transformer) with high probability to specifically process their long-sequence behavioral data, while simultaneously fusing the output of the first expert (Depth FM) with lower weights. The output fusion layer generates a preliminary click-through rate prediction value after weighting according to the selection probability, for example, 0.72, indicating that the predicted probability of the user clicking the advertisement is 72% under the condition of ignoring other constraints.

[0286] Step S6: The real-time feedback calibration module 600 does not participate in the decision-making for this prediction, but performs online calibration based on the feedback data from the previous or earlier ad display. Assuming that the feedback from the previous ad display has arrived, the online gradient update unit updates the parameters of the router network according to the feedback gradient, and simultaneously sends a calibration signal to the dynamic feature fusion module.

[0287] Step S7: The sorting output unit 703 of the multi-objective optimization decision module 700 obtains 10 advertisements from the candidate advertisement pool (the candidate set must contain at least 2 advertisements). The multi-dimensional index calculation unit 701 calculates the predicted click-through rate correction value, expected revenue value, and negative feedback risk value for each advertisement. Assume that an electronics product advertisement has a high expected revenue (high bid) but also a high negative feedback risk (users have repeatedly closed similar advertisements), while a food and beverage advertisement has a moderate expected revenue but extremely low negative feedback risk. The multi-objective optimization unit 702 sets weight coefficients according to the current platform operation strategy (e.g., ...). The overall score of each advertisement is calculated. Ultimately, the food and beverage advertisement's overall score surpasses that of the electronics advertisement, and it is selected as the final recommended advertisement. The sorting output unit 703 outputs the recommendation result to the advertising delivery terminal, and the food and beverage advertisement is successfully displayed in the ad slot.

[0288] Example 2

[0289] This embodiment is basically the same as Embodiment 1, with the main difference being the design of the adaptive route prediction network.

[0290] In Example 1, four parallel expert subnetworks were set up. In practical applications, the number and type of expert subnetworks can be expanded according to the needs of the business scenario. For example, in a scenario where video advertising is the main delivery method, a dedicated expert subnetwork for video content features can be added, using a three-dimensional convolutional network (3D CNN) to process video frame sequence features.

[0291] Furthermore, in addition to selecting experts based on input features, the router network in this embodiment also introduces a context-based adaptive expert activation threshold adjustment mechanism. Specifically, the entropy value of the selection probability distribution p output by the router network... It is used to assess the uncertainty of the current prediction task. When the entropy value is high (i.e., the probabilities of multiple experts' choices are relatively close), it indicates that there is no absolutely dominant expert in the current scenario. In this case, the threshold τ of sparse activation is appropriately lowered to activate more experts to participate in the fusion, thereby improving the robustness of the prediction. When the entropy value is low (i.e., the probability of a certain expert is significantly dominant), the threshold τ is increased to encourage sparser activation and improve computational efficiency. This dynamic threshold adjustment method further enhances the intelligence and scenario adaptability of routing decisions.

[0292] Example 3

[0293] This embodiment is basically the same as Embodiment 1, with the main difference being the design of the multi-objective optimization decision module.

[0294] In Example 1, multi-objective optimization employs a weighted summation method. This example uses a more advanced Pareto front search method. Specifically, for a set of candidate ads, the system first generates target points for all candidate ads in a target space encompassing three dimensions: predicted click-through rate, expected revenue, and negative feedback risk. Then, a non-dominated ranking algorithm (such as the fast non-dominated ranking method in the NSGA-II framework) is used to find the Pareto front formed by this set of points (i.e., a set of ads where no other ad is inferior to the current ad in all metrics and superior to the current ad in at least one metric). Finally, the system selects an ad from the Pareto front for deployment based on the operational strategy.

[0295] The advantage of the Pareto front search method is that it does not require pre-setting the weight coefficients of the indicators, but directly outputs all non-dominated solutions (i.e., the Pareto optimal solution set), from which the ad delivery platform selects based on the actual situation. This method avoids the bias that may be caused by subjectively setting weights and has a more solid theoretical foundation for optimization.

[0296] Example 4

[0297] This embodiment is basically the same as Embodiment 1, with the main difference being the configuration of the biometric feature collection type.

