Marketing advertisement intelligent putting method and system based on AI

Abstract

The invention discloses an AI-based intelligent marketing advertisement putting method and system, and the method comprises the steps: collecting and fusing heterogeneous user behavior data streams from a plurality of independent platforms in real time, and generating a user behavior feature map with a unified space-time mark; on the basis of the user behavior characteristic spectrum, a deep auto-encoder and a space-time diagram convolutional network are used for joint modeling, and fine-grained scene-decoupled user interest preference vectors are generated; inputting the user interest preference vector into a pre-trained generative adversarial network, and dynamically generating a personalized advertisement material highly matched with the current user interest preference and scene; and taking the personalized advertisement material as a candidate arm, and dynamically generating an advertisement putting strategy including an advertisement display form, a display opportunity and a display channel in combination with a current user state and historical putting feedback data. By utilizing the embodiment of the invention, the accuracy and the real-time performance of advertisement putting can be improved, and the advertisement conversion rate and the user experience are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of AI, in particular to an AI-based marketing advertisement intelligent delivery method and system. BACKGROUND

[0002] With the rapid development of digital marketing, advertisement delivery faces challenges such as user behavior data fragmentation, high scene dynamics, and complex personalized needs. Existing technologies usually rely on single platform data or static user portraits, making it difficult to real-time fuse cross-platform heterogeneous behavior data (such as clicks, browsing, cross-device scenarios, etc.), resulting in coarse-grained user interest modeling. In addition, traditional advertisement generation methods are mostly based on fixed templates or artificial rules, lacking the ability to dynamically adapt to real-time user interests and scenes, and the optimization of delivery strategies is often limited to historical data statistics, making it impossible to achieve real-time decision-making with context awareness. SUMMARY

[0003] The purpose of the present application is to provide an AI-based marketing advertisement intelligent delivery method and system to solve the problems in the prior art, improve the accuracy and real-time performance of advertisement delivery, and improve advertisement conversion rate and user experience.

[0004] One embodiment of the present application provides an AI-based marketing advertisement intelligent delivery method, which comprises: real-time collection and fusion of heterogeneous user behavior data streams from multiple independent platforms, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / dwell time patterns, and cross-device context scene information, generating a unified spatiotemporal labeled user behavior feature map; based on the user behavior feature map, using a deep autoencoder and a spatiotemporal graph convolution network to jointly model, decoupling user long-term and short-term interest preferences and their coupling relationship with specific spatiotemporal scenes, generating fine-grained, scene-decoupled user interest preference vectors; input the user interest preference vector into a pre-trained generative adversarial network, combine real-time context scene constraints in the generator, through a discriminator-driven adversarial optimization process, dynamically generate personalized advertisement materials highly matched with current user interest preferences and scenes; use the personalized advertisement materials as candidate arms, combine current user state and historical delivery feedback data, use a context-aware multi-armed bandit model for real-time online learning and decision-making, dynamically generate advertisement delivery strategies including advertisement display form, display timing, and display channel.

[0005] Optionally, the real-time acquisition and fusion of heterogeneous user behavior data streams from multiple independent platforms, the data stream at least includes explicit click / conversion behavior sequence, implicit browsing / stay time pattern and cross-device context scene information, generates a unified spatio-temporal labeled user behavior feature map, including: Through the distributed stream processing engine, real-time ingestion of multi-platform user behavior data stream, time stamp calibration and space coordinate mapping of explicit behavior sequence and implicit behavior pattern, generation of spatio-temporal aligned original behavior trajectory set; Based on the original behavior trajectory set, using cross-device identity resolution algorithm to associate the fragmented behavior of the same user, constructing user global behavior chain through device fingerprint and behavior pattern similarity double verification; The global behavior chain is input into the heterogeneous data fusion module, and the semantic conflict resolution mechanism driven by the knowledge graph is used to align the behavior semantics of different platforms, and the disambiguated behavior atomic operation sequence is generated; The behavior atomic operation sequence is spatio-temporal graph convolutional coded, each operation is mapped to a graph node with spatio-temporal label, the behavior transition probability is represented by edge weight, and the unified spatio-temporal labeled user behavior feature map is output.

[0006] Optionally, based on the user behavior feature map, a deep autoencoder and a spatio-temporal graph convolutional network are jointly modeled to decouple the user's long-term and short-term interest preferences and their coupling relationship with specific spatio-temporal scenes, and a fine-grained, scene-decoupled user interest preference vector is generated, including: Extract the spatio-temporal subgraph from the user behavior feature map and input it into the double-channel deep autoencoder, wherein the time sequence channel uses LSTM to encode the behavior sequence pattern, and the space channel uses GAT to encode the behavior association topology, and outputs the long-term interest hidden vector and the short-term interest hidden vector respectively; The long-term interest hidden vector and the short-term interest hidden vector are input into the scene decoupling module, and the scene-related features and the scene-independent features are separated through adversarial domain adaptation, and the decoupled interest component group is generated; The interest component group is allocated a dynamic weight, and the activation weight of each component in the current spatio-temporal scene is calculated using the attention mechanism to generate a scene-adaptive interest fusion vector; The interaction between the interest fusion vector and the scene node is constructed through the spatio-temporal graph convolutional network, and the neighbor node influence weight is updated using the graph diffusion algorithm to output the fine-grained scene-enhanced interest vector; The scene-enhanced interest vector is subjected to orthogonalization constraint training to ensure that different interest dimensions are linearly independent, and finally a scene-decoupled user interest preference vector is generated.

[0007] Optionally, the user interest preference vector is input into the pre-trained generative adversarial network, the real-time context scene constraint is combined in the generator, and through a discriminator-driven adversarial optimization process, a personalized advertising material highly matched with the current user interest preference and scene is dynamically generated, including: The user interest preference vector is input into the generator of the pre-trained generative adversarial network, and a real-time context scene vector is injected, the interest and scene features are fused through a conditional control gate to generate an initial advertising material latent representation; A multi-scale evaluation module is constructed in the discriminator, wherein the structural scale verifies the visual logical consistency of the advertising material, the semantic scale matches the interest preference vector, and the scene scale checks the context adaptation degree, and outputs a three-dimensional evaluation score; An adversarial loss function is constructed according to the three-dimensional evaluation score, the generator is optimized through a gradient penalty mechanism, and the material latent representation is iteratively adjusted until it passes the multi-scale verification threshold; The optimized material latent representation is input into a diversity enhancement decoder, and a personalized multi-modal candidate advertising material is generated by using style transfer and content disentangling technology.

[0008] Optionally, the personalized advertising material is taken as a candidate arm, combined with the current user state and historical delivery feedback data, and a context-aware multi-armed bandit model is used for real-time online learning and decision-making to dynamically generate an advertising delivery strategy including advertising display form, display timing and display channel, including: The personalized advertising material is mapped into a candidate arm of a multi-armed bandit, and the feature vector of each arm includes material type, interest matching degree and scene adaptation degree; Context features are constructed based on the current state of the user, and the expected return value of each candidate arm is calculated through Thompson sampling; A hierarchical reward feedback mechanism is designed: short-term rewards record click-through rates, and long-term rewards track conversion funnel depth, and the arm value model is dynamically updated through time difference learning; The uncertainty entropy value of the candidate arm is calculated according to the expected return value, and when the entropy value exceeds the scene mutation threshold, an exploration strategy is activated, and the exposure weight of the candidate arm is redistributed according to the entropy value proportion; The updated arm value model and exploration weight are fused to generate a three-dimensional advertising delivery strategy including display form, display timing and display channel.

[0009] Another embodiment of the application provides an AI-based marketing advertising intelligent delivery system, which comprises: A collection module is configured to collect and fuse heterogeneous user behavior data streams from multiple independent platforms in real time, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / stay duration patterns and cross-device context scene information, and generate a unified spatiotemporal labeled user behavior feature map. a decoupling module configured to, based on the user behavior feature graph, utilize a deep auto-encoder and a spatio-temporal graph convolution network to jointly model, decouple long-term and short-term interest preferences of a user and coupling relations between the user and specific spatio-temporal scenes, and generate a fine-grained, scene-decoupled user interest preference vector; a generation module configured to input the user interest preference vector into a pre-trained generative adversarial network, combine real-time context scene constraints in a generator, and dynamically generate personalized advertising materials highly matched with current user interest preferences and scenes through a discriminator-driven adversarial optimization process; a decision module configured to use the personalized advertising materials as candidate arms, combine current user states and historical delivery feedback data, utilize a context-aware multi-armed bandit model to perform real-time online learning and decision-making, and dynamically generate an advertising delivery strategy including advertising display forms, display opportunities, and display channels.

[0010] Still another embodiment of the present application provides a storage medium having a computer program stored therein, wherein the computer program is configured to execute the method described in any of the above embodiments when running.

[0011] Still another embodiment of the present application provides an electronic device comprising a memory having a computer program stored therein and a processor configured to execute the computer program to perform the method described in any of the above embodiments.

