Intelligent optimization system for advertisement exposure frequency based on deep learning

Abstract

The application belongs to the technical field of deep learning and artificial intelligence, and specifically relates to an intelligent optimization system for advertisement exposure frequency based on deep learning. The system normalizes and encodes user identification, advertisement object identification, creative feature, delivery scene feature, request time slice, exposure result and interaction feedback in the advertisement request log, constructs a time sequence heterogeneous graph containing user nodes, advertisement object nodes, creative nodes, scene nodes and time window nodes according to a rolling time window, generates an associated fatigue representation of a candidate advertisement object through a relationship perception graph neural network, generates an individual exposure memory representation of a user continuous exposure sequence through an exposure memory network, and then fuses the two types of representations through a frequency control gate generator to output a frequency control result of allowed exposure, reduced-weight exposure or suspended exposure. The system integrates cross-advertisement object association and user time sequence attenuation into the same discrimination process, thereby reducing implicit repeated exposure.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and artificial intelligence technology, specifically a deep learning-based intelligent optimization system for advertising exposure frequency. Background Technology

[0002] Existing ad exposure frequency control is mostly deployed in the ad request decision chain. Its conventional implementation uses user identifier, ad object identifier, ad campaign identifier, and time window as a composite key to accumulate exposure counts in a cache or log table. Upon receiving an ad request, the system reads the current user's exposure count within a daily, weekly, or custom rolling period and compares it to a preset frequency cap. If the cap is not reached, the candidate ad continues to participate in the ranking; if the cap is reached, the candidate ad is filtered or its ranking weight is reduced. Some systems also write interaction results such as clicks, conversions, blocking, and skips into user profiles or ad object profiles, and use historical exposure counts as input features for the ranking model. Count updates typically employ request-side writing, exposure callback confirmation, and window expiration clearing. Frequency thresholds are pre-configured according to ad campaigns, media resources, or user segments. For situations where the same creative is split into multiple ad objects, or similar creatives appear concurrently in different delivery scenarios, conventional frequency control still accumulates counts separately according to the ad object number. When ad object numbers change, creatives are reused, or themes are migrated, existing counts are difficult to carry over between objects. Such solutions are simple to implement and have a stable engineering process, but frequency control mainly relies on the cumulative count under independent advertising objects. The threshold is usually set by historical experience, the campaign cycle and the average response curve. It is difficult to uniformly represent repeated exposure of users under similar advertising creatives, similar campaign themes or the same campaign scenarios.

[0003] To address the coarse-grained nature of fixed-threshold frequency control, some advertising systems introduce deep learning models to predict click, visit, or conversion probabilities, concatenating user features, ad features, scenario features, historical exposure counts, and interaction feedback into a sample vector. During training, the model uses post-exposure interaction results as supervisory signals to learn response scores; in online inference, the system determines whether to display candidate ads based on the response scores and current frequency counts. Other solutions employ sequence neural networks to record recent user exposure trajectories, extracting short-term interest or fatigue changes in order of request time. The modeling process for these solutions typically consists of offline sample construction, deep learning training, online feature concatenation, real-time scoring, and rule truncation. The text, image tags, landing page themes, and delivery scenarios of the ad objects are often used as ordinary input features in the prediction. Because these features are usually compressed into the same vector, the connections, interaction edge types, and time window relationships between ad objects do not form a propagable data structure, and model parameter updates struggle to distinguish between fatigue propagation between similar ads and independent responses between dissimilar ads. Although the models can utilize a variety of features, most still model at the level of a single ad object or a single ad campaign. The semantic proximity relationships between ad creatives, the shared relationships between ad placement scenarios, and the fatigue migration relationships between similar user groups are not incorporated into a unified graph structure. The frequency states obtained by the deep learning models are still biased towards local historical records.

