User intention prediction and accurate advertisement putting system based on deep learning

Abstract

The present application relates to the technical field of deep learning, and more particularly to a user intention prediction and accurate advertisement delivery system based on deep learning, which comprises the following steps: aggregating cross-platform behavior logs, obtaining intention primitive streams, and providing a basis for subsequent intention prediction; constructing a heterogeneous graph, extracting adjacent node features, quantitatively analyzing and generating a graph, and expanding the potential audience range of advertisements; by integrating the user's current scene fingerprint feature vector, calculating the semantic coupling coefficient and processing the intention propagation path, obtaining the intention confidence, and making the advertisement delivery fit the user's actual needs; by identifying the high-confidence intention vector demand characteristics, extracting the matching material segments, generating personalized advertisement prototypes, and enhancing the relevance between the advertisement and the user; through the federal learning gateway broadcast to the cooperative media platform, each platform decodes and delivers, while protecting the user's privacy, realizing cross-domain collaborative delivery, integrating multi-platform resources, and giving consideration to safety and advertisement coverage effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a user intention prediction and accurate advertisement placement system based on deep learning. BACKGROUND

[0002] In the context of the current booming digital marketing, frontier technologies such as deep learning are constantly reshaping the pattern of the advertisement placement field, bringing new development opportunities to the advertising industry. However, in the current application of advertisement placement mode, there are still the following problems: traditional advertisement placement is often based on simple user tags or single platform behavior data to speculate user intentions, which is difficult to fully and accurately grasp the complex and changing needs of users; many advertisement placements use a one-size-fits-all approach, without fully considering the personalized needs and preferences of users, resulting in inaccurate matching of advertisement content to users; traditional advertisement placement ignores the scenario information of users, and the needs and behaviors of users in different scenarios may differ greatly; in the digital age, user behaviors are scattered across multiple platforms, but there are data barriers between different platforms, making it difficult to share and integrate data; as users' emphasis on privacy protection continues to increase, traditional advertisement placement methods may have privacy leakage risks in the process of data collection and use, causing users to resist advertisements. Therefore, the present application proposes a user intention prediction and accurate advertisement placement system based on deep learning. SUMMARY

[0003] The purpose of the present application is to solve the problems in the background art, and the user intention prediction and accurate advertisement placement system based on deep learning is proposed.

[0004] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0005] The user intention prediction and accurate advertisement placement system based on deep learning comprises an intention primitive stream extraction module, an intention propagation graph evolution module, an intention confidence scene adaptation module, an advertisement gene recombination generation module, and a privacy cross-domain collaborative placement module.

[0006] The intention primitive stream extraction module aggregates user cross-platform behavior logs; adopts a behavior primitive decomposition method to disassemble the original behavior sequence into basic behavior units, and then generates an intention primitive stream.

[0007] The intention propagation graph evolution module constructs a heterogeneous graph with users as nodes based on the intention primitive stream; extracts the intention propagation features of adjacent nodes from the heterogeneous graph through a time series graph convolution network, and finally generates an intention propagation graph.

[0008] Intention confidence scenario adaptation module: obtain intention propagation graph, inject user's current scene fingerprint feature vector into graph node in real time; calculate and analyze semantic coupling coefficient through attention gate mechanism, and process intention propagation path, output high confidence intention vector calibrated by scene;

[0009] Advertisement gene recombination generation module: identify demand features in high confidence intention vector, extract matching material segments from advertisement gene library; recombine gene segments through generative adversarial constraint mechanism to generate personalized advertisement prototype;

[0010] Privacy cross-domain collaborative delivery module: inject differential privacy noise layer into personalized advertisement prototype to generate noisy advertisement feature package; broadcast the feature package to cooperative media platforms through the federal learning gateway, and each platform decodes the feature package based on local user group and performs delivery.

[0011] Further, the intention primitive flow extraction module aggregates user cross-platform behavior logs to obtain a behavior sequence vector set; the original behavior sequence is decomposed into basic behavior units by using a behavior primitive decomposition method, and then the process of generating an intention primitive flow includes:

[0012] A unified data access layer is established, which is connected through API interfaces with each platform, real-time crawls behavior log data generated by users on different platforms, and associates user identity through device unique identifier: for search words, record the search string input by the user, search time and search platform; by burying points in the video playback page, record the time difference from starting to play to leaving, and obtain the video stay duration; by setting scroll event listening on the product detail page, record the proportion of the page position the user scrolls to and the total page length, and thus obtain the product page scroll depth; after preliminary cleaning of the crawled various log data to remove duplicate and error data, store them according to user ID to form a cross-platform behavior log set of each user;

[0013] Structural analysis is performed on the generated cross-platform behavior log set to convert unstructured data into machine-readable vector form: map search words to keyword IDs in the behavior feature dictionary and record their occurrence frequency and timestamp; discretize video stay duration into three categories of short, medium and long, and associate video content labels; convert product page scroll depth into continuous values of 0-1 and match product classification attributes; output the preprocessed behavior sequence vector set as the input of subsequent primitive decomposition;

[0014] Based on the pre-processed behavior sequence vector set, a behavior feature dictionary containing search term keywords, video content labels, and commodity classification attributes is constructed, and the original behavior sequence is analyzed: the core intent words in the user input are identified by using a TF-IDF keyword extraction algorithm, and the same core keywords appearing in continuous search operations are aggregated, and finally merged into one search primitive; the video watching behavior is divided according to the cosine similarity of the video content labels, and when the similarity of adjacent video labels is lower than a preset behavior judgment threshold, a new watching primitive is determined; the commodity browsing behavior boundary is determined according to the change of the commodity classification hierarchy, and when the user operation is monitored to switch from one classification to a different hierarchical classification, a new browsing primitive is generated; the decomposed basic behavior units are standardized and coded, and are given a unique primitive ID and a behavior type label, forming a structured intent primitive stream.

[0015] Further, the intent propagation graph evolution module includes the following steps:

[0016] Each user is taken as a node in the graph, and the node attributes are determined; the intent primitive streams of different users are analyzed by using an association rule mining algorithm, and if the same primitive appears in the primitive sequence of two users, or there is a semantically associated primitive, a connection between the user nodes is established; the edge type is divided according to the primitive association property, and the edges directly sharing the same primitive are marked as direct propagation edges, and the edges with semantic association are marked as indirect propagation edges; at the same time, a time attribute is added to each node to record the time distribution of the intent primitives, and the number of primitives is counted by hour.