[0012] Compared with the prior art, the AI-based marketing advertising intelligent delivery method provided by the present application can collect and fuse heterogeneous user behavior data streams from multiple independent platforms in real time, generate a unified spatio-temporal labeled user behavior feature graph, based on the user behavior feature graph, utilize a deep auto-encoder and a spatio-temporal graph convolution network to jointly model, generate a fine-grained, scene-decoupled user interest preference vector, input the user interest preference vector into a pre-trained generative adversarial network, dynamically generate personalized advertising materials highly matched with current user interest preferences and scenes, use the personalized advertising materials as candidate arms, combine current user states and historical delivery feedback data, and dynamically generate an advertising delivery strategy including advertising display forms, display opportunities, and display channels, thereby improving the accuracy and real-time performance of advertising delivery and improving advertising conversion rates and user experience. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 FIG. 1 is a hardware structure block diagram of a computer terminal for an AI-based marketing advertising intelligent delivery method according to an embodiment of the present application; Figure 2 FIG. 2 is a flowchart of an AI-based marketing advertising intelligent delivery method according to an embodiment of the present application; Figure 3Another flowchart of an AI-based marketing advertisement intelligent putting method provided by an embodiment of the present application is shown in FIG. 6. Figure 4 Another flowchart of an AI-based marketing advertisement intelligent putting method provided by an embodiment of the present application is shown in FIG. 6. Figure 5 A structural diagram of an AI-based marketing advertisement intelligent putting system provided by an embodiment of the present application is shown in FIG. 7. DETAILED DESCRIPTION

[0014] The embodiments described below with reference to the drawings are exemplary and are only used to explain the present application, and cannot be explained as a limitation of the present application.

[0015] An AI-based marketing advertisement intelligent putting method is provided by an embodiment of the present application, which can be applied to an electronic device, such as a computer terminal, specifically, a general computer, etc.

[0016] The AI-based marketing advertisement intelligent putting method is described in detail below by taking a computer terminal as an example. Figure 1 A hardware structural diagram of a computer terminal of an AI-based marketing advertisement intelligent putting method provided by an embodiment of the present application is shown in FIG. 2. Figure 1 As shown in the figure, the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0017] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can make the processor execute any kind of AI-based marketing advertisement intelligent putting method.

[0018] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0019] The internal memory provides an environment for the running of the computer program in the non-volatile storage medium, which, when executed by the processor, can make the processor execute any kind of AI-based marketing advertisement intelligent putting method.

[0020] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 The structure shown in FIG. 2 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0021] It should be appreciated that the processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0022] Referring to Figures 2-4 Embodiments of the present application provide an AI-based marketing advertisement intelligent delivery method, which can include the following steps: S201, real-time collection and fusion of heterogeneous user behavior data streams from multiple independent platforms, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / stay duration patterns and cross-device context scene information, generation of a unified spatiotemporal labeled user behavior feature map; Specifically, the distributed stream processing engine can be used to real-time ingest multi-platform user behavior data streams, time stamp calibration and space coordinate mapping of explicit behavior sequences and implicit behavior patterns, generation of a set of spatiotemporally aligned original behavior trajectories; The system deploys a data ingestion layer based on a distributed stream processing engine (e.g., Apache Flink or Spark Streaming). This engine connects multiple independent platforms (e.g., e-commerce platforms, social media, offline POS systems) in real-time through pre-configured API interfaces (e.g., RESTful API, Kafka Connector). The data streams from each platform are transmitted in a specific format (e.g., JSON, Protobuf) and contain explicit behavior sequences (e.g., clicking the "Buy Now" button, completing payment) and implicit behavior patterns (e.g., page dwell time of 120 seconds, video watch progress of 70%). Due to potential millisecond-level discrepancies in platform clocks (e.g., platform A timestamp delayed by 300 milliseconds), the engine calls a Network Time Protocol (NTP) service to calibrate the timestamps of all behavior data to the microsecond level, ensuring uniformity of the global time reference (e.g., unified to UTC+0 time zone). Meanwhile, for behaviors involving geographic locations (e.g., GPS coordinates "39.9042°N, 116.4074°E" obtained by an APP), the engine converts data from different positioning sources (GPS / IP positioning) into a unified spatial grid encoding (e.g., Geohash encoding "wx4g0" representing a certain business district in Beijing) through spatial coordinate mapping algorithms (e.g., WGS84 coordinate system to Mercator projection). The calibrated behavior data is reorganized into a "user ID-time stamp-spatial grid" triplet, forming an initial Raw Behavior Trajectory Set, with each trajectory representing a user's behavior segment in continuous space-time.

[0023] To handle high-concurrency data streams, the distributed engine adopts a sharded parallel processing architecture. For example, 128 Parallel Processing Units (PPUs) are set up, each responsible for data from a specific user shard (e.g., shards based on user ID hash values). Within each PPU, there is a behavior classifier module: explicit behaviors (e.g., "add to cart," "share link") are labeled as High-Intent Signals (HIS), and implicit behaviors (e.g., "page scrolling speed 0.5 seconds / screen," "mouse hover area heat value 85") are labeled as Latent-Interest Signals (LIS). The classified behaviors are input into a space-time alignment processor, which includes two core sub-modules: Time aligner: based on a sliding time window (e.g., 5-minute window), aggregate behaviors, and reorder the out-of-order events within the window according to the calibrated timestamps (e.g., adjust "browse first, click later" to the correct timing).

[0024] Spatial Mapper: associate ambiguous locations (e.g. IP-located "Chaoyang District, Beijing") to precise spatial grids (e.g. Geohash "wx4g09" represents Sanlitun) through Geo-Fence Matching Algorithm, while supplementing context labels for location-less behaviors (e.g. "Wi-Fi name Starbucks_798" maps to "Coffee Shop Context").

[0025] Final output is a Spatiotemporal-Aligned RawTrajectory Set (SARTS) with data structure: {User ID, Timestamp Sequence [T1, T2…Tn], Spatial Grid Sequence [G1, G2…Gn], Behavior Type Sequence [B1, B2…Bn]}.

[0026] Spatiotemporal alignment process needs to address the issue of uneven behavior density. For example, a user triggers 10 browsing events per second on a platform (high-frequency implicit behavior), while only 1 payment per day on another platform (low-frequency explicit behavior). The system introduces an Adaptive Sampling Mechanism (ASM): for high-frequency behaviors (such as dwell time), importance-based downsampling (retain samples with standard deviation greater than 15% threshold), for low-frequency behaviors (such as payment), full retention. At the same time, through the Behavior Weight Allocator (BWA), different behaviors are given confidence weights (such as GPS location behavior space weight = 1.0, IP location weight = 0.7). All processing is completed in the In-Memory Computing Layer of the stream processing engine, with a delay control within 100 milliseconds, ensuring the real-time nature of the trajectory set.

[0027] Based on the original behavior trajectory set, the Cross-Device Identity Resolution Algorithm is used to associate fragmented behaviors of the same user, and the user's global behavior chain is constructed through double verification of device fingerprint and behavior pattern similarity; The core of the Cross-Device Identity Resolution Algorithm (CDIR) is Device Fingerprint Clustering (DFC). The system extracts Device-Level Features (DLF) from the original behavior trajectory, including: Hardware Fingerprint (HF): e.g. device model "iPhone12", screen resolution "1170x2532", operating system version "iOS15.4".

[0028] Software Fingerprint (SF): e.g. browser User Agent "Mozilla / 5.0 (Macintosh; Intel Mac OS X 10_15_7)", APP SDK version "3.2.1".

[0029] Network Fingerprint (NF): e.g. IP segment "192.168.1.0 / 24", MAC address "a1:b2:c3:d4:e5:f6".

[0030] The algorithm adopts Hierarchical Clustering Model (HCM) to calculate the feature overlap between devices using Jaccard Similarity Coefficient (JSC). For example, if the JSC value of two devices = 0.85 (threshold > 0.8), they are classified into the same cluster (Cluster), and are preliminarily determined as the same user.

[0031] The device clustering result needs to be verified by Behavioral Pattern Similarity Verification (BPSV) to exclude false positives. The system extracts user behavior features (Behavioral Features, BF): Temporal Pattern (TP): e.g. daily active period distribution (80% active probability from 9am to 6pm).

[0032] Operational Habit (OH): e.g. click position preference (70% click rate on the right side of the screen), average swipe speed (200 pixels / second).

[0033] Content Preference (CP): e.g. browsing "digital product" category accounts for 45%, searching keyword "4K display" frequency 3 times / week.

[0034] The Dynamic Time Warping (DTW) algorithm is used to calculate the similarity of behavior sequences across different devices (a DTW distance < 0.2 is considered a match). For example, if a user searches for "coffee machine" on their phone at 9:00 AM and browses a coffee machine product page on their PC at 9:05 AM, the DTW distance is 0.15 (the time difference is within the tolerance window), and this is considered the same user behavior chain.

[0035] After passing the double verification, the system builds the User Omnichannel BehaviorChain (UBC). The data structure of each behavior chain is as follows: { MasterUserID: "U-7X9K3", / / Global user ID Devices: [ {DeviceID: "D-001", Type: "Mobile", Fingerprint: {HF: "...", SF: "...", NF: "..."}}, {DeviceID: "D-002", Type: "PC", Fingerprint: {...}} ], BehaviorSegments: [ {Device: "D-001", Time: "2023-08-20 09:00:00", Grid: "wx4g0", Action: "Search: coffee machine"}, {Device: "D-002", Time: "2023-08-20 09:05:00", Grid: "wx4g1", Action:"View:ProductID=789"} ] }.

[0036] To handle conflicts (such as multiple users sharing a device), the system sets up a confidence arbiter (CA): when the weighted value of the device fingerprint similarity (JSC) and the behavior pattern similarity (DTW) (weight ratio 6:4) is greater than 0.75, the association is accepted; otherwise, the manual review process is triggered.

[0037] The global behavior chain is input into the heterogeneous data fusion module, and the semantic conflict resolution mechanism driven by the knowledge graph is used to align the behavior semantics of different platforms to generate a disambiguated behavior atomic operation sequence. The core of the Heterogeneous Data Fusion Module (HDFM) is the pre-built Advertising Behavior Knowledge Graph (ABKG). This graph contains three types of entities: Action Entity: such as "Click", "Purchase", "Scroll".