[0004] The main technical problem with existing technologies is that the frequency status of ad exposure lacks a unified data representation that simultaneously expresses cross-ad object association and user temporal decay, making it difficult for frequency control judgment of candidate ads to reflect implicit repeated exposure. The technical reason is that fixed counting methods only record the number of exposures between the same user and the same ad object, failing to pass on repeated touches from similar creatives, related themes, or the same scenarios to the current candidate ad. While single-sequence models can record the historical exposure order of users, they lack sufficient sequence evidence in new ad objects, low-exposure users, or low-sample creative scenarios, making it difficult to supplement fatigue states with similar objects and users. For continuous ad request flows, once the frequency status lacks cross-object propagation paths and time intervals, the contact traces of the same user on multiple similar ad objects are split into multiple isolated records. The model can only infer frequency risk from local samples of the current candidate object, and the fatigue recovery degree corresponding to different exposure intervals is also difficult to participate in frequency control judgment. This problem is mainly manifested in the incomplete input basis of the candidate ad frequency status, with missing graph associations and temporal decay simultaneously affecting the gating results. As a result, the system is prone to misjudging similar exposures scattered across different ad objects as independent reach in the ad request flow, and it is also prone to treating long-term scattered exposures and short-term intensive exposures similarly, resulting in unstable basis for generating the allowed exposure, downgraded exposure and suspended exposure intervals of candidate ads. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent optimization system for advertising exposure frequency based on deep learning, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based intelligent optimization system for ad exposure frequency includes a sample preprocessor, a temporal heterogeneous graph builder, a relation-aware graph neural network, an exposure memory network, and a frequency control gating generator. The sample preprocessor normalizes and encodes the user identifier, ad object identifier, creative features, delivery scenario features, request time slice, exposure results, and interaction feedback in the ad request log; The temporal heterogeneous graph builder generates a heterogeneous graph containing user nodes, advertising object nodes, creative nodes, scene nodes, and time window nodes according to a rolling time window. The relation-aware graph neural network outputs an association fatigue representation of candidate advertising objects based on different types of connection edges; The exposure memory network outputs an individual exposure memory representation based on the user's continuous exposure sequence; The frequency control gating generator integrates the associated fatigue characterization and the individual exposure memory characterization to generate the exposure frequency control result of the candidate advertising object.

[0007] Preferably, the temporal heterogeneous graph builder includes a node mapper and an edge relation generator; The node mapper maps users, advertising objects, advertising creatives, delivery scenarios, and time segments that appear within the same scrolling time window to graph nodes with type identifiers. The edge relationship generator establishes exposure edges, click edges, conversion edges, and negative feedback edges between user nodes and advertising object nodes, establishes content ownership edges between advertising object nodes and creative nodes, establishes scene association edges between advertising object nodes and scene nodes, and writes time segmentation tags and interaction result tags for each connection edge.

[0008] Preferably, the relation-aware graph neural network includes an edge type encoding layer, a relation-based message aggregation layer, and a cross-window state update layer; The edge type encoding layer generates edge type vectors for exposure edges, click edges, conversion edges, negative feedback edges, content ownership edges, and scene association edges, respectively. The sub-relationship message aggregation layer calculates the adjacent node messages according to the edge type vector; The cross-window state update layer performs gated updates on the node messages of the current scrolling time window and the node embeddings of adjacent historical scrolling time windows, generating an associated fatigue representation that includes the similarity influence relationship of advertising objects and the user fatigue migration relationship.

[0009] Preferably, the exposure memory network includes an exposure sequence organizer, a time interval encoder, and a temporal attention encoder; The exposure sequence organizer extracts the exposure records of candidate advertising objects and their similar advertising objects in multiple time segments according to the user identifier, and forms an exposure sequence sorted by the requested time. The time interval encoder maps the time interval between adjacent exposure records into a decay position vector; The temporal attention encoder jointly encodes the exposure sequence and the decay position vector, outputting an individual exposure memory representation that distinguishes between short-period dense exposures and long-period dispersed exposures.