[0017] Further, the intent propagation graph evolution module extracts the intent propagation features of adjacent nodes from the heterogeneous graph by using a time series graph convolution network to generate the intent propagation graph, and the process includes the following steps:

[0018] After the heterogeneous graph is constructed, the intensity of the intent propagation between users is further quantified: the co-occurrence frequency of the primitives in the intent primitive streams of any two user nodes is counted, and the co-occurrence frequency is the ratio of the co-occurrence number to the total number of primitives of the two users; a conditional probability model is used to calculate the conditional probability Q(u2|u1) of the occurrence of the intent primitive u2 of the user U2 after the occurrence of the intent primitive u1 of the user U1, as the primitive transfer probability from the user U1 to the user U2; the initial value of the edge weight is set as the weighted sum of the primitive co-occurrence frequency and the primitive transfer probability, the weight coefficient is optimized by using a grid search algorithm, and the optimal combination is determined; the edge weight updating mechanism is triggered by the intent propagation graph evolution module every hour: the co-occurrence frequency and the transfer probability are recalculated, and the edge weight is adjusted by using an exponential smoothing method, so that a complete heterogeneous graph is constructed;

[0019] The constructed heterogeneous graph is input into a time series graph convolution network, and is divided into multiple time series subgraphs according to a 1-hour time window, and an adjacency matrix and a node feature matrix are constructed; wherein each time series subgraph contains user nodes, edges and edge weights of the time period; the adjacency matrix and the node feature matrix of the input subgraph are input into the graph convolution layer, which aggregates the feature vectors of adjacent nodes for each node, and combines the edge weights to generate spatial features through weighted summation; the time convolution layer introduces a gating cycle unit to fuse the node features of adjacent time windows to generate embedding vectors containing space-time information; the embedding vectors are reduced through a fully connected layer to extract all intention propagation features representing the intention propagation patterns of adjacent nodes;

[0020] According to the intention propagation features output by the time series graph convolution network; adjust the weight of the edge according to the latest intention propagation feature; at the same time, based on the intention propagation feature, the user nodes are clustered and analyzed, and the user groups with similar intention propagation patterns are identified, and the users with similar intention propagation patterns are classified into a class to form different user groups; based on the adjusted edge weight, important connections are screened, and edges exceeding a preset filtering threshold are retained to generate an intention propagation graph, wherein nodes represent users or user groups, edges represent intention propagation paths, and the thickness of the edge represents the intention propagation strength.

[0021] Further, the intention confidence scene adaptation module obtains the intention propagation graph, and the process of injecting the current scene fingerprint feature vector of the user into the graph node in real time includes:

[0022] Build a real-time interaction interface with the intention propagation graph storage system; after obtaining the intention propagation graph, collect the current scene fingerprint feature vector of the user: through the SDK of the mobile device end or the Web end, collect the geographic location information in real time; use high-precision positioning technology combined with map data to convert the geographic location into a semantic scene label; use the encoding rule to convert it into a semantic scene label feature vector; at the same time, collect the device sensor data; convert it into a motion state feature vector through the encoding rule; splice the semantic scene label feature vector and the motion state feature vector to form a complete scene fingerprint feature vector; package all collected scene fingerprint feature vectors in a unified format, and inject them into the corresponding nodes of the intention propagation graph through a real-time data stream pipeline. Add scene dimension information to each node.

[0023] Further, the intention confidence scene adaptation module calculates and analyzes the semantic coupling coefficient through the attention gate mechanism, and processes the intention propagation path to output the scene-calibrated high-confidence intention vector.

[0024] The relationship between the scene fingerprint and the intent propagation path is further analyzed by building an attention gate architecture. The architecture takes the feature vectors of the scene fingerprint and the intent propagation path as input data streams. The node feature matrix and edge weight information of the target intent propagation path are obtained from the intent propagation graph. The feature vectors of all nodes on the path are mean-pooled or weighted aggregated to obtain the node aggregated feature vector. The edge weight feature vector is generated by combining the edge weight information of the path. The node aggregated feature vector and the edge weight feature vector are concatenated to form the intent propagation path feature vector. The scene fingerprint feature vector is obtained and input into the attention gate architecture along with the intent propagation path feature vector. In the attention gate architecture, the semantic coupling coefficients between the scene fingerprint and the intent propagation path feature vectors are calculated through an attention mechanism. At the same time, a dynamic activation threshold is set. For paths with a semantic coupling coefficient higher than the dynamic activation threshold, the confidence is weighted and diffused according to the numerical value of the semantic coupling coefficient. For low-coupling paths, noise interference is suppressed by reducing their weights, and finally a high-confidence intent vector calibrated by the scene is output.

[0025] Furthermore, the advertisement gene recombination generation module identifies the demand features in the high-confidence intent vector and extracts matching material segments from the advertisement gene library. The process of generating personalized advertisement prototypes by recombining gene segments through a generative adversarial constraint mechanism includes:

[0026] The high-confidence intent vector is probabilistically decomposed using a variational autoencoder, and the implicit demand features are extracted by sampling the latent space.

[0027] N core dimensions with strong association with user behavior are selected from the extracted implicit demand feature dimensions using mutual information entropy calculation. The core dimension set is obtained. Based on the core dimension set, all core dimensions are sorted in descending order according to the numerical value of the mutual information entropy, and the top N core dimensions are extracted.

[0028] A dynamic weight is assigned to each core dimension. The core dimensions are associated with industry terms using a knowledge graph to form a demand label system. Based on the formed demand label system, matching material segments, i.e., gene segments, are extracted from the advertisement gene library: copywriting genes, visual genes, and interactive genes.

[0029] A deep hashing algorithm is used to convert the demand labels into 128-bit binary codes.

[0030] A feature code vector is defined for the gene segment. The binary code of the demand label is matched with the feature code of the gene segment using the Hamming distance, and the material with a distance less than a preset Hamming distance threshold is preferentially returned. The matching results are secondarily screened, and the candidate gene segments are output.

[0031] Based on the screened candidate gene fragments of each dimension, the generator adopts hybrid attention to cross-modal encode the copywriting, visual, and interactive gene fragments; in the encoding process, self-attention is used to capture the internal correlation of gene fragments of the same type; cross-attention is used for semantic alignment of different types of gene fragments; multi-head attention is used to splice features of each modality to generate recombined gene fragments;

[0032] The discriminator sets double constraints, including a content discriminator and a scene discriminator; the recombined gene fragments are used as the generated samples of the discriminator; the content discriminator verifies the logical coherence of the generated samples and the demand label; the scene discriminator evaluates the adaptability of the generated samples to the current scene of the user; by continuously optimizing the generator and the discriminator, the personalized advertisement prototype is finally output.