[0038] Object Entity: such as "Product_Page", "Video_AD_15s".

[0039] Scenario Entity: such as "Home_Page", "Checkout_Process".

[0040] The graph defines semantic relationships through Property Edges, for example: "Click -(has_intent)->Purchase_Intent"; "Scroll -(indicates_interest)->Content_Engagement".

[0041] The Semantic Conflict Resolution Mechanism (SCRM) workflow is as follows: Behavior atomization decomposition: The original action in the global behavior chain (such as "double-clicking a product image") is decomposed into Atomic Operations (AO). For example, it is decomposed into: {Operation Type: "Click", Target Object: "Product_Image", Interaction Method: "Double"}.

[0042] Graph semantic mapping: The Graph Neural Network Matcher (GNNM) is used to query the knowledge graph. For example, "Add to Wishlist" on an e-commerce platform is mapped to the atomic operation "Save_Item", and "Bookmark Post" on a social platform is mapped to the same operation, eliminating platform terminology differences.

[0043] Conflict resolution rules: Intent conflict: e.g. a behavior is labeled as "high-value conversion" on platform A and "ordinary browsing" on platform B. The system calculates the Intent Confidence Score (ICS). If ICS > 0.8, the high-value label is adopted.

[0044] Scenario conflict: e.g. "clicking on banner ads" has a weight of 1.0 in the home page scenario and a weight of 0.3 in the payment page scenario (user may have been mistaken).

[0045] Output the Disambiguated Atomic Operation Sequence (DAOS). Each atomic operation contains standardized fields, such as: { OperationID: "AO-8832", Type: "CLICK", / / Standardized operation type (uppercase English) Object: "VIDEO_AD", / / Standardized object Scene: "SOCIAL_FEED", / / Standardized scenario Parameters: { Duration: 2.3, / / Additional parameters (such as duration 2.3 seconds) Position: {X: 0.65, Y: 0.72} / / Normalized coordinates }, Platform: "WeChat", / / Source platform Timestamp: "2023-08-20T09:00:00Z" / / ISO8601 time format }。

[0046] The sequence is strictly sorted by timestamp, forming a unified user behavior timeline across platforms for downstream feature graph construction.

[0047] Spatiotemporal graph convolutional encoding is performed on the atomic operation sequence, mapping each operation to a graph node with spatiotemporal labels, and representing behavior transition probabilities through edge weights, outputting a unified spatiotemporal labeled user behavior feature graph.

[0048] Construction process of Spatiotemporal Graph Convolutional Encoder (ST-GCE): Graph Node Generation: Convert each atomic operation in DAOS into a graph node. The Node Attribute Vector (NAV) contains: Behavior Type Encoding (e.g. CLICK=001, VIEW=002); Temporal-Spatial Marking (Timestamp converted into seconds of the day "32400" represents 9am, Spatial Grid Geohash converted into integer encoding "892736"); Scene Encoding (e.g. Home Page=1001, Product Detail Page=1002); Object Attributes (e.g. Ad Type=Video, Product Category=Electronics).

[0049] Example Node: Node_ID: N-557, NAV: [001, 32400, 892736, 1001, 301,205].

[0050] Edge Weight Calculation (EWC) defines the relationship between nodes: Temporal Edge: Connects consecutive operation nodes of the same user. Weight calculation formula: W_temporal = 1 / (time difference (seconds) + 1).

[0051] For example, operation A (T1=9:00:00) and operation B (T1=9:00:05) have a time difference of 5 seconds, weight = 1 / (5+1)=0.1667.

[0052] Spatial Edge: Connects operation nodes within the same spatial grid (across users). Weight = Spatial grid user density normalized value (e.g. weight = 0.1 if there are 10 people in the grid).

[0053] Semantic Edge: Based on knowledge graph to calculate behavior semantic relevance. For example, "clicking on an ad" and "adding to cart" have a path length of 2 in the knowledge graph, weight = 1 / 2=0.5.

[0054] Feature aggregation through multi-layer ST-GCN: First layer convolution: Aggregate single-user temporal neighbors (e.g. node N-557 aggregates the features of its predecessor node N-556 and successor node N-558); Second layer convolution: Aggregate spatial grid neighbors (e.g. aggregate the latest behavior of all users under the same Geohash encoding "wx4g0"); The third convolution layer aggregates semantically similar neighbors (such as aggregating all “ad click” behavior node features).

[0055] The final output is the User Behavior Feature Graph (UBFG). The data structure example is as follows: { UserID: "U-7X9K3", Nodes: [ {NodeID: "N-557", Attributes: [0.21, 1.8, -0.4, ...]}, / / 64-dimensional feature vector {NodeID: "N-558", Attributes: [...]} ], Edges: [ {Source: "N-557", Target: "N-558", Weight: 0.1667, Type: "Temporal"}, {Source: "N-557", Target: "N-901", Weight: 0.32, Type: "Semantic"} ] }.

[0056] The graph also carries spatiotemporal tags (such as the time code 32400 and spatial code 892736 of node N-557), becoming the input basis for downstream interest modeling.

[0057] First, a distributed data collection system is used to capture user behavior data across multiple platforms in real time. This data includes not only explicit behaviors like clicks and purchases, but also implicit behavioral characteristics like page dwell time and scroll depth, combined with contextual information such as device type and location. A spatiotemporal alignment algorithm is used to eliminate temporal and spatial deviations between different data sources, constructing a feature map encompassing the complete user trajectory. This addresses the fragmentation of cross-platform user behavior data. Unified spatiotemporal modeling restores the user's true behavioral path, providing a high-quality data foundation for subsequent interest mining. Furthermore, the inclusion of implicit behaviors compensates for the information missing from traditional click data, providing a more comprehensive and holistic user portrait.

[0058] S202, based on the user behavior feature map, using a deep autoencoder and a spatiotemporal graph convolutional network to jointly model the user's long-term and short-term interest preferences and their coupling relationship with specific spatiotemporal scenarios, and generating a fine-grained, scenario-decoupled user interest preference vector. Specifically, this may include: S2021, extract the space-time subgraph from the user behavior feature map, input the dual-channel deep autoencoder, wherein the time sequence channel uses LSTM to encode the behavior sequence mode, and the space channel uses GAT to encode the behavior correlation topology, and respectively outputs the long-term interest hidden vector and the short-term interest hidden vector; The system first extracts the space-time subgraph related to the target scene according to the current analyzed user behavior feature map (containing all behavior nodes and transition relations of the user with time and space labels). For example, if the user's interest in the evening of weekdays needs to be analyzed, the behavior nodes (such as "clicking on electronic product advertisements", "staying on food page for 15 seconds", "searching for restaurants near subway stations") and their connection edges with time stamps from 18:00 to 22:00 and space positions at home or commuting paths are extracted. The subgraph is input into a dual-channel deep autoencoder (Dual-Channel Deep Autoencoder). The architecture consists of two parallel branches: a temporal channel (Temporal Channel) and a spatial channel (Spatial Channel). The core of the temporal channel is a long short-term memory network (Long Short-Term Memory, LSTM), which is designed to capture the time dependence in the behavior sequence. LSTM processes user behavior nodes in time steps, for example, converting the user's "browse mobile phone -> compare prices -> collect goods" behavior for three consecutive days into a sequence input. The forget gate (Forget Gate) inside LSTM determines how much historical information to retain (such as the influence of yesterday's price comparison behavior on today), the input gate (Input Gate) determines the amount of new information of the current behavior (such as collecting goods), and the output gate (Output Gate) generates the current hidden state. By stacking two layers of LSTM units (128 neurons per layer), the system finally outputs a 128-dimensional long-term interest latent vector (Long-term Interest Latent Vector) from the last time step, which encodes the user's stable preference pattern across time periods (such as continuous interest in electronic products).

[0059] The core of the spatial channel is the Graph Attention Network (GAT) for modeling the non-temporal correlation topology between behavior nodes. GAT regards the spatio-temporal subgraph as a graph structure, and each behavior node (such as "clicking on advertisement A" and "staying on page B") is initialized as a 256-dimensional feature vector (containing attributes such as behavior type, duration, and location). GAT calculates the correlation strength between nodes through a multi-head attention mechanism: first, an attention coefficient is calculated for each node pair (such as node A and node B), which is generated by a small neural network (containing 64 hidden units) and input as a concatenated vector of the features of the two nodes. For example, the user "searching for restaurants" and "viewing discount coupons" may have a high attention coefficient of 0.85 due to spatial proximity (within 1 km of the same mall). After parallel calculation of 8 independent attention heads (each head outputs a 64-dimensional feature), the results are concatenated and input into a linear layer (containing 128 neurons) to finally generate a spatial feature representation for each node. Through iterative aggregation of two layers of GAT (the first layer aggregates direct neighbors and the second layer aggregates second-order neighbors), the system outputs a 128-dimensional short-term interest latent vector that focuses on the spatial correlation between the user's recent behaviors (such as the sudden desire for food during commuting).

[0060] The collaboration of the two channels is achieved through joint training: the decoder of the autoencoder attempts to reconstruct the original spatio-temporal subgraph from the long-term and short-term latent vectors. The reconstruction loss is calculated using the Mean Squared Error (MSE), for example, requiring the decoder to restore the exact time interval and location offset of the user "18:03 clicks on the advertisement -> 18:05 stays on the detail page". During training, the Adam optimizer (learning rate 0.001) is used to minimize the reconstruction loss to force the latent vectors to retain key information. The long-term interest latent vector usually converges to low-frequency features (such as the user's brand loyalty to electronic products), while the short-term interest latent vector captures high-frequency fluctuations (such as the impulsive consumption tendency triggered by temporary promotions). Both vectors are L2 normalized (with a length of 1) to avoid scale bias in subsequent calculations.