[0010] Preferably, the temporal heterogeneous graph builder further includes a creative semantic consistency processor; The creative semantic consistency processor extracts the semantic embeddings of the advertising creative text, landing page summary, material tags, and the campaign theme. Based on the adjacency relationship between the semantic embeddings, it generates creative nearest neighbor edges between advertising object nodes and performs joint constraint encoding on the creative nearest neighbor edges and the content attribution edges. This allows advertising object nodes with the same type of creative to share the fatigue propagation path in the relation-aware graph neural network, while advertising object nodes with different themes retain independent sets of connection edges.

[0011] Preferably, the sub-relationship message aggregation layer includes a negative feedback suppression channel and a positive interaction retention channel; The negative feedback suppression channel performs joint encoding on the negative feedback edge and its connected adjacent exposure edge, and writes the joint encoding result into the fatigue component of the advertising object node; The positive interaction retention channel separates and aggregates the click edge and the conversion edge, and writes the aggregation result into the response component of the ad object node; The cross-window state update layer generates a relational frequency state of candidate advertising objects in the current request time slice based on the fatigue component and the response component.

[0012] Preferably, the temporal attention encoder includes an advertising object attention head and a time decay attention head; The advertising object attention head assigns attention weights to candidate advertising objects, creative neighboring advertising objects, and non-neighboring advertising objects respectively; The time decay attention head adjusts the participation weight of different historical exposure records in the user's exposure sequence according to the decay position vector; The temporal attention encoder gates and splices the outputs of the two attention heads to form an individual exposure memory representation corresponding to the candidate advertising object.

[0013] Preferably, the frequency control gating generator includes a characterization alignment layer, a fusion gating layer, and a frequency interval discrimination layer; The representation alignment layer projects the relational frequency state and the individual exposure memory representation to the same feature space; The fusion gating layer generates a graph representation participation coefficient and a sequence representation participation coefficient based on the historical sample density of the candidate advertising object, and performs weighted fusion on the projected representations. The frequency interval discrimination layer outputs the allowed exposure interval, the reduced-weight exposure interval, or the suspended exposure interval based on the fused frequency risk characterization.

[0014] Preferably, the fusion gating layer further includes a sparse state recognizer; The sparse state recognizer generates sparse state codes based on the number of historical exposures of the user, the number of historical exposures of the candidate ad object, the number of creative nearest neighbor edges, and the number of scene-related edges. The fusion gating layer increases the graph representation participation coefficient when sparse state encoding represents low-sample states and limits the variation range of the sequence representation participation coefficient. When sparse state encoding represents high-sample states, it increases the sequence representation participation coefficient based on the user's continuous exposure sequence.

[0015] Preferably, the frequency interval discrimination layer further includes a cross-object frequency control consistency checker; The cross-object frequency control consistency checker calls the creative nearest neighbor edge, the relational frequency state and the sparse state code to perform consistency check on the frequency interval of the same user on the candidate ad object and its creative nearest neighbor ad object; When the frequency range of a candidate ad object is inconsistent with the frequency range of a creative neighbor ad object, the cross-object frequency control consistency checker re-determines the frequency control gating value of the candidate ad object based on the neighbor edge weight and the graph representation participation coefficient.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By adopting a technical solution that integrates temporal heterogeneous graphs and exposure memory, the frequency of ad exposures is transformed from a single count record into a composite representation with correlations and temporal order. The sample preprocessor encodes users, ad objects, creatives, delivery scenarios, time slices, exposure results, and interaction feedback from the ad request log into computable features. The temporal heterogeneous graph builder organizes these features into user nodes, ad object nodes, creative nodes, scenario nodes, and time window nodes, preserving different interaction relationships through exposure edges, click edges, conversion edges, negative feedback edges, content attribution edges, and scenario association edges. The relation-aware graph neural network aggregates adjacency messages according to edge type, enabling exposure fatigue on similar ad objects, negative feedback trends in similar user groups, and repeated reach traces under scenario association to enter the correlation fatigue representation of candidate ad objects. The exposure memory network then generates individual exposure memory representations according to the user's continuous exposure sequence and adjacent exposure time intervals, distinguishing between short-cycle dense exposures and long-cycle scattered exposures at the encoding level. The frequency control gating generator aligns and fuses the graph-side correlation fatigue representations with the sequence-side individual exposure memory representations to generate allowed exposure intervals, reduced-weight exposure intervals, or paused exposure intervals. Cross-window state updates further embed nodes from adjacent historical windows into the current graph representation, preventing frequency states from being completely reset during window switching. Edge type encoding and relational message aggregation ensure that exposure, clicks, conversions, and negative feedback enter candidate ad objects along different connection paths, reducing representation bias caused by the mixing of heterogeneous interaction signals. The graph structure provides fatigue propagation paths across ad objects and user groups, while the sequence structure preserves the exposure order of individual users in consecutive requests. Both together correct the frequency risk representation of candidate ad objects. This processing chain ensures that the frequency judgment of candidate ads is constrained by both cross-object relationships and user temporal states, solving the problem of difficulty in frequency control identification when similar ads are scattered across different objects. It also transforms the generation basis of frequency intervals from isolated counting to a joint judgment of associated fatigue and individual memory.