[0033] Further, the privacy-preserving cross-domain collaborative delivery module injects a differential privacy noise layer into the personalized advertisement prototype to generate a noisy advertisement feature package; through the federal learning gateway, the feature package is broadcast to the cooperative media platforms, and each platform decodes the feature package based on the local user group and performs the delivery process, which includes:

[0034] Based on the privacy protection requirements of the advertisement delivery scene, the data sensitivity in different scenes is quantitatively evaluated, and the privacy budget parameter of differential privacy is set; the high-confidence intention vector is normalized and mapped to the numerical range [a, b];

[0035] According to the Laplace mechanism of differential privacy, a noise vector with the same dimension as the normalized high-confidence intention vector is generated; the Laplace mechanism samples a noise value from the Laplace distribution and adds it to each dimension of the normalized high-confidence intention vector;

[0036] The high-confidence intention vector with added noise is fused with the multi-modal advertisement features of the personalized advertisement prototype using a hierarchical encoding strategy to generate a noisy advertisement feature package: the bottom layer uses an autoencoder to compress the multi-modal features into a low-dimensional embedding; the middle layer dynamically adjusts the weight of the modal-level features through an attention mechanism; and the top layer applies differential privacy protection to the compressed features;

[0037] After generating the noisy advertisement feature package, a federal learning gateway is built to realize cross-platform collaborative delivery; the noisy advertisement feature package is encrypted and broadcast to each cooperative media platform through the federal learning gateway; after receiving the encrypted noisy advertisement feature package, each media platform uses a pre-set decryption key to decrypt it; according to the behavior feature data of the local user group, the noisy advertisement feature package is decoded; and each media platform performs the advertisement delivery task in the local user group according to the decoded advertisement feature package.

[0038] Compared with the prior art, the beneficial effects of the present application are: by aggregating cross-platform behavior logs and adopting behavior primitive decomposition, the user's behavior patterns in different scenarios can be comprehensively captured, complex behaviors can be decomposed into basic units, structured intent primitive streams can be generated, accurate and detailed user intent basic data can be provided for subsequent modules, and the accuracy of intent prediction can be improved; a heterogeneous graph with users as nodes is constructed, and a time sequence graph convolution network is used to extract features, the intent propagation relationship between users is intuitively presented, the intent propagation intensity and mode change are reflected in real time, and user groups and important propagation paths are identified, and potential user demand associations are mined; by injecting a scene fingerprint feature vector in real time and processing it through an attention gate mechanism, the intent confidence can be calibrated in combination with the current scene of the user, the fit between the intent prediction and the actual scene can be effectively improved, and the attractiveness and conversion rate of the advertisement can be improved; by extracting demand features through multiple technologies and recombining gene fragments to generate personalized advertisement prototypes, the user demand is matched, the advertisement content is highly customized, the relevance and targeting of the advertisement and the user are enhanced, and the user experience is improved; by injecting a differential privacy noise layer and using a federated learning gateway to deliver, cross-domain collaboration is realized on the premise of protecting user privacy, data security requirements are met, and multiple platform resources can be integrated to expand the range of advertisement delivery. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A module diagram of the user intent prediction and precise advertisement delivery system based on deep learning proposed by the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0041] Reference Figure 1 The user intent prediction and precise advertisement delivery system based on deep learning includes an intent primitive stream extraction module, an intent propagation graph evolution module, an intent confidence scene adaptation module, an advertisement gene recombination generation module, and a privacy-preserving cross-domain collaborative delivery module.

[0042] The intent primitive stream extraction module: aggregates user cross-platform behavior logs (search words, video dwell time, product page scroll depth); adopts a behavior primitive decomposition method to decompose the original behavior sequence into basic behavior units, and then generates an intent primitive stream.

[0043] The intent propagation graph evolution module: based on the intent primitive flow, a heterogeneous graph with users as nodes is constructed; the intent propagation features of adjacent nodes are extracted from the heterogeneous graph through a time series graph convolution network, and finally an intent propagation graph is generated;

[0044] The intent confidence scene adaptation module: obtain the intent propagation graph, and inject the user's current scene fingerprint feature vector (geographic location, device sensor data) into the graph node in real time; calculate and analyze the semantic coupling coefficient through the attention gate mechanism, and process the intent propagation path, output the high-confidence intent vector calibrated by the scene;

[0045] The advertisement gene recombination generation module: identifies the demand features in the high-confidence intent vector, extracts matching material segments (copywriting genes, visual genes, and interactive genes) from the advertisement gene library; recombines the gene segments through a generative adversarial constraint mechanism to generate a personalized advertisement prototype;

[0046] The privacy cross-domain collaborative delivery module: injects a differential privacy noise layer into the personalized advertisement prototype (protects the original high-confidence intent vector), generates a noisy advertisement feature package; through a federated learning gateway, the feature package is broadcast to cooperative media platforms, and each platform decodes the feature package based on the local user group and performs delivery.

[0047] It should be further explained that in the specific implementation process, the intent primitive flow extraction module aggregates user cross-platform behavior logs to obtain a behavior sequence vector set; the process of generating an intent primitive flow includes:

[0048] A unified data access layer is established, which is connected through the API interface of each platform (such as search engines, video platforms, and e-commerce platforms), real-time crawls the behavior log data generated by users on different platforms, and associates the user identity through a device unique identifier (such as IMEI, MAC address): for search words, record the user input search string, search time, and search platform; by setting a scroll event listener on the product detail page, record the proportion of the page position the user scrolls to and the total page length, and thus obtain the product page scroll depth; after preliminary cleaning of the crawled various log data to remove duplicate and error data, store them according to user ID to form a cross-platform behavior log set for each user;

[0049] Structural analysis is performed on the generated cross-platform behavior log set, unstructured data (such as free text search words, video stay duration values) is converted into machine-readable vector form: the search words are mapped to keyword IDs in the behavior feature dictionary, and the occurrence frequency and timestamp are recorded; the video stay duration is discretized into three categories: short (less than 30 seconds), medium (30-120 seconds), and long (more than 120 seconds), and the video content label is associated; the product page scroll depth is converted into a continuous value of 0-1, and the product classification attribute is matched; the pre-processed behavior sequence vector set is output as the input of the subsequent primitive decomposition;

[0050] Based on the pre-processed behavior sequence vector set, a behavior feature dictionary containing search keywords, video content labels, and product classification attributes is constructed, and the original behavior sequence is analyzed: the core intent words in the user input are identified by using the TF-IDF keyword extraction algorithm, and the same core keywords appearing in continuous search operations are aggregated, and finally combined into one search primitive; video viewing behavior is divided according to the cosine similarity of video content labels, and when the similarity of adjacent video labels is lower than the preset behavior judgment threshold, a new viewing primitive is determined; the boundaries of product browsing behavior are determined according to the changes of product classification levels, and when the user's operation is monitored to switch from one classification to a different level classification, a new browsing primitive is generated; the decomposed basic behavior units are standardized coded, and are given a unique primitive ID and a behavior type label, forming a structured intent primitive stream.