[0061] S2022, input the long-term interest latent vector and the short-term interest latent vector into the scene decoupling module, separate the scene-related features and scene-independent features through adversarial domain adaptation, and generate a group of decoupled interest components; The core goal of the Scene Disentanglement Module is to separate the scene dependency in the interest expressions. The module receives the long-term and short-term interest latent vectors (each of 128 dimensions), which are concatenated into a 256-dimensional Joint Interest Vector. This vector is input into a Shared Feature Extractor, which consists of three fully connected layers (each with 256 neurons, using ReLU activation functions), outputting a 256-dimensional Shared Features. The Shared Feature Extractor serves both adversarial tasks: the Domain Classification Task and the Interest Reconstruction Task. The goal of the Domain Classification Task is to distinguish the scene category (e.g., "home", "office", "mall") where the features come from, while the Interest Reconstruction Task requires preserving the user's core interests (e.g., electronics preference) regardless of the scene changes.

[0062] Adversarial Domain Adaptation is achieved through a Gradient Reversal Layer (GRL). The GRL is located between the Shared Feature Extractor and the Domain Classifier. The Domain Classifier is a three-layer neural network (128→64→number of scene categories), e.g., when the system defines 5 scenes (home, office, commute, mall, other), the output layer has 5 neurons. During forward propagation, the GRL directly passes the features; during backward propagation, the GRL backpropagates the gradient of the domain classification loss multiplied by a negative coefficient (e.g., -0.5). This means that the Shared Feature Extractor is trained to "fool" the Domain Classifier - i.e., to make the feature distributions of different scenes as similar as possible. For example, a user's interest in coffee will be encoded as the same feature pattern in the office (high frequency of purchases) and mall (occasional consumption) scenes, thus eliminating scene noise (e.g., additional purchase frequency in the office scene due to work stress).

[0063] Meanwhile, the shared features are input into an interest decoder, which consists of two independent branches: a scene-dependent branch and a scene-agnostic branch. The scene-dependent branch contains two fully connected layers (256→128→64) and outputs a 64-dimensional scene-dependent component, which is supervised by a domain classification loss (to be recognized by the classifier). The scene-agnostic branch has the same structure and outputs a 64-dimensional scene-agnostic component, which is trained to restore the user's core interest (e.g., the user always prefers high-end brands regardless of the scene) through a reconstruction loss. Finally, the two components are concatenated with the shared features to form a disentangled interest component group containing 256-dimensional shared features, 64-dimensional scene-dependent features, and 64-dimensional scene-agnostic features, totaling 384 dimensions. In training, a weight balancing strategy is adopted, with the domain classification loss weight set to 0.3 and the reconstruction loss weight set to 0.7, ensuring that interest fidelity is prioritized over scene separation.

[0064] S2023, dynamically allocate weights to the interest component group, calculate the activation weights of each component in the current spatiotemporal scene using an attention mechanism, and generate a scene-adaptive interest fusion vector; The core of the dynamic weight allocation module is a scene-driven attention mechanism. The module receives the disentangled interest component group (384 dimensions) and the current spatiotemporal scene vector (Current Spatiotemporal Scene Vector, 128 dimensions). The scene vector consists of three parts: 1) temporal features (e.g., hour encoding as a 24-dimensional sine / cosine vector); 2) spatial features (e.g., GPS coordinates converted to a 50-dimensional region embedding through geocoding); 3) device features (e.g., mobile phone / computer converted to a 54-dimensional vector through an embedding layer). After concatenation, the three parts are compressed to a 128-dimensional scene vector through a fully connected layer (128 neurons). This vector serves as the query vector of the attention mechanism.

[0065] The attention calculation consists of three steps: Key-value mapping: the 384-dimensional vector of the interest component group is generated through three independent fully connected layers: Key vector (Key Vector): 384→128 dimensions, representing the matchability of the component; Value Vector: 384→128 dimensions, carrying actual interest information; Scene Gate Vector: 384→128 dimensions, controlling scene sensitivity.

[0066] Correlation Scoring: Calculate the dot product of query vector (scene vector) and key vector, generate original attention score. For example, the user is currently in the mall (scene vector), the "promotion sensitive" component of the key vector may get a high score of 0.9.

[0067] Gate Adjustment: Multiply the original score with the scene gate vector element by element, suppress the components that conflict with the current scene. For example, in the office scene, the "leisure and entertainment" component of the scene gate output is close to 0, which will reduce its attention score to below 0.1.

[0068] Final score is normalized by Softmax to get the activation weight of each component (Activation Weight), and the total weight is 1.

[0069] Weighted fusion process: The value vector of each component is multiplied by the corresponding activation weight and summed to generate a 128-dimensional scene-adaptive interest fusion vector. For example, in the commuting scene (query vector), the weight of short-term interest component (such as "immediate discount") may reach 0.6, the weight of long-term interest component (such as "brand loyalty") is 0.3, and the weight of scene-related component (such as "subway station surrounding services") is 0.1. The vector is added to the original joint interest vector through residual connection (Residual Connection) to ensure that the basic interest is not lost. Finally, Layer Normalization (layer normalization) is used to stabilize the numerical distribution, and the output vector dimension remains 128.

[0070] S2024, build the interaction relationship between the interest fusion vector and the scene node through the spatio-temporal graph convolution network, update the neighbor node influence weight using the graph diffusion algorithm, and output the fine-grained scene-enhanced interest vector; Spatio-Temporal Graph Convolutional Network (ST-GCN) is used to model the interaction between user interest and external scene elements. The network constructs a dynamic graph structure: nodes include: 1) User Interest Node, which carries the scene adaptive interest fusion vector (128 dimensions) generated in the last step; 2) Scene Element Nodes, such as nearby "shopping mall promotion activities", "weather status", "social trends", etc., each node is represented by a pre-trained 64-dimensional embedding. The edges are divided into two categories: 1) influence edges from user nodes to scene nodes (representing the sensitivity of interest to the scene); 2) association edges between scene nodes (such as "rainy day" and "take-out service" strong association).

[0071] The graph convolution operation updates the weights using the Graph Diffusion Algorithm: Initialize influence weights: the edge weights from user nodes to each scene node are calculated by dot product. For example, the similarity between the user interest vector and the "discount activity" scene vector is 0.75, and the initial weight is set to 0.75.

[0072] Multi-order diffusion propagation: perform three-order diffusion: First-order diffusion: scene nodes receive directly connected user interests (weight 0.75) and update their own state (such as "discount activity" node value increase); Second-order diffusion: updated scene nodes affect their neighbors (such as "discount activity" enhances the "same brand" node weight 0.6); Third-order diffusion: indirectly connected nodes (such as "payment method") obtain weak influence (weight 0.2) through intermediate nodes.

[0073] Weight redistribution: after diffusion, recalculate the edge weights from user nodes to scene nodes. New weight = initial weight x (1 + average influence of neighbor nodes). For example, the scene node with initial weight 0.75, because its neighbor "payment method" is activated, the new weight is raised to 0.82.

[0074] User interest node update: aggregate features from all connected scenario nodes (weighted sum), e.g. user interest vector incorporates information from "discount activity (weight 0.82)" "weather (weight 0.3)" "social trend (weight 0.4)". The aggregated result is concatenated with the original interest vector (128+64=192 dimensions), and a gating mechanism (containing sigmoid function and fully connected layer) is used to filter noise, finally output a 128-dimensional scene-enhanced interest vector. For example, the original user interest "electronic products" is significantly enhanced in the "price sensitivity" dimension after incorporating the "time-limited discount" scenario.

[0075] S2025, orthogonalization constraint training is performed on the scene-enhanced interest vector to ensure that different interest dimensions are linearly independent, and finally a scene-decoupled user interest preference vector is generated.

[0076] The core goal of orthogonalization constraint training is to eliminate the dimension redundancy of the interest vector. The system constructs an orthogonal loss function that acts on the scene-enhanced interest vector (128 dimensions). The loss function calculates the correlation matrix between the dimensions of the vector: the batch training data (e.g. 1024 user interest vectors) are stacked column by column into a matrix X (size 1024x128), and the covariance matrix C = XᵀX (size 128x128) is calculated. In the ideal state, C should be close to a diagonal matrix (non-diagonal elements are 0).

[0077] The orthogonalization loss is defined as the Frobenius norm of the non-diagonal elements of the covariance matrix C. Specifically, the diagonal elements of C are set to zero to obtain matrix C0.

[0078] The square sum of all elements of C0 is calculated, and the square root is taken to obtain the loss value: orthogonal loss = √(∑ᵢ∑ⱼ(c0ᵢⱼ)²).

[0079] For example, if the correlation coefficient of two dimensions (e.g. "brand loyalty" and "price sensitivity") is 0.8, then c0ᵢⱼ=0.8, and its square 0.64 will be included in the loss. The training goal is to minimize the loss (ideal value is 0).

[0080] The training process jointly optimizes the orthogonal loss (weight 0.2) and the main task loss (such as the cross-entropy loss of ad click rate prediction, weight 0.8). Projected Gradient Descent is used: after each gradient update, the interest vector is subjected to Singular Value Decomposition (SVD) and projected into the orthogonal basis space. The final output of the scene-disentangled user interest preference vector satisfies: The correlation coefficient of any two interest dimensions (such as dimension 3 "food preference" and dimension 7 "sports interest") is less than 0.1; The variance of each dimension is close to 1 (after standardization); The 128-dimensional vector is used as input for downstream ad generation, ensuring that different interests (such as shopping and entertainment) are orthogonal in the vector space, avoiding model confusion of users' multi-faceted interest features.