[0298] In different advertising scenarios, the difficulty of collecting different types of biometric information and their predictive value vary. This embodiment provides a configurable biometric collection scheme, allowing the system to adaptively enable or disable specific biometric collection channels according to different scenarios:

[0299] In native mobile applications, eye tracking (via the front-facing camera) and facial micro-expression recognition are relatively easy to implement and can be configured to be enabled by default.

[0300] In wearable device scenarios (such as smartwatches), collecting skin conductance signals is relatively convenient and can be prioritized.

[0301] In laboratory or specialized equipment scenarios (such as VR / AR glasses), high-precision eye tracking and EEG signal acquisition can be enabled;

[0302] In scenarios where user privacy is a strong concern, only non-visual biometrics (such as skin conductance response) can be collected, or the biometrics collection module can be completely disabled, reverting to the traditional CTR prediction mode.

[0303] This configurable design enables the system of the present invention to flexibly adapt to the needs of different scenarios and different user groups, and has good engineering practical value.

[0304] The embodiments described herein are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An advertising click-through rate prediction system, characterized in that, include: The data acquisition module (100) is used to collect user data of at least one user in real time. The user data includes user behavior data, user attribute data, contextual data and user biometric signal data. The user biometric signal data includes at least one type of biometric signal selected from the following: eye tracking signal, facial micro-expression signal, skin conductance response signal and electroencephalogram signal. A multimodal feature extraction module (200) is connected to the data acquisition module (100) and is used to extract features from the user behavior data, user attribute data and contextual data respectively to obtain behavior feature vectors, attribute feature vectors and contextual feature vectors, and to perform time-frequency domain joint analysis on the user biometric signal data to extract at least one biometric feature vector. A dynamic feature fusion module (300), connected to the multimodal feature extraction module (200), is used to input the behavioral feature vector, the attribute feature vector, the context feature vector and the at least one biometric feature vector into an adaptive attention fusion network, calculate the cross-attention weights between the multiple feature vectors through a multi-head self-attention mechanism, and generate a fused feature vector based on the cross-attention weights. The adaptive attention fusion network includes a learnable feature weight allocation layer (303) for dynamically adjusting the fusion weights of the multiple feature vectors according to the real-time biometric signals of different users at the prediction time. The graph neural network encoding module (400) is connected to the dynamic feature fusion module (300) and is used to construct a heterogeneous graph structure with the fused feature vector as node features, using at least one of the following relationships: historical interaction relationship between users and advertisements, social relationship between users, and content association relationship between advertisements. The heterogeneous graph structure is then encoded by a multi-layer graph neural network to obtain a graph embedding feature vector containing structural information. An adaptive route prediction module (500), connected to the graph neural network encoding module (400), is used to input the graph embedding feature vector into an adaptive route prediction network. The adaptive route prediction network includes multiple parallel expert prediction subnetworks and at least one router network (502). The router network (502) dynamically selects and activates at least one expert prediction subnetwork based on the real-time changes of the graph embedding feature vector and the user biometric signal data, and weights and fuses the outputs of the activated expert prediction subnetworks to generate a preliminary click-through rate prediction value. The real-time feedback calibration module (600) is connected to the data acquisition module (100) and the adaptive route prediction module (500) respectively. It is used to generate feedback tags based on the user's subsequent real click behavior data after the current advertisement exposure, and to perform online incremental updates on the decision parameters of the router network (502) of the adaptive route prediction module (500) based on the feedback tags. At the same time, it generates calibration signals to feed back to the dynamic feature fusion module (300) to adjust the learnable feature weight allocation layer (303) in the adaptive attention fusion network. A multi-objective optimization decision module (700), connected to the real-time feedback calibration module (600), is used to acquire at least two ad candidate sets and, based on the preliminary click-through rate prediction value and the calibration signal output by the real-time feedback calibration module (600), calculate a comprehensive placement value score for each ad in the at least two ad candidate sets. The comprehensive placement value score is calculated using a multi-objective optimization function based on at least two indicators among the predicted click-through rate, expected revenue, and user negative feedback risk. The module also sorts the ads in the at least two ad candidate sets according to the comprehensive placement value score and outputs an ad recommendation list.