[0051] It should be further pointed out that, in the specific implementation process, the intent propagation graph evolution module constructs a heterogeneous graph with users as nodes based on the intent primitive stream; the intent propagation features of adjacent nodes are extracted from the heterogeneous graph by a time series graph convolution network, and finally the process of generating the intent propagation graph includes:

[0052] Each user is regarded as a node in the graph, and the node attributes are determined, including basic information such as user age and historical behavior preference data such as high-frequency access platform type; the intent primitive streams of different users are analyzed by using an association rule mining algorithm, and if the same primitive appears in the primitive sequence of two users, for example, both contain a notebook computer primitive, or there is a semantically associated primitive, for example, a smartphone and a phone case, a connection between the user nodes is established; the edge type is divided according to the primitive association property, and the edges directly sharing the same primitive are marked as direct propagation edges, and the edges with semantic association are marked as indirect propagation edges; at the same time, a time attribute is added to each node to record the time distribution of the intent primitive, and the number of primitives is counted by hour;

[0053] After the construction of the heterogeneous graph is completed, the intention propagation strength between users is further quantified to provide a data basis for edge weight calculation: the co-occurrence frequency of the intention primitives in the intention primitive stream of any two user nodes is counted, and the co-occurrence frequency is the ratio of the co-occurrence number to the total number of intention primitives of the two users; the conditional probability model is used to calculate the conditional probability Q(u2|u1) of the occurrence of the intention primitive u2 of the user U2 after the occurrence of the intention primitive u1 of the user U1, as the primitive transfer probability from the user U1 to the user U2; the initial value of the edge weight is set as the weighted sum of the co-occurrence frequency and the primitive transfer probability, the weight coefficient is optimized by the grid search algorithm, and the optimal combination is determined; wherein the optimal combination refers to determining the weight coefficients of the co-occurrence frequency and the primitive transfer probability when calculating the initial value of the edge weight, so as to achieve the most accurate goal of the edge weight; the edge weight updating mechanism is triggered by the intention propagation graph evolution module every hour: the co-occurrence frequency and the transfer probability are recalculated, and the exponential smoothing method is used to adjust the edge weight, giving higher weight to the data within 1 hour, so that the edge weight can reflect the strength change of the intention propagation between users in real time; the updated edge weight reflects the change of the intention propagation strength between users in real time, thereby constructing a complete heterogeneous graph;

[0054] The constructed heterogeneous graph is input into the time series graph convolution network, and is divided into multiple time series subgraphs according to a 1-hour time window, and an adjacency matrix (based on node connection relationship) and a node feature matrix (based on node attribute preprocessing) are constructed; wherein each time series subgraph contains user nodes, edges and edge weights in the time period; the adjacency matrix is constructed according to the connection relationship between the user nodes in the time series subgraph (i.e. the existence of the edge); the node feature matrix is obtained by preprocessing the determined node attributes (such as user age, gender, basic information and historical behavior preference data); the adjacency matrix and the node feature matrix of the input subgraph are input into the graph convolution layer, which aggregates the feature vectors of adjacent nodes for each node, and combines the edge weight to generate spatial features through weighted summation; the time convolution layer introduces a gated recurrent unit to process the feature changes of different time series subgraphs, fuse the node features of adjacent time windows, and generate embedding vectors containing spatio-temporal information; the embedding vectors are reduced through the fully connected layer, and the intention propagation features representing the intention propagation patterns of adjacent nodes are extracted;

[0055] According to the intention propagation features output by the time sequence diagram convolution network, the features reflect the intention propagation mode between adjacent nodes, and provide a basis for edge weight adjustment and user clustering; the weight of the edge is adjusted according to the latest intention propagation features, so that it can better reflect the strength and direction of the intention propagation; at the same time, the user nodes are clustered and analyzed based on the intention propagation features, the user groups with similar intention propagation modes are identified, the users with similar intention propagation modes are classified into a category, and different user groups are formed; important connections are screened based on the adjusted edge weight, and edges exceeding a preset filtering threshold (the filtering threshold is used to screen edges that can reflect the core intention propagation relationship and have a certain strength of the edge weight, and the edge weight value limit is set to highlight the key propagation path) are retained, and then an intention propagation graph is generated, wherein the nodes represent users or user groups, the edges represent intention propagation paths, and the thickness of the edges represents the intention propagation strength; at the same time, the intention propagation graph evolution module generates a graph snapshot every 24 hours, compares different time point snapshots, analyzes the user group clustering changes, the core path migration and the edge weight fluctuation; at the same time, the intention propagation graph is added with an interactive attribute, so that the intention primitive flow details can be viewed by clicking the node, and the time sequence evolution can be observed by dragging the time axis, and the visualization of the intention propagation rule is further realized.

[0056] It needs to be further explained that in the specific implementation process, the intention confidence scene adaptation module obtains the intention propagation graph, and injects the user's current scene fingerprint feature vector into the graph node in real time; the process of calculating and analyzing the semantic coupling coefficient through the attention gate mechanism and processing the intention propagation path to output the high-confidence intention vector calibrated in the scene includes:

[0057] A real-time interactive interface with the intention propagation graph storage system is built to ensure that the latest graph data can be quickly obtained, and the graph data obtained through the interface will be used as the basis for subsequent injection of the user's current scene fingerprint; after obtaining the intention propagation graph, the user's current scene fingerprint feature vector is collected: through the SDK of the mobile device end or the Web end, the geographic location information is collected in real time; by using high-precision positioning technology combined with map data, the geographic location is converted into a semantic scene label, such as a shopping mall; by using coding rules (such as one-hot encoding, word embedding, etc.), it is converted into a semantic scene label feature vector; at the same time, device sensor data such as acceleration sensor and gyroscope data are collected to determine the user's motion state; through the coding rule, it is converted into a motion state feature vector; the semantic scene label feature vector and the motion state feature vector are spliced to form a complete scene fingerprint feature vector; all collected scene fingerprint feature vectors are packaged in a unified format, and injected into the corresponding nodes of the intention propagation graph through a real-time data stream pipeline, and the scene dimension information is added to each node, so that the graph nodes can reflect the user's intention state in different scenes;