[0081] A dual-channel neural network is used to process the temporal pattern and spatial correlation of user behavior. The long-term interest channel captures stable preference features, and the short-term interest channel identifies immediate behavioral motivations. Through adversarial training, scene-related and scene-independent features are separated, avoiding excessive influence of specific scenes on interest modeling. This breaks the limitation of traditional interest models that simply stack long-term and short-term features, achieving truly scene-adaptive interest expression. The decoupled interest vector can reflect both the user's essential preferences and dynamically adapt to different scene requirements.

[0082] S203, input the user interest preference vector into the pre-trained generative adversarial network. In the generator, combine real-time context scene constraints, and through the discriminator-driven adversarial optimization process, dynamically generate personalized ad materials that match the current user interest preference and scene; specifically, it can include: S2031, input the user interest preference vector into the generator of the pre-trained generative adversarial network, and simultaneously inject a real-time context scene vector. Through a conditional control gate, integrate interest and scene features to generate an initial ad material latent representation; The system receives a user interest preference vector (UIPV) from an upstream module. This is a high-dimensional numerical vector (typically 128 or 256 dimensions) generated by a deep learning model, quantifying the user's fine-grained interest preferences. For example, a value of 0.87 in the 10th dimension of the vector indicates a high interest in "outdoor sports equipment," while a value of 0.45 in the 35th dimension indicates a moderate interest in "luxury handbags." At the same time, the system collects a real-time context scene vector (RCSV) in real-time. This vector is dynamically generated by sensors and a context analysis engine, containing multi-dimensional information about the current environment: device type (1 for mobile, 0 for PC), geographic location (GPS coordinates encoded as a latitude-longitude grid value), network status (0 for 4G, 1 for 5G, 2 for WiFi), time period (1 for weekday daytime, 2 for nighttime, 3 for weekend), ambient noise level (normalized to 0-1), etc. These two key inputs are synchronized and fed into the generator (Gen) module of a pre-trained generative adversarial network (GAN). The generator has a conditional control gate (CCG) designed inside. The core function of the CCG is to dynamically fuse the UIPV and RCSV. Its workflow is as follows: First, the CCG calculates the feature affinity matrix (FAM) of the two vectors, evaluating the relevance of interest dimensions and scene dimensions through dot product operations (for example, the affinity score between "outdoor interest" and "current GPS located in a park" is as high as 0.92); then, based on the FAM, a dynamic fusion weight mask (DFWM) is generated, which determines the strength of the influence of each interest feature on the scene (for example, high-affinity features are boosted to 1.2 times their original weight, while low-affinity features are reduced to 0.8 times their original weight); finally, the CCG performs a weighted fusion operation and outputs a condition-enhanced interest vector (CEIV), which retains the core interest features and deeply embeds the scene adaptation signals.

[0083] The core structure of the generator is a Deep Convolutional Neural Network (DCNN) or a Transformer Decoder (TD). Taking the DCNN architecture as an example, it contains multiple Upsampling Convolution Blocks (UCBs). The Conditional Enhancement Interest Vector (CEIV) first enters a Fully Connected Embedding Layer (FCEL) and is converted into an Initial Latent Code (ILC) with low dimension and high density. This latent code serves as the seed input for the generation process. Subsequently, the ILC passes through multiple UCBs in sequence. Each UCB contains: a Transposed Convolution Layer (TCL) to gradually expand the feature map size (e.g., from 4x4 pixels to 64x64 pixels); a Batch Normalization Layer (BNL) to stabilize the training process; a Conditional Instance Normalization Layer (CINL) to specifically modulate the style parameters (such as hue, texture) of the feature map, with the modulation coefficients directly derived from the CEIV; and an activation function layer (such as ReLU or LeakyReLU). During the process of gradually expanding the feature map, the CEIV is continuously injected into each UCB through a Conditional Projection Connection (CPC), ensuring that the scene constraints are involved in the entire feature generation process. Finally, the last UCB outputs a rough Ad Material Latent Representation (AMLR). At this point, the AMLR is a three-dimensional tensor (e.g., 128x128x3, corresponding to width, height, and RGB channels), whose numerical distribution has not been optimized. It can be regarded as a "primary sketch" of the ad material, containing basic composition and content outline, but with blurred details and possibly uncoordinated logic.

[0084] To improve the quality of the initial latent representation, the generator usually connects a residual refinement module (RRM) at the end. The RRM is stacked by multiple residual dense blocks (RDBs). Each RDB aggregates multi-level features through dense skip connection (DSC) and calculates the importance weight of each feature channel (e.g., red channel weight 0.9, blue channel weight 0.7) using channel attention mechanism (CAM) to preferentially retain key color information. After multiple rounds of feature extraction and fusion by the RRM, the AMLR gradually repairs structural defects (such as twisted object boundaries), enhances detail clarity (such as text sharpening), and strengthens consistency with the CEIV (such as ensuring that the recommended product category matches the interest label). The output end compresses the channel number to the target format (such as 3-channel RGB for static images and time-series frame sequences for video ads) through a 1x1 convolution layer. At this point, the generator completes the forward propagation and outputs the initial ad material latent representation (IAMLR). This representation has not been verified by the discriminator and needs to go through an adversarial optimization loop for iterative improvement.

[0085] S2032, a multi-scale evaluation module is constructed in the discriminator, wherein the structural scale verifies the visual logical consistency of the ad material, the semantic scale matches the interest preference vector, the scene scale checks the context adaptation degree, and a three-dimensional evaluation score is output. The initial ad material latent representation IAMLR is input into the discriminator (Discriminator, Dis). The discriminator is not simply a judgment of "true or false", but a multi-scale evaluation module (Multi-scale Evaluation Module, MEM) is constructed to make fine scoring from three orthogonal dimensions: 1. Structural Scale (SS): Focuses on the underlying visual plausibility and logical consistency of the material. This module first extracts multi-level visual feature maps through a pre-trained convolutional feature extractor (such as the shallow layers of VGG19 or ResNet50). Then, the Structural Consistency Analysis Subnet (SCAS) calculates the Self-Similarity Matrix (SSM) of the feature maps to detect whether the local pattern repetition is abnormal (such as the presence of mechanically copied textures); the Spatial Relation Reasoning Module (SRRM) uses a Graph Neural Network (GNN) to model the topological relationships between objects (such as "title text" should be located above "product picture"), and deducts points for violating preset spatial rules (such as overlapping text and pictures); the Physical Plausibility Validator (PPV) outputs the number of violations based on the physical law prior in the generated adversarial network training data (such as light and shadow direction consistency, perspective proportion). Finally, the SS score (range 0 to 1) integrates the above indicators, and the higher the score, the more natural and reasonable the visual material is.

[0086] 2. Semantic Scale (SemS): The core module evaluates the semantic alignment between the ad content and the user's interest preference. This module adopts a Multimodal Semantic Alignment Network (MSAN). First, a Visual Semantic Encoder (VSE) encodes the IAMLR into a deep feature, outputting a Visual Semantic Vector (VSV), such as a 1024-dimensional vector generated by global average pooling. At the same time, an Interest Preference Parser (IPP) decodes and semantically maps the input original user interest preference vector UIPV (e.g., mapped to a Parsed Interest Vector PIV of the same dimension as VSV through a fully connected layer). The key step is Cross-modal Similarity Calculation (CSC): Cosine Similarity (CS) or Bilinear Attention (BA) is used to measure the matching degree of VSV and PIV. For example, if UIPV shows that the user's interest value for "electric cars" is 0.9, while VSV detects that the material mainly shows "gasoline cars", the similarity will be significantly reduced. In addition, a Conceptual Conflict Detector (CCD) scans the repulsive relationship between the material elements (obtained through target detection) and the interest labels (such as the user explicitly rejecting "carbonated drinks", but the material appears a coke icon), and a significant score is deducted if a conflict is detected. The SemS score (0 to 1) reflects the interest matching accuracy.

[0087] 3. Context Scale (CS): Strictly check the adaptability of the advertisement material to the real-time scene. This module receives the real-time context scene vector RCSV as the conditional input. The context rule engine (CRE) loads the pre-defined business rule library (such as "subway commuting scene disables audio advertisements", "night mode needs to enable dark theme"), checks the compliance of IAMLR piece by piece, and triggers the flag if it violates the rules. The context feature adaptation subnet (CFAS) calculates the matching degree of the material visual attributes (such as the dominant color luminance value, the dynamic element complexity) and the scene parameters: for example, through the luminance analyzer (LA) to calculate the average pixel luminance (APL) of the material, if APL is 0.8 (high brightness) and the current ambient light sensor reports the brightness value as 0.1 (dark environment), the adaptation degree is low; the bandwidth adaptivity evaluator (BAE) predicts the material loading time according to the network state value (in RCSV) (such as tolerating high-definition video in 5G environment, and prioritizing static pictures in 3G environment), and deducts the score if it exceeds the time limit. Finally, the CS score (0 to 1) integrates the rule compliance rate and the feature adaptation degree.

[0088] The discriminator summarizes the scores of the three scales (SS_Score, SemS_Score, CS_Score) to form a three-dimensional evaluation score vector (3D Evaluation Score Vector, 3DESV), which is directly used as the basis for subsequent adversarial optimization.