2. The advertising click-through rate prediction system according to claim 1, characterized in that, The multimodal feature extraction module (200) includes: The biometric encoding unit (204) is used to preprocess and extract time-frequency domain features from the user's biometric signal data. Specifically, it extracts gaze duration, gaze count, pupil diameter change and saccade path features from the eye-tracking signal, extracts micro-expression temporal features from the facial micro-expression signal through a three-dimensional convolutional neural network, extracts time-domain statistical features and frequency-domain power spectrum features from the skin conductance level change curve from the skin conductance response signal, and extracts power spectral density and functional connectivity strength features from multiple brain region frequency bands from the electroencephalogram (EEG) signal. The feature alignment unit (205) is connected to the biometric encoding unit (204) and is used to align the biometric vector with other feature vectors in the temporal dimension and output the temporally aligned multimodal feature tensor.

3. The advertising click-through rate prediction system according to claim 1, characterized in that, The adaptive attention fusion network includes: The feature embedding layer (301) is used to map each input feature vector to an embedding space of the same dimension to obtain the embedding feature matrix; A multi-head self-attention layer (302) is connected to the feature embedding layer (301) and is used to perform multi-head self-attention calculation on the embedded feature matrix to generate attention-weighted feature representations. The multi-head self-attention layer (302) adopts a dimensionality reduction calculation mechanism to reduce computational complexity. The feature weight allocation layer (303), connected to the multi-head self-attention layer (302), is used to dynamically calculate the fusion weight vector of each modality feature based on the biometric state recognition result generated by the user's biometric signal data, wherein the biometric state recognition result is used to characterize the user's current level of attention concentration and emotional arousal level; The feature fusion layer (304) is connected to the feature weight allocation layer (303) and is used to perform weighted fusion on the attention-weighted feature representation according to the fusion weight vector to generate a fusion feature vector.

4. The advertising click-through rate prediction system according to claim 1, characterized in that, The graph neural network encoding module (400) includes: Heterogeneous graph construction unit (401) is used to construct a heterogeneous graph structure based on user nodes, advertising nodes and context nodes. The types of edges in the heterogeneous graph include multiple types such as user history click edges, user-ad exposure edges, advertising content similarity edges and user social relationship edges. The node feature initialization unit (402) is connected to the heterogeneous graph construction unit (401) and is used to use the fused feature vector output by the dynamic feature fusion module (300) as the initial feature vector of the user node in the heterogeneous graph. The neighbor aggregation unit (403) is connected to the node feature initialization unit (402) and is used to perform type-aware neighbor aggregation operation according to the type of edge through each layer of the multi-layer graph neural network, wherein different types of edges use different aggregation functions and attention weights; The pooling encoding unit (404), connected to the neighbor aggregation unit (403), is used to output the user node embedding vector output from the last layer of the multilayer graph neural network as the graph embedding feature vector.

5. The advertising click-through rate prediction system according to claim 1, characterized in that, The adaptive route prediction network includes: A state encoder (501) is used to jointly encode the temporal changes of the graph embedding feature vector and the user biometric signal data to generate a routing state vector; The router network (502), connected to the state encoder (501), includes a gating mechanism network for calculating the selection probability of each of the plurality of parallel expert prediction subnetworks according to the routing state vector, and activating at least one expert prediction subnetwork based on the selection probability, wherein a single expert prediction subnetwork is activated when the maximum selection probability is greater than a preset threshold, otherwise multiple expert prediction subnetworks are activated. Multiple parallel expert prediction subnetworks, each with different network structures and parameter configurations, are used to handle different types of prediction scenarios. An output fusion layer (504), connected to the router network (502) and the plurality of parallel expert prediction subnetworks, is used to perform a weighted summation of the outputs of the activated expert prediction subnetworks according to the selection probabilities to generate a preliminary click-through rate prediction value.