[0058] By building an attention gate architecture, the relationship between the scene fingerprint and the intent propagation path is further analyzed; wherein the architecture takes the feature vectors of the scene fingerprint and the intent propagation path as input data streams; the node feature matrix and edge weight information (the weighted sum of the element co-occurrence frequency and the transition probability) of the target intent propagation path are obtained from the intent propagation graph; the feature vectors of all nodes on the path are mean-pooled or weighted aggregated (the weight is the edge weight), to obtain the node aggregated feature vector; the edge weight feature vector is generated in combination with the edge weight information (the weight of the direct propagation edge and the indirect propagation edge) of the path; it can be understood that the edge weight information includes two types of data: the numerical level edge weight information is the weighted sum of the element co-occurrence frequency and the transition probability, which is used to quantify the propagation strength); the type level edge weight information is the classification of the direct propagation edge and the indirect propagation edge, which is used for semantic coupling analysis; the node aggregated feature vector and the edge weight feature vector are spliced to form the intent propagation path feature vector; the scene fingerprint feature vector is obtained, and the scene fingerprint feature vector and the intent propagation path feature vector are input into the attention gate architecture; in the attention gate architecture, the semantic coupling coefficients between the scene fingerprint and each intent propagation path feature vector are calculated through the attention mechanism, wherein the purpose of calculating the semantic coupling coefficients is to measure the semantic association degree between the scene fingerprint and the intent propagation path, and the role of the intent propagation path in the intent propagation is determined according to the semantic association degree; at the same time, a dynamic activation threshold is set; for the paths with a semantic coupling coefficient higher than the dynamic activation threshold, the confidence is weighted and diffused according to the numerical size of the semantic coupling coefficient, to enhance the role of these intent propagation paths in the intent propagation; for the low coupling paths, noise interference is suppressed by reducing the weight, and finally the high-confidence intent vector after scene calibration is output.

[0059] It needs to be further explained that in the specific implementation process, the advertisement gene recombination generation module identifies the demand features in the high-confidence intent vector, extracts the matching material segments from the advertisement gene library; the process of generating personalized advertisement prototypes by generating adversarial constraint mechanism to recombine gene segments includes:

[0060] The high-confidence intent vector is probabilistically decomposed by using a variational autoencoder, and the implicit demand features are extracted by sampling the latent space (wherein the implicit demand features are the latent features obtained by probabilistically decomposing the high-confidence intent vector by the variational autoencoder VAE, and the essence is the deep demand in the user intent which is not explicitly expressed but can be identified by data mining): obtaining a high-confidence intent vector x e R D , wherein D is the original dimension of the high-confidence intent vector, and R is the real number set; the high-confidence intent vector x e R D is decomposed into implicit demand features Z e R l (wherein l is the dimension of the implicit demand features, and l < D), and the specific steps are: the encoder maps the input high-confidence intent vector x to the latent variable distribution q φ(Z|x) (φ is the encoder parameter), output mean μ φ (x) and log variance

[0061] μ φ = W μ · x + κ μ ,

[0062] where W μ , W σ ∈ R l×D is the encoder weight matrix, κ μ , κ σ ∈ R l is the encoder bias vector;

[0063] The decoder reconstructs the input high-confidence intention vector x from the latent variable Z ~ q φ (Z|x) (φ is the encoder parameter), output mean μ ψ (Z) :

[0064] μ ψ = W ψ · Z + κ ψ ,

[0065] where W ψ ∈ R l×D is the decoder weight matrix, κ ψ is the decoder bias vector, and ψ is the decoder parameter;

[0066] The reconstruction loss is combined with the KL divergence to constrain the latent variable distribution to be close to the prior distribution p(Z) = M(0, I) (M(0, I) is a normal distribution, 0 is the mean vector, and I is the identity matrix):

[0067]

[0068] where Lo is the loss function, E is the expectation operator, Z ~ q φ represents that the random variable Z follows the conditional distribution generated by the encoder q φ , and β is the KL divergence weight hyperparameter used to control the sparsity of the latent variable;

[0069] Using mutual information entropy calculation, N core dimensions are selected from the l implicit demand feature dimensions that are strongly associated with user behavior: define the user behavior feature y ∈ R m , where m is the user behavior feature dimension, such as clicks, dwell time, etc.; the correlation strength is measured by calculating the mutual information entropy between the implicit demand feature and the user behavior vector:

[0070] MI(Z; y) = H(Z) + H(y) + H(Z, y),

[0071] In the formula, H is the information entropy, and H(Z,y) is the joint entropy;

[0072] Then obtain the core dimension set ζ = {dim1, dim2, ..., dim N}, where dim1, dim2, ..., dim N These are the 1st core dimension, the 2nd core dimension, ..., the Nth core dimension, where N is the number of core dimensions. Based on the set of core dimensions, all core dimensions are sorted in descending order according to their mutual information entropy values, and the first N core dimensions (e.g., 8) are extracted.

[0073] dim for each core dimension i Assign dynamic weights w to ∈ζ (where i is the core dimension index) i (τ) (where τ is the time step, representing the dynamic update timing of the weight), reflecting its importance changes over time or user behavior: based on the core dimension dim i Recent contribution γ i (τ) (e.g., the click-through rate increase of the core dimension in the most recent T ad generation) and historical weight w i (τ-1), updated using exponential smoothing:

[0074] w i (τ)=χ·w i (τ-1)+(1-χ)·γ i (τ),

[0075] In the formula, χ is a smoothing coefficient used to control the decay rate of historical weights;

[0076] By linking core dimensions with industry terminology through a knowledge graph, a demand tagging system is formed: a heterogeneous knowledge graph G = (v, e) is used to represent the association between core dimensions and industry terminology, where the set of graph nodes v = v dim ∪v ter (v dim v is the core dimension node set. ter (This is a set of industry terminology nodes), where the graph edge e represents the correlation strength between the core dimension and the industry terminology; for each core dimension dim i By aggregating information from neighboring industry terminology nodes using a graph neural network, a demand label t is generated. i :

[0077]

[0078] In the formula, h is the embedding vector of the core dimension node. ter This is the embedding vector of the industry term node, belonging to industry term node v. ter Elements in the set; Λ(dimi ) is a set of neighbor industry term nodes of the core dimension; θ is a graph neural network parameter;

[0079] It can be understood that by forming the demand label system, the abstract demand dimension can be converted into specific and measurable indicators, facilitating subsequent extraction of matching material segments in the advertisement gene library. Meanwhile, the demand label system can automatically switch between 5-8 basic labels (coarse granularity) and 15-20 subdivided labels (fine granularity) according to real-time computing resources and delivery precision requirements, balancing retrieval efficiency and matching accuracy through dynamic granularity adjustment.