[0089] S2033, according to the three-dimensional evaluation score, build an adversarial loss function, optimize the generator through gradient penalty mechanism, and iteratively adjust the material latent representation until it passes the multi-scale verification threshold; The core of adversarial optimization is to build a multi-objective adversarial loss function (MOALF). This function is composed of three parts weighted together: Structural adversarial loss (SAL): based on SS score calculation. The discriminator's structural score for real high-quality advertisement materials is expected to be high (target value close to 1), and the goal of the generated material is to maximize the SS score. SAL adopts the form of least squares loss (LSL): SAL = (SS_Score - 1)^2, that is, the generator needs to push the SS score close to 1.

[0090] Semantic Adversarial Loss (SemAL): based on SemS score. Its calculation introduces Interest Matching Distance (IMD): IMD = 1 - SemS_Score. SemAL is defined as SemAL = max(0, IMD - Margin), where Margin is a pre-set threshold (e.g. 0.2), only when the matching is insufficient beyond the threshold, the penalty is applied, avoiding over-optimization.

[0091] Context Adversarial Loss (CAL): based on CS score. Designed as Hinge Loss (HL): CAL = max(0, Threshold_CS - CS_Score), where Threshold_CS is the context adaptation threshold (e.g. 0.7). When the CS score is below this line, a loss is generated.

[0092] Finally, MOALF = a * SAL + β * SemAL + γ * CAL, the weight coefficients a, β, γ (initial values such as 0.4, 0.3, 0.3) are dynamically adjusted according to business priorities.

[0093] To prevent the optimization process from oscillating or collapsing, the loss function integrates a Gradient Penalty Mechanism (GPM). The core of GPM is to add a regularization term in the loss function: Relative Gradient Penalty (RGP). Its calculation process is: randomly sample two adjacent points θ1 and θ2 in the generator parameter space, calculate their corresponding IAMLR1 and IAMLR2, and the score difference of the discriminator ΔS = |Dis(IAMLR1) - Dis(IAMLR2)|. At the same time, calculate the Euclidean distance of the parameter space Δθ = ||θ1 - θ2||. RGP is defined as RGP = λ * max(0, (ΔS / Δθ) - 1)^2, where λ is the penalty intensity coefficient (e.g. 10). This penalty term forces the small change of generator parameters to cause the score change not to exceed the linear growth, effectively smoothing the optimization path, avoiding gradient explosion or mode collapse.

[0094] The optimization process adopts an alternating training strategy (ATS): fixing the discriminator parameters, calculating the gradient of the MOALF on the generator parameters through the backpropagation algorithm (BP), and updating the generator weights using an adaptive optimizer (such as Adam or RMSProp) (learning rate, for example, 0.0001). After updating, the generator generates new IAMLR, and the discriminator scores again. This cycle continues, and the three-dimensional score is recorded for each iteration. The system is preset with a multi-scale validation threshold (MVT), for example, requiring SS_Score>0.85, SemS_Score>0.80, and CS_Score>0.75. When the scores of consecutive N rounds (such as 3 rounds) of iterations all exceed the MVT, or the maximum number of iterations (such as 100 times) is reached, the optimization is terminated. The final output of the optimized ad material latent representation (OAMLR) meets the acceptable standards in structure, semantics, and scene.

[0095] S2034, input the optimized ad material latent representation into the diversity-enhanced decoder, and generate personalized multi-modal candidate advertising materials by using style transfer and content disentanglement technology.

[0096] The optimized ad material latent representation OAMLR enters the diversity-enhanced decoder (DED). The core task of DED is to introduce controllable style changes while ensuring that the core content (such as recommended products and key copy) strictly matches user interests, and generate multiple groups of candidate materials for subsequent delivery strategy selection. To this end, DED adopts a content-style disentanglement architecture (CSDA). The architecture includes two parallel branches: The content preservation pathway (CPP) is composed of multiple residual convolution layers (RCL) and directly processes OAMLR. Its design feature is a low-pass filtering characteristic (LFC), which preserves the low-frequency structural information in OAMLR that is strongly related to interest semantics (such as product outlines and core slogans) by limiting the convolution kernel receptive field and reducing nonlinear operations.

[0097] The Style Modulation Pathway (SMP) receives a Style Noise Vector (SNV). The SNV is a low-dimensional vector (e.g., 32-dimensional) randomly sampled from a Gaussian distribution, representing the source of style variation. The SMP uses a multilayer perceptron (MLP) to map the SNV into a Style Parameter Set (SPS), which consists of a set of affine transformation parameters (scaling factor γ and offset β).

[0098] The two signals are combined in the Disentangled Fusion Layer (DFL). DFL employs Adaptive Instance Normalization (AdaIN) technology: First, the feature maps output by the CPP undergo per-channel instance normalization (mean subtraction and standard deviation division) to remove style information. Next, the normalized features are affine transformed (channel-by-channel scaling and offsetting) using the gamma and beta parameters provided by the SMP to infuse the new style. This process ensures that the content channel maintains the core ad information (e.g., product model, promotional terms), while the style channel independently controls the visual presentation (e.g., changing the color scheme from cool to warm, the font from a formal serif to a lively sans serif, and the background music from piano to electronic music).

[0099] To generate Multimodal Candidate Ad Materials (MCAM), DED supports multiple output heads: Static image decoding head: The fused feature map is decoded into an RGB image (such as a 1200x628 pixel banner image) through a transposed convolution chain.

[0100] Video sequence generation head: Integrates a 3D convolution layer (3DCL) and a recurrent generation unit (RGU) to output 15-second short videos (720p resolution, 25 frames per second) with dynamic transitions and audio-visual synchronization.

[0101] Interactive Rich Media Header: Connects to the UI Element Generator (UIEG) to output layout code (HTML5+CSS3) for interactive components such as clickable buttons, carousels, and forms.

[0102] By injecting different SNVs into the same OAMLR, DED quickly generates K groups (e.g., K = 5) of candidate materials with different styles but consistent content. For example, the same sports shoe advertisement can generate "extreme sports style" (high saturation, dynamic shots), "technology future" (cool color, geometric lines), "warm life" (warm color, family scene), and other versions. All candidate materials carry metadata (such as style code, interest matching degree, scene adaptation degree), which are input into the downstream multi-armed bandit model for optimal placement decision.

[0103] The generator receives the interest vector and real-time scene features, and dynamically adjusts the generation strategy through conditional control gates. The discriminator evaluates the material quality from three dimensions of visual logic, semantic matching, and scene adaptation, drives the generator to iteratively optimize, and finally outputs creative content that meets both user interest and current scene adaptation, realizing the leap of advertising creativity from static templates to dynamic generation, and ensuring that the generated content has both aesthetic quality and commercial value through adversarial optimization. The multi-dimensional evaluation mechanism avoids semantic deviation or scene mismatch in generated content.

[0104] S204, the personalized advertising material is used as a candidate arm, combined with the current user state and historical placement feedback data, and a multi-armed bandit model with context awareness is used for real-time online learning and decision-making to dynamically generate an advertising placement strategy including advertising display form, display timing, and display channel.

[0105] Specifically, the personalized advertising material can be mapped to a candidate arm of a multi-armed bandit, and the feature vector of each arm includes material type, interest matching degree, and scene adaptation degree. After the generation of personalized advertising materials, the system converts them into candidate arms (Candidate Arm) in the multi-armed bandit (Multi-Armed Bandit, MAB) model. Each candidate arm corresponds to an advertising material scheme, and its feature vector (Feature Vector) is composed of three core dimensions: Creative Type (CT): describes the physical form of the advertisement, such as Video Ad, Carousel Image, and Interactive Popup. The system maps the material format to a numerical value through a pre-defined classification encoder, such as Video Ad encoded as numerical value 1 and Carousel Image as numerical value 2.

[0106] Interest Match Score (IMS): quantifies the relevance of the material to the user's interest preference vector. It is calculated by the cosine similarity between the semantic embedding vector of the ad material (output by the generative adversarial network) and the user's interest preference vector, with a value range of 0 to 1 (0 represents no correlation, 1 represents complete match).

[0107] Context Adaptation Score (CAS): evaluates the compatibility of the material with the current spatio-temporal context. For example, a user receives an audio ad (context adaptation score 0.9) during the commute period (context label: mobile state) is better than a long text ad (adaptation score 0.3). The score is generated by comparing the environmental features output by the context decoupling module with the material scene constraints.

[0108] Example: a video ad for coffee lovers, whose feature vector can be represented as [CT=1, IMS=0.85, CAS=0.92]. The system builds a feature matrix for all candidate arms as the input basis of the multi-armed bandit machine.

[0109] Based on the user's current state, the context features are constructed, and the expected reward value of each candidate arm is calculated by Thompson sampling; User Current State (UCS) is a dynamic vector that contains real-time behavior features (such as the current browsing page category), device status (mobile / PC), and environmental parameters (city mall information based on GPS positioning). The system integrates the following data through the context feature builder: Real-time behavior snapshot: page dwell time, scrolling speed, etc. in the current session, captured in real time by the stream processing engine.

[0110] Historical preference decay factor: weights recent behavior based on a time decay function (such as exponential decay), avoiding outdated interests from interfering.

[0111] Context enhancement label: such as "weekday evening home scene" and "weekend outdoor scene", output by the spatio-temporal graph convolution network.

[0112] Thompson Sampling (TS) is used to calculate the expected reward value (ERV) of the candidate arm: Each candidate arm maintains a Beta Distribution parameter: the number of successes α (such as clicks) and the number of failures β.

[0113] For the current context feature, TS randomly samples a reward probability value from the beta distribution of each arm.

[0114] The arm with the highest sampled value is chosen as the current optimal candidate, but the sampling results of other arms are retained for exploration.

[0115] Example: The historical feedback data of a certain video ad is α = 120 clicks and β = 30 ignores, and its beta distribution shape indicates a high click probability. When the user is in the "commuting mobile scenario", TS may sample a reward value of 0.88; while the same ad in the "office scenario" may have a sampling value of 0.65.