6. The advertising click-through rate prediction system according to claim 1, characterized in that, The real-time feedback calibration module (600) includes: The feedback data acquisition unit (601) is used to collect the user's actual click behavior data within a preset time window after the current advertisement is exposed. The actual click behavior data includes whether a click was made and the time when the click occurred. An online gradient update unit (602) is connected to the feedback data acquisition unit (601) and is used to calculate the loss gradient based on the difference between the actual click behavior data and the preliminary click rate prediction value, and to use the loss gradient to perform online gradient descent update on the parameters of the router network (502), wherein the online gradient update is performed in real time after each feedback arrives without retraining the entire prediction network. A calibration signal generation unit (603) is connected to the online gradient update unit (602) and is used to generate a calibration signal according to the sign and magnitude of the loss gradient. The calibration signal contains adjustment direction and adjustment step size indication information for the feature weight allocation layer (303) in the dynamic feature fusion module (300). The delayed feedback processing unit (604) is used to cache the delayed click feedback data that arrives after the preset time window, and perform parameter updates in batches after the cache quantity threshold is met.

7. The advertising click-through rate prediction system according to claim 1, characterized in that, The multi-objective optimization decision module (700) includes: The multi-dimensional indicator calculation unit (701) is used to calculate the predicted click-through rate correction value based on the preliminary click-through rate prediction value and the calibration signal for each candidate advertisement, calculate the expected revenue value based on the advertiser's bidding information and the advertisement type, and calculate the user negative feedback risk value based on the user's historical negative feedback data. A multi-objective optimization unit (702) is connected to the multi-dimensional indicator calculation unit (701) and is used to comprehensively calculate the predicted click-through rate correction value, the expected revenue value and the user negative feedback risk value through a preset multi-objective optimization function to obtain the comprehensive placement value score of each candidate advertisement. The multi-objective optimization function adopts the Pareto optimization strategy, which allows the advertiser to dynamically adjust the weight coefficient of each indicator according to different business objectives. The sorting output unit (703) is connected to the multi-objective optimization unit (702) and is used to sort the candidate advertisements from high to low according to the comprehensive placement value score, select at least one advertisement with the highest score to generate an advertisement recommendation list and output it to the advertisement placement terminal.

8. The advertising click-through rate prediction system according to claim 1, characterized in that, The system also includes: The distributed deployment architecture includes at least one central server and multiple edge computing nodes. The data acquisition module (100) and the multimodal feature extraction module (200) are deployed on the edge computing node side to complete the acquisition and preliminary feature extraction processing of user biometric signals at the near end of the user terminal. The dynamic feature fusion module (300), the graph neural network encoding module (400), the adaptive route prediction module (500), the real-time feedback calibration module (600), and the multi-objective optimization decision module (700) are deployed on the central server side. The privacy protection unit, deployed on the edge computing node side, is used to perform localized differential privacy processing on the collected user biometric signal data. After feature extraction is completed, the original biometric signal data is deleted, and only the feature vector is encrypted and transmitted to the central server.

9. The advertising click-through rate prediction system according to claim 1, characterized in that, The data acquisition module (100) includes: The behavioral data collection unit is used to collect in real time the user's historical click records, browsing duration, search query terms, purchase records and interaction operation sequences on at least one Internet platform; The attribute data acquisition unit is used to obtain the user's age, gender, region, device type, and long-term interest tags; The context data acquisition unit is used to collect real-time information on the time, location, weather, network status, and page context of the current advertising display scenario. A biometric acquisition unit is used to acquire user biometric signal data through at least one sensing device configured on a user terminal device, wherein the sensing device includes one or more of a camera, an eye-tracking sensor, a skin conductance sensor, and an electroencephalogram (EEG) signal acquisition device.

10. The advertising click-through rate prediction system according to claim 5, characterized in that, The multiple parallel expert prediction subnetworks include: The first expert subnetwork, using a deep factorization machine architecture, is used to handle click-through rate prediction in scenarios with sparse discrete features. The second expert subnetwork, employing a deep cross-network architecture, is used to handle click-through rate prediction in high-order feature interaction scenarios. The third expert subnetwork, employing a sequence transformer architecture, is used to handle click-through rate prediction in user behavior sequence scenarios. The fourth expert subnetwork adopts a multi-gated hybrid expert network architecture to handle click-through rate prediction in multiple tasks and scenarios. The router network (502) dynamically selects and activates one or more of the first to fourth expert sub-networks according to the routing state vector, and performs weighted fusion of the outputs of the activated expert sub-networks.

Images