[0080] Based on the formed demand label system, matching material segments, i.e., gene segments g j (where j is the gene segment index): The gene library adopts a distributed storage architecture, the copywriting gene is classified according to the sentiment tendency (positive, neutral, negative) and semantic theme, the visual gene contains hierarchical vector materials (splittable product components, scene background, decorative elements), and the interactive gene is encapsulated as a nestable dynamic component (such as a floating pop-up window, a progress bar animation);

[0081] The demand label t i is converted into a 128-bit binary code Φ i :

[0082]

[0083] In the formula, sign is a sign function, W h is a hash network weight matrix, κ h is a hash network bias vector, is an embedding vector of the demand label;

[0084] The feature code vector Φ j of the gene segment is defined; the binary code of the demand label and the feature code of the gene segment are matched in Hamming distance, and the material with a distance less than a preset Hamming distance threshold J is preferentially returned:

[0085]

[0086] In the formula, Dis(Φ j ,Φ j ) is the Hamming distance, r is the binary code bit index, and Π is an indicator function (1 if the condition is true, otherwise 0);

[0087] The matching results are secondarily screened, and candidate gene segments are output, wherein 5-8 candidate gene segments are retained for each dimension;

[0088] Based on the screened candidate gene fragments of each dimension, the generator adopts mixed attention to encode the copywriting, visual and interactive gene fragments in a cross-modal manner; in the encoding process, self-attention is used to capture the internal correlation of gene fragments of the same type, i.e., to calculate the self-attention matrix of gene fragments of the same type (such as the set of copywriting gene fragments); cross-attention is used to perform semantic alignment on gene fragments of different types, i.e., to calculate the cross-attention of gene fragments of different types (such as copywriting gene fragments and visual gene fragments); and multi-head attention is used to splice the features of each modality to generate the recombined gene fragments g rec ;

[0089] The discriminator sets double constraints, including a content discriminator and a scene discriminator; the recombined gene fragments g rec are taken as the generated samples of the discriminator; the content discriminator verifies the logical coherence between the generated samples and the demand label t:

[0090]

[0091] wherein, Lcon is the loss function of the content discriminator; E is an expectation operator; (g rec ,t) ~ p real represents that the random variable pair (g rec ,t) obeys the real data distribution p real , i.e., the joint distribution of the real advertisement prototype and its corresponding demand label; represents the expectation of the random variable pair (g gen ,t) under the sample distribution p rec generated by the generator; Dcon is the content discriminator (input is the pair of gene fragments and demand labels);

[0092] The scene discriminator evaluates the adaptability of the generated samples to the current scene us of the user:

[0093]

[0094] wherein, Lsce is the loss function of the scene discriminator; represents the expectation of the random variable pair (g gen ,us) under the real data distribution p rec ; represents the expectation of the random variable pair (g gen ,us) under the sample distribution p rec generated by the generator;

[0095] The total loss function Ltot is used to balance the content and scene verification: wherein, is the weight hyperparameter of the scene discriminator;

[0096] By continuously optimizing the generator and discriminator, the combination of gene fragments generated by the generator can simultaneously meet the requirements of the content discriminator and the scene discriminator, ultimately outputting a personalized advertising prototype.

[0097] It should be further explained that, in the specific implementation process, the privacy-enhancing cross-domain collaborative delivery module injects a differential privacy noise layer into the personalized advertising prototype to generate a noisy advertising feature package; the feature package is then broadcast to the cooperating media platforms through a federated learning gateway, and each platform decodes the feature package based on its local user group and executes the delivery process, which includes:

[0098] Based on the privacy protection needs of advertising scenarios, this paper quantitatively assesses the data sensitivity under different scenarios and sets the privacy budget parameter for differential privacy: The privacy protection level PL is defined, and the sensitivity index SI is used to measure the data sensitivity (a higher SI value indicates more sensitive data, SI∈{1,2,3}, where 1 is low sensitivity, 2 is medium sensitivity, and 3 is high sensitivity). The privacy budget parameter for differential privacy is denoted as ε, and the formula for setting the privacy budget parameter is as follows:

[0099]

[0100] In the formula, α is the adjustment coefficient for the actual scenario; where the global sensitivity of the high-confidence intent vector is determined through sensitivity analysis;

[0101] The high-confidence intent vector is normalized and mapped to the numerical range [a, b] to better incorporate noise: Let the high-confidence intent vector be V = (υ1,υ2,…,υ d ), where υ1,υ2,…,υ d Let be the 1st, 2nd, ..., dth elements of the high-confidence intent vector, respectively; where d is the dimension of the high-confidence intent vector; determine the maximum value max(V) and minimum value min(V) of the high-confidence intent vector V; normalize each element in the high-confidence intent vector V to obtain the normalized high-confidence intent vector. All its elements are in the range [a, b] (e.g., [0, 1]), and the normalization formula is: In the formula, c is the index of an element in the vector;

[0102] Based on the Laplace mechanism of differential privacy, a high-confidence intent vector is generated and normalized. Noise vectors n of the same dimension L =(n L(1) ,n L(2) ,…,n L(d) (where n) Lc ~Lap(0,η)):

[0103] In the formula, η is the noise scaling parameter, Δf is the sensitivity of the high-confidence intent vector, ε is the privacy budget parameter, and d is the dimension of the high-confidence intent vector; the Laplace mechanism samples noise values ​​from the Laplace distribution and adds them to each dimension of the normalized high-confidence intent vector:

[0104] A hierarchical encoding strategy is employed to fuse the high-confidence intent vector with added noise into the multimodal advertising features of the personalized advertising prototype (where multimodal advertising features are the other multimodal features in the personalized advertising prototype besides the high-confidence intent vector), generating a noisy advertising feature package: the bottom layer uses an autoencoder to compress the multimodal features into low-dimensional embeddings; the middle layer dynamically adjusts the modal-level feature weights through an attention mechanism (e.g., visual elements have increased weights in beauty ads); the top layer applies differential privacy protection to the compressed features, i.e., recalculates the sensitivity and adds noise; understandably, generating a noisy advertising feature package through encoding ensures that the core information of the advertisement is preserved while protecting the privacy of the original intent vector.

[0105] After generating the noisy ad feature package, a federated learning gateway is built to achieve cross-platform collaborative delivery. An encrypted communication protocol is used to ensure the security of data transmission. The noisy ad feature package is encrypted and then broadcast to various partner media platforms through the federated learning gateway. Upon receiving the encrypted noisy ad feature package, each media platform decrypts it using a pre-defined decryption key. Based on the behavioral characteristics data of the local user group, the noisy ad feature package is decoded. During decoding, the parameters of the local model are used to adapt and adjust the noisy ad feature package to better suit the preferences and needs of local users. Each media platform executes ad delivery tasks within its local user group based on the decoded ad feature package, while simultaneously recording relevant data during the delivery process, such as impressions and click-through rates, to provide a basis for subsequent evaluation and optimization.