[0116] Design a hierarchical reward feedback mechanism: short-term reward records click rate, long-term reward tracks conversion funnel depth, and dynamically update arm value model through temporal difference learning. The Hierarchical Reward Mechanism splits user behavior feedback into two layers: Short-Term Reward (STR): Records immediate behavior after ad exposure, such as Click-Through Rate (CTR). CTR is calculated by dividing the number of clicks by the number of exposures, and the value is normalized to the range of 0-1.

[0117] Long-Term Reward (LTR): Tracks the depth of the user's subsequent conversion path (i.e. Conversion Funnel Depth, CFD). For example: Level 1: Add to cart (reward value 0.3); Level 2: Payment page access (reward value 0.6); Level 3: Complete purchase (reward value 1.0).

[0118] Temporal Difference Learning (TD Learning) dynamically updates the arm value model: Define the Value Function Q(s,a) to represent the expected total reward of choosing arm a in state s.

[0119] After each user feedback (such as a click), calculate the TD error: the difference between the actual reward and the estimated reward.

[0120] Update the Q value by backpropagating the error through the TD(λ) algorithm (λ is the decay factor, default value 0.7).

[0121] Example: A certain advertisement exposure, user clicks (STR=0.8) and purchases within 24 hours (LTR=1.0), the system calculates the comprehensive reward R=0.5×STR+0.5×LTR=0.9. If the original Q value is 0.75, the TD error is 0.15, then update the Q value to 0.7515 with a learning rate of 0.01.

[0122] According to the expected return value, the uncertainty entropy value of the candidate arm is calculated, and when the entropy value exceeds the scene mutation threshold, the exploration strategy is activated, and the candidate arm exposure weight is redistributed according to the entropy value proportion; The uncertainty entropy value (Uncertainty Entropy Value, UEV) measures the confidence of the candidate arm return prediction: For each candidate arm beta distribution (α, β), calculate its information entropy:

[0123] Where k is the success / failure event, and p is the probability estimate. The entropy value ranges from 0 to 1 (0 indicates certainty, and 1 indicates high uncertainty).

[0124] The scene mutation threshold (Context Shift Threshold, CST) is a dynamic set alarm line (default value 0.8), when the environment feature change rate exceeds the preset (such as the user suddenly switches the device), CST automatically down to 0.6.

[0125] The activation logic of the exploration strategy (Exploration Strategy): If the UEV of a certain arm is greater than CST, the system starts the entropy weight allocator (Entropy-Based Weight Allocator): Calculate the total entropy (Total Entropy, TE) of all high-entropy arms; redistribute the exposure traffic according to the formula Weight_i = UEV_i / TE.

[0126] Low-entropy arms (UEV < CST) still allocate traffic according to the Thompson sampling result.

[0127] Example: A certain new material due to insufficient data (α=2, β=1) causes UEV=0.92 (exceeds CST=0.8). The system calculates its entropy weight as 0.4, the original optimal arm weight decreases from 0.9 to 0.5, and the remaining 0.1 is allocated to other arms. This balances exploitation (Exploitation) and exploration (Exploration).

[0128] Fusion of the updated arm value model and exploration weight, generate a three-dimensional advertisement placement strategy including display form, display timing, and display channel.

[0129] The generation of the three-dimensional advertising strategy (3D Advertising Strategy) is divided into three steps: Display Format (DF): Extract the content type (CT) from the candidate arm feature vector, which is directly mapped to the physical form (such as CT=1→video advertisement).

[0130] Display Timing (DT): According to the behavior prediction model (Behavior Prediction Model) in the user state: If the user's current stay duration is greater than the threshold (such as 8 seconds), it will be displayed immediately; if the user is in the page transition period, it will be delayed to the next page loading and displayed 3 seconds later.

[0131] Display Channel (DC): Selected by the scene adaptation degree (CAS) maximization principle: Mobile high-CAS materials are preferentially delivered to information streams (such as WeChat Moments); PC high-CAS materials are preferentially delivered to sidebar ad slots.

[0132] The execution process of the Strategy Fusion Engine: Input: Thompson-sampled optimal arm, candidate set after exploration weight adjustment, real-time user state; Decision: With a 90% probability, select the TS optimal arm, and with a 10% probability, randomly select other arms according to the exploration weight (exploration ratio can be adjusted); Output: Structured strategy instructions such as {DF: "video", DT: "page_load+3s", DC: "wechat_moments"}.

[0133] Example: The user is in the post-work shopping scenario (UCS label "evening_shopping"), the system selects a video advertisement (CT=1), and pushes it through the TikTok short video channel (DC strategy) 3 seconds after the user enters the product detail page (DT strategy).

[0134] Each advertising material is considered as an arm of a slot machine, and the exploration and utilization are balanced through the Thompson sampling algorithm. A hierarchical reward mechanism is designed, focusing on short-term click rates and immediate feedback, and tracking conversion rates and other deep indicators in the long term. The advertising strategy is dynamically adjusted according to the user state changes, realizing the self-evolution ability of the advertising strategy, optimizing the long-term marketing value while ensuring the immediate effect. Through the uncertainty perception mechanism, the user interest drift is captured in time, avoiding the efficiency decay caused by strategy solidification.

[0135] It can be seen that the real-time acquisition and fusion of heterogeneous user behavior data streams from multiple independent platforms generate a unified spatiotemporal labeled user behavior feature graph; based on the user behavior feature graph, a deep autoencoder and a spatiotemporal graph convolution network are jointly modeled to generate a fine-grained, scene-decoupled user interest preference vector; the user interest preference vector is input into a pre-trained generative adversarial network to dynamically generate personalized advertising materials highly matched with the current user interest preference and the scene; the personalized advertising materials are taken as candidate arms, and the current user state and historical delivery feedback data are combined to dynamically generate an advertising delivery strategy including advertising display forms, display timing and display channels, so that the precision and real-time performance of advertising delivery can be improved, and the advertising conversion rate and user experience can be improved.

[0136] Another embodiment of the application provides an AI-based marketing advertising intelligent delivery system, referring to Figure 5 , the system can include: The acquisition module 501 is configured to acquire and fuse real-time heterogeneous user behavior data streams from multiple independent platforms, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / stay duration patterns and cross-device context scene information, and generate a unified spatiotemporal labeled user behavior feature graph; The decoupling module 502 is configured to jointly model a deep autoencoder and a spatiotemporal graph convolution network based on the user behavior feature graph, decouple the long-term and short-term interest preferences of the user and the coupling relationship between the long-term and short-term interest preferences and specific spatiotemporal scenes, and generate a fine-grained, scene-decoupled user interest preference vector; The generation module 503 is configured to input the user interest preference vector into a pre-trained generative adversarial network, combine real-time context scene constraints in the generator, and dynamically generate personalized advertising materials highly matched with the current user interest preference and the scene through a discriminator-driven adversarial optimization process; The decision module 504 is configured to take the personalized advertising materials as candidate arms, combine the current user state and historical delivery feedback data, and use a context-aware multi-arm bandit model to perform real-time online learning and decision-making, and dynamically generate an advertising delivery strategy including advertising display forms, display timing and display channels.

[0137] The embodiment of the application further provides a storage medium, and the storage medium stores a computer program, wherein the computer program is set to execute the steps in any of the method embodiments.

[0138] Specifically, in the embodiment, the storage medium can be set to store a computer program for executing the following steps: S201, real-time collection and fusion of heterogeneous user behavior data streams from multiple independent platforms, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / stay duration patterns and cross-device context scene information, generation of a unified spatiotemporal labeled user behavior feature graph; S202, based on the user behavior feature graph, joint modeling of a deep autoencoder and a spatiotemporal graph convolution network, decoupling of long-term and short-term interest preferences of a user and coupling relationships between the user and specific spatiotemporal scenes, generation of a fine-grained, scene-decoupled user interest preference vector; S203, input of the user interest preference vector into a pre-trained generative adversarial network, combination of real-time context scene constraints in a generator, dynamic generation of personalized advertising materials highly matched with current user interest preferences and scenes through an adversarial optimization process driven by a discriminator; S204, combination of the personalized advertising materials as candidate arms, current user states and historical delivery feedback data, real-time online learning and decision making through a context-aware multi-armed bandit model, dynamic generation of an advertising delivery strategy including advertising display forms, display opportunities and display channels.

[0139] The embodiment of the application also provides an electronic device including a memory and a processor, the memory storing a computer program, and the processor being configured to run the computer program to execute the steps in any of the method embodiments.

[0140] Specifically, the electronic device can further include a transmission device connected to the processor and an input / output device connected to the processor.

[0141] Specifically, in the embodiment, the processor can be configured to execute the following steps through the computer program: S201, real-time collection and fusion of heterogeneous user behavior data streams from multiple independent platforms, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / stay duration patterns and cross-device context scene information, generation of a unified spatiotemporal labeled user behavior feature graph; S202, based on the user behavior feature graph, joint modeling of a deep autoencoder and a spatiotemporal graph convolution network, decoupling of long-term and short-term interest preferences of a user and coupling relationships between the user and specific spatiotemporal scenes, generation of a fine-grained, scene-decoupled user interest preference vector; S203, input of the user interest preference vector into a pre-trained generative adversarial network, combination of real-time context scene constraints in a generator, dynamic generation of personalized advertising materials highly matched with current user interest preferences and scenes through an adversarial optimization process driven by a discriminator; S204, the personalized advertisement material is taken as a candidate arm, and a context-aware multi-arm bandit model is used to perform real-time online learning and decision-making in combination with the current user state and historical delivery feedback data, so as to dynamically generate an advertisement delivery strategy including an advertisement display form, a display time and a display channel.