[0106] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0107] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0108] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1.A deep learning-based user intent prediction and accurate advertisement placement system, characterized by comprising: The module comprises an intention primitive flow extraction module, an intention propagation graph evolution module, an intention confidence scene adaptation module, an advertisement gene recombination generation module, and a privacy cross-domain collaborative delivery module. The intention primitive flow extraction module aggregates user cross-platform behavior logs, and uses a behavior primitive decomposition method to decompose original behavior sequences into basic behavior units, and then generates intention primitive flow. The intention primitive flow extraction module connects each platform API through a data access layer, obtains user cross-platform behavior logs and associates user identities, cleans and structures the user cross-platform behavior logs, and converts them into vector form to output a preprocessed behavior sequence vector set, divides search primitives, viewing primitives and browsing primitives based on the behavior sequence vector set to form basic behavior units, and encodes the basic behavior units into intention primitive flow. The intention propagation graph evolution module constructs a heterogeneous graph with users as nodes based on the intention primitive flow, extracts intention propagation features of adjacent nodes from the heterogeneous graph through a time series graph convolution network, and finally generates an intention propagation graph. The intention propagation graph evolution module constructs a heterogeneous graph with users as nodes, determines node attributes, analyzes intention primitive flow of different users, divides edge types including direct propagation edges and indirect propagation edges, adds time attributes to each node, and forms a heterogeneous graph containing node attributes, edge types and time attributes. The intention confidence scene adaptation module obtains the intention propagation graph, injects a user's current scene fingerprint feature vector into the graph node in real time, calculates and analyzes semantic coupling coefficients through an attention gate mechanism, processes the intention propagation path, and outputs a high-confidence intention vector calibrated in the scene. The advertisement gene recombination generation module identifies demand features in the high-confidence intention vector, extracts matching material segments from an advertisement gene library, and generates personalized advertisement prototypes by recombining gene segments through a generative adversarial constraint mechanism. The privacy cross-domain collaborative delivery module injects differential privacy noise into the personalized advertisement prototype to generate a noisy advertisement feature package, and broadcasts the feature package to cooperative media platforms through a federated learning gateway. Each platform decodes the feature package based on the local user group and performs delivery. 2.The deep learning based user intent prediction and accurate advertisement launching system according to claim 1, characterized in that: The intention primitive flow extraction module aggregates user cross-platform behavior logs, and obtains a behavior sequence vector set. The process of using a behavior primitive decomposition method to decompose original behavior sequences into basic behavior units and then generating intention primitive flow includes: A unified data access layer is established to connect with the API interfaces of each platform, and user behavior log data generated on different platforms is real-time scraped, and the user identity is associated through a device unique identifier: for search words, record the user input search string, search time, and search platform; by setting a scroll event listener on the product detail page, record the proportion of the page position the user scrolls to and the total page length to obtain the product page scroll depth; after preliminary cleaning of the scraped log data of various types to remove duplicate and error data, the data is classified and stored according to user ID to form a cross-platform behavior log set for each user. The generated cross-platform behavior log set is structurally parsed to convert unstructured data into machine-readable vector form: search words are mapped to keyword IDs in a behavior feature dictionary, and their occurrence frequencies and timestamps are recorded; video stay duration is discretized into three categories of short, medium, and long, and associated with video content labels; the product page scroll depth is converted into a continuous value of 0-1, and matched with product classification attributes; and the pre-processed behavior sequence vector set is output as input for subsequent primitive decomposition; Based on the pre-processed behavior sequence vector set, a behavior feature dictionary containing search keywords, video content labels, and product classification attributes is constructed to analyze the original behavior sequence: core intent words in user input are identified by using a TF-IDF keyword extraction algorithm, and the same core keywords appearing in continuous search operations are aggregated, ultimately merging into one search primitive; video viewing behavior is divided according to the cosine similarity of video content labels, and when the similarity of adjacent video labels is lower than a preset behavior judgment threshold, a new viewing primitive is determined; the boundaries of product browsing behavior are determined according to the changes in product classification hierarchy, and when a user switches from one classification to a different hierarchical classification through monitoring user operations, a new browsing primitive is generated; the decomposed basic behavior units are standardized and coded, with unique primitive IDs and behavior type labels assigned, forming a structured intent primitive stream. 3.The deep learning based user intent prediction and accurate advertisement launching system according to claim 1, characterized in that: The intent propagation graph evolution module constructs a heterogeneous graph with users as nodes based on the intent primitive stream, including: Each user is treated as a node in the graph, and node attributes are determined; different users' intent primitive streams are analyzed by using an association rule mining algorithm, and if the same primitive appears in the primitive sequence of two users, or there is a semantically associated primitive, a connection between the user nodes is established; the edge type is divided according to the primitive association property, and the edge type is marked as a direct propagation edge if the same primitive is directly shared, and as an indirect propagation edge if there is a semantic association; the time attribute is added to each node to record the time distribution of the intent primitive, and the number of primitives is counted by hour. 4.The deep learning based user intent prediction and accurate advertisement launching system according to claim 1, characterized in that: The intent propagation graph evolution module extracts the intent propagation features of adjacent nodes from the heterogeneous graph through a time series graph convolution network to generate the intent propagation graph, including: After the heterogeneous graph is constructed, the intensity of intent propagation between users is further quantified: the number of co-occurrence of primitives in the intent primitive streams of any two user nodes is counted, and the co-occurrence frequency is the ratio of the number of co-occurrence to the total number of primitives of the two users; a conditional probability model is used to calculate the conditional probability Q(u2|u1) of the occurrence of the intent primitive u2 of user U2 after the occurrence of the intent primitive u1 of user U1 as the primitive transfer probability from user U1 to user U2; the initial value of the edge weight is set as the weighted sum of the primitive co-occurrence frequency and the primitive transfer probability, and the weight coefficient is optimized by a grid search algorithm to determine the optimal combination; the edge weight updating mechanism is triggered every hour by the intent propagation graph evolution module: the co-occurrence frequency and the transfer probability are recalculated, and the edge weight is adjusted by using an exponential smoothing method, thereby constructing a complete heterogeneous graph; The constructed heterogeneous graph is input into a time series graph convolution network, and is divided into multiple time series subgraphs according to a 1-hour time window, and an adjacency matrix and a node feature matrix are constructed; wherein each time series subgraph corresponds to a 1-hour time window, and contains user nodes, edges and edge weights within the time window; the adjacency matrix and the node feature matrix of the input subgraph are input into the graph convolution layer, which aggregates the feature vectors of adjacent nodes for each node, and combines the edge weights to generate spatial features through weighted summation; the time convolution layer introduces a gating cycle unit to fuse the node features of adjacent time windows to generate embedding vectors containing space-time information; the embedding vectors are reduced in dimension through a fully connected layer, and the intention propagation features representing the intention propagation patterns of adjacent nodes are extracted; According to the intention propagation features output by the time series graph convolution network; adjust the weight of the edge according to the latest intention propagation feature; at the same time, based on the intention propagation feature, the user nodes are clustered and analyzed, and the user groups with similar intention propagation patterns are identified, the users with similar intention propagation patterns are classified into a class, and different user groups are formed; based on the adjusted edge weight, important connections are screened, and edges exceeding a preset filtering threshold are retained, and then an intention propagation graph is generated, wherein nodes represent users or user groups, edges represent intention propagation paths, and the thickness of the edge represents the intention propagation strength. 5.The deep learning based user intent prediction and accurate advertisement launching system according to claim 1, characterized in that: The intention confidence scene adaptation module obtains the intention propagation graph, and the process of injecting the user's current scene fingerprint feature vector into the graph node in real time includes: Build a real-time interaction interface with the intention propagation graph storage system; after obtaining the intention propagation graph, collect the user's current scene fingerprint feature vector: through the SDK of the mobile device end or the Web end, real-time collection of geographic location information; use high-precision positioning technology combined with map data to convert the geographic location into a semantic scene label; use the encoding rule to convert it into a semantic scene label feature vector; at the same time, collect device sensor data; convert it into a motion state feature vector through the encoding rule; concatenate the semantic scene label feature vector and the motion state feature vector to form a complete scene fingerprint feature vector; package all collected scene fingerprint feature vectors in a unified format, and inject them into the corresponding nodes of the intention propagation graph through a real-time data stream pipeline, and add scene dimension information to each node. 6.The deep learning based user intent prediction and accurate advertisement launching system according to claim 1, characterized in that: The intention confidence scene adaptation module calculates and analyzes the semantic coupling coefficient through the attention gate mechanism, and processes the intention propagation path, and the process of outputting the scene-calibrated high-confidence intention vector includes: The relationship between the scene fingerprint and the intent propagation path is further analyzed by building an attention gate architecture, wherein the architecture takes the feature vectors of the scene fingerprint and the intent propagation path as input data streams; the node feature matrix and edge weight information of the target intent propagation path are obtained from the intent propagation graph; the feature vectors of all nodes on the path are mean-pooled or weighted aggregated to obtain a node aggregated feature vector; the edge weight feature vector is generated by combining the edge weight information of the path; the node aggregated feature vector and the edge weight feature vector are spliced to form an intent propagation path feature vector; the scene fingerprint feature vector is obtained, and the scene fingerprint feature vector and the intent propagation path feature vector are input into the attention gate architecture; in the attention gate architecture, the semantic coupling coefficients of the scene fingerprint and each intent propagation path feature vector are calculated through an attention mechanism; at the same time, a dynamic activation threshold is set; for the paths with a semantic coupling coefficient higher than the dynamic activation threshold, the confidence is weighted and diffused according to the numerical size of the semantic coupling coefficient; for the low-coupling paths, noise interference is suppressed by reducing the weight, and finally a high-confidence intent vector calibrated by the scene is output. 7.The deep learning based user intent prediction and accurate advertisement launching system according to claim 1, characterized in that: The advertisement gene recombination generation module identifies the demand features in the high-confidence intent vector, extracts matching material segments from the advertisement gene library; the process of generating personalized advertisement prototypes by recombining gene segments through a generative adversarial constraint mechanism includes: The high-confidence intent vector is probabilistically decomposed using a variational autoencoder, and the implicit demand features are extracted by sampling in the latent space; N core dimensions strongly associated with user behavior are selected from the extracted implicit demand feature dimensions using mutual information entropy calculation; the core dimension set is obtained; based on the core dimension set, all core dimensions are sorted in descending order according to the numerical value of the mutual information entropy, and the first N core dimensions are extracted; A dynamic weight is assigned to each core dimension; the core dimensions are associated with industry terms using a knowledge graph to form a demand label system; based on the formed demand label system, matching material segments, i.e., gene segments, are extracted from the advertisement gene library: copywriting genes, visual genes, and interactive genes; A deep hashing algorithm is used to convert the demand labels into 128-bit binary codes; The feature code vector of the gene segment is defined; the binary code of the demand label is matched with the feature code of the gene segment in Hamming distance, and the material with a distance less than a preset Hamming distance threshold is preferentially returned; the matching result is secondarily screened, and the candidate gene segment is output; Based on the candidate gene segments of each dimension screened out, the generator uses hybrid attention to cross-modal encode the copywriting, visual, and interactive gene segments; in the encoding process, self-attention is used to capture the internal association of gene segments of the same type; cross-attention is used for semantic alignment of gene segments of different types; multi-head attention is used to splice the features of each modality to generate recombined gene segments; The discriminator sets double constraints, including a content discriminator and a scene discriminator; the reorganized gene segment is taken as a generated sample of the discriminator; the content discriminator verifies the logical coherence of the generated sample and the demand label; the scene discriminator evaluates the adaptability of the generated sample to the current scene of the user; through continuous optimization of the generator and the discriminator, the personalized advertisement prototype is finally output. 8.The deep learning based user intent prediction and accurate advertisement launching system according to claim 1, characterized in that: The privacy-preserving cross-domain collaborative delivery module injects a differential privacy noise layer into the personalized advertisement prototype to generate a noisy advertisement feature package; Through the federal learning gateway, the feature package is broadcast to the cooperative media platforms, and each platform decodes the feature package based on the local user group and performs the delivery process, which includes: Based on the privacy protection requirements of the advertisement delivery scene, the data sensitivity in different scenes is quantitatively evaluated, and the privacy budget parameter of differential privacy is set; the high-confidence intention vector is normalized and mapped to the numerical range [a, b]; According to the Laplace mechanism of differential privacy, a noise vector with the same dimension as the normalized high-confidence intention vector is generated; the Laplace mechanism samples noise values from the Laplace distribution and adds them to each dimension of the normalized high-confidence intention vector; The layered encoding strategy is used to fuse the high-confidence intention vector with added noise and the multi-modal advertisement features of the personalized advertisement prototype to generate a noisy advertisement feature package: the bottom layer uses an autoencoder to compress the multi-modal features into low-dimensional embeddings; the middle layer dynamically adjusts the feature weights of the modal level through an attention mechanism; the top layer applies differential privacy protection to the compressed features; After the generation of the noisy advertisement feature package is completed, a federal learning gateway is built to realize cross-platform collaborative delivery; the noisy advertisement feature package is encrypted and broadcast to each cooperative media platform through the federal learning gateway; after receiving the encrypted noisy advertisement feature package, each media platform uses the pre-set decryption key to decrypt it; according to the behavior feature data of the local user group, the noisy advertisement feature package is decoded; each media platform performs the advertisement delivery task in the local user group according to the decoded advertisement feature package.

Images