[0142] The above describes the structure, features and effects of the present application in detail according to the embodiments shown in the drawings. The above description is only the preferred embodiments of the present application, but the present application is not limited to the embodiments shown in the drawings. Any change or modification made according to the concept of the present application, or equivalent embodiments with equivalent changes, shall be within the protection scope of the present application.

Claims

1. An AI-based marketing advertisement intelligent delivery method, characterized in that, The method comprises: Real-time acquisition and fusion of heterogeneous user behavior data streams from multiple independent platforms, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / stay duration patterns and cross-device context scene information, generating a unified spatiotemporal labeled user behavior feature graph; Based on the user behavior feature graph, a deep autoencoder and a spatiotemporal graph convolution network are jointly modeled to decouple the long-term and short-term interest preferences of the user and the coupling relationship between the user and a specific spatiotemporal scene, and a fine-grained, scene-decoupled user interest preference vector is generated; The user interest preference vector is input into a pre-trained generative adversarial network, the real-time context scene constraint is combined in the generator, and through a discriminator-driven adversarial optimization process, a personalized advertising material highly matched with the current user interest preference and the scene is dynamically generated; The personalized advertising material is taken as a candidate arm, the current user state and historical delivery feedback data are combined, a context-aware multi-armed bandit model is used for real-time online learning and decision-making, and an advertising delivery strategy including advertising display form, display timing and display channel is dynamically generated.

2. The method of claim 1, wherein, The real-time acquisition and fusion of heterogeneous user behavior data streams from multiple independent platforms, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / stay duration patterns and cross-device context scene information, generating a unified spatiotemporal labeled user behavior feature graph, comprises: Through a distributed stream processing engine, multi-platform user behavior data streams are real-time ingested, the explicit behavior sequences and the implicit behavior patterns are timestamp calibrated and spatial coordinate mapped, and a spatiotemporally aligned original behavior trajectory set is generated; Based on the original behavior trajectory set, a cross-device identity resolution algorithm is used to associate the fragmented behaviors of the same user, a user global behavior chain is constructed through device fingerprint and behavior pattern similarity double verification; The global behavior chain is input into a heterogeneous data fusion module, a knowledge graph driven semantic conflict resolution mechanism is used to align the behavior semantics of different platforms, and a disambiguated behavior atomic operation sequence is generated; The behavior atomic operation sequence is spatiotemporal graph convolution coded, each operation is mapped to a graph node with spatiotemporal labels, the behavior transition probability is represented by edge weight, and a unified spatiotemporal labeled user behavior feature graph is output.

3. The method of claim 2, wherein, The based on the user behavior feature graph, the deep autoencoder and the spatiotemporal graph convolution network are jointly modeled to decouple the long-term and short-term interest preferences of the user and the coupling relationship between the user and a specific spatiotemporal scene, and a fine-grained, scene-decoupled user interest preference vector is generated, comprising: From the user behavior feature graph, a spatiotemporal subgraph is extracted and input into a double-channel deep autoencoder, wherein the time sequence channel uses LSTM to encode the behavior sequence pattern, and the space channel uses GAT to encode the behavior association topology, and long-term interest hidden vectors and short-term interest hidden vectors are respectively output; The long-term interest hidden vectors and the short-term interest hidden vectors are input into a scene decoupling module, scene-related features and scene-independent features are separated through adversarial domain adaptation, and a decoupled interest component group is generated; The interest component group is dynamically weighted and allocated, and an attention mechanism is used to calculate the activation weight of each component in the current spatiotemporal scene, and a scene-adaptive interest fusion vector is generated. An interaction relationship between an interest fusion vector and a scene node is constructed through a spatio-temporal graph convolution network, a neighbor node influence weight is updated using a graph diffusion algorithm, and a fine-grained scene-enhanced interest vector is output; The scene-enhanced interest vector is subjected to orthogonalization constraint training to ensure that different interest dimensions are linearly independent, and finally a scene-decoupled user interest preference vector is generated.

4. The method of claim 3, wherein, The user interest preference vector is input into a pre-trained generative adversarial network, real-time context scene constraints are combined in the generator, and through a discriminator-driven adversarial optimization process, a personalized advertising material that is highly matched with the current user interest preference and the scene is dynamically generated, including: The user interest preference vector is input into the generator of the pre-trained generative adversarial network, and at the same time, a real-time context scene vector is injected, interest and scene features are fused through a conditional control gate, and an initial advertising material latent representation is generated; A multi-scale evaluation module is constructed in the discriminator, wherein the structural scale verifies the visual logic consistency of the advertising material, the semantic scale matches the interest preference vector, and the scene scale checks the context adaptation degree, and a three-dimensional evaluation score is output; An adversarial loss function is constructed according to the three-dimensional evaluation score, the generator is optimized through a gradient penalty mechanism, and the material latent representation is iteratively adjusted until it passes the multi-scale verification threshold; The optimized material latent representation is input into a diversity-enhanced decoder, and a personalized multi-modal candidate advertising material is generated using style transfer and content disentangling techniques.

5. The method of claim 4, wherein, The personalized advertising material is taken as a candidate arm, combined with the current user state and historical delivery feedback data, and a context-aware multi-armed bandit model is used for real-time online learning and decision-making to dynamically generate an advertising delivery strategy including advertising display form, display timing and display channel, including: The personalized advertising material is mapped to a candidate arm of the multi-armed bandit, and the feature vector of each arm includes material type, interest matching degree and scene adaptation degree; Context features are constructed based on the current state of the user, and the expected return value of each candidate arm is calculated through Thompson sampling; A hierarchical reward feedback mechanism is designed: short-term rewards record click-through rates, and long-term rewards track conversion funnel depth, and the arm value model is dynamically updated through time difference learning; The uncertainty entropy value of the candidate arm is calculated according to the expected return value, and when the entropy value exceeds the scene mutation threshold, an exploration strategy is activated, and the exposure weight of the candidate arm is redistributed in proportion to the entropy value; The updated arm value model and exploration weight are fused to generate a three-dimensional advertising delivery strategy including display form, display timing and display channel.

6. An AI-based marketing advertisement intelligent delivery system, characterized in that, The system comprises: A collection module for collecting and fusing real-time heterogeneous user behavior data streams from multiple independent platforms, the data streams at least including explicit click / conversion behavior sequences, implicit browsing / stay duration patterns and cross-device context scene information, and generating a unified spatio-temporal labeled user behavior feature graph; A decoupling module for decoupling the user long-term and short-term interest preferences and their coupling relationship with a specific spatio-temporal scene based on the user behavior feature graph, using a deep autoencoder and a spatio-temporal graph convolution network joint modeling, and generating a fine-grained, scene-decoupled user interest preference vector; The generating module is configured to input the user interest preference vector into a pre-trained generative adversarial network, combine real-time context scene constraints in a generator, and dynamically generate personalized advertising materials highly matched with the current user interest preference and the scene through a discriminator-driven adversarial optimization process. The decision module is configured to use the personalized advertising materials as candidate arms, combine the current user state and historical delivery feedback data, and use a context-aware multi-arm bandit model to perform real-time online learning and decision-making, and dynamically generate an advertising delivery strategy including an advertising display form, a display time, and a display channel.

7. The system of claim 6, wherein, The collection module is specifically configured to: timestamp calibration and spatial coordinate mapping of explicit behavior sequences and implicit behavior patterns are performed on the multi-platform user behavior data stream through a distributed stream processing engine to generate a set of original behavior trajectories aligned in time and space; based on the set of original behavior trajectories, a cross-device identity resolution algorithm is used to associate fragmented behaviors of the same user, and a user global behavior chain is constructed through double verification of device fingerprints and behavior pattern similarity; the global behavior chain is input into a heterogeneous data fusion module, a knowledge graph-driven semantic conflict resolution mechanism is used to align behavior semantics of different platforms, and a disambiguated behavior atomic operation sequence is generated; the behavior atomic operation sequence is subjected to spatio-temporal graph convolution coding, each operation is mapped to a graph node with spatio-temporal labels, behavior transition probabilities are represented through edge weights, and a unified spatio-temporal labeled user behavior feature graph is output.

8. The system of claim 7, wherein, The decoupling module is specifically configured to: temporal and spatial subgraphs are extracted from the user behavior feature graph and input into a double-channel deep autoencoder, wherein the time sequence channel uses LSTM to encode the behavior sequence pattern, and the spatial channel uses GAT to encode the behavior association topology, and long-term interest hidden vectors and short-term interest hidden vectors are respectively output; the long-term interest hidden vectors and the short-term interest hidden vectors are input into a scene decoupling module, scene-related features and scene-independent features are separated through adversarial domain adaptation, and a group of decoupled interest components is generated; the group of interest components is subjected to dynamic weight allocation, and an attention mechanism is used to calculate the activation weights of each component in the current spatio-temporal scene to generate a scene-adaptive interest fusion vector; an interaction relationship between the interest fusion vector and the scene node is constructed through a spatio-temporal graph convolution network, neighbor node influence weights are updated using a graph diffusion algorithm, and a fine-grained scene-enhanced interest vector is output; the scene-enhanced interest vector is subjected to orthogonalization constraint training to ensure that different interest dimensions are linearly independent, and finally a scene-decoupled user interest preference vector is generated.

9. A storage medium, characterized by The storage medium has stored therein a computer program, which is configured to execute the method of any one of claims 1-5 when run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory has stored therein a computer program, and the processor is configured to execute the computer program to execute the method of any one of claims 1-5.

Images