[0017] 2. At the deep technical mechanism level, the Creative Semantic Consistency Processor generates creative nearest neighbor edges based on the semantic embedding of ad creative text, landing page summary, material tags, and campaign theme. These creative nearest neighbor edges are jointly constrained and encoded with content attribution edges, ensuring that similar creative ad objects share fatigue propagation paths, while ad objects with significantly different themes retain independent sets of connected edges, reducing fatigue transmission between unrelated ads from the source of the graph structure. The relational message aggregation layer sets up negative feedback suppression channels and positive interaction retention channels, writing negative feedback edges and adjacent exposure edges into the fatigue component, and click edges and conversion edges into the response component, preventing different interaction signals from overlapping in the same vector. The temporal attention encoder encodes candidate ad objects, creative nearest neighbor ad objects, non-nearest neighbor ad objects, and exposure records at different time intervals separately through ad object attention heads and time decay attention heads, ensuring that sequence memory retains object similarity and time decay differences. The fusion gating layer adjusts the participation ratio of graph representation and sequence representation based on sample density. Low-sample states rely more on graph relationship supplementation, while high-sample states rely more on continuous exposure sequence correction, reducing frequency control offset in new ad objects or low-exposure user scenarios. The sparse state recognizer writes the number of historical exposures of the user, the number of historical exposures of the candidate ad object, the number of creative nearest neighbor edges, and the number of scene-related edges into the sparse state code, giving the adjustment of the fusion ratio a clear data source. The frequency interval discrimination layer outputs different frequency control intervals on the frequency risk representation after fusion, ensuring that the results of upstream relationship modeling and sequence modeling fall into a unified decision space. The cross-object frequency control consistency checker redetermines the gate value based on the creative nearest neighbor edges, relational frequency states, and sparse state codes, ensuring that the frequency control intervals of the same user on the nearest neighbor ad objects remain traceably consistent, reducing the probability of conflicting frequency control results between similar ad objects, and allowing changes in the weight of creative nearest neighbor edges to be fed back into the gate value correction process of the candidate ad objects. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the overall operation of a deep learning-based intelligent optimization system for ad exposure frequency. Figure 2 Flowchart for the construction of temporally heterogeneous graphs and the processing of relation-aware graph neural networks; Figure 3 The flowchart for generating memory networks and frequency control gating is shown. Detailed Implementation

[0019] In one embodiment, a deep learning-based intelligent optimization system for ad exposure frequency includes a sample preprocessor, a temporal heterogeneous graph builder, a relationship-aware graph neural network, an exposure memory network, and a frequency control gating generator. The sample preprocessor normalizes and encodes user identifiers, ad object identifiers, creative features, delivery scenario features, request time slices, exposure results, and interaction feedback from ad request logs. The temporal heterogeneous graph builder generates a heterogeneous graph containing user nodes, ad object nodes, creative nodes, scenario nodes, and time window nodes according to a rolling time window. The relationship-aware graph neural network outputs association fatigue representations of candidate ad objects based on different types of connection edges. The exposure memory network outputs individual exposure memory representations based on continuous user exposure sequences. The frequency control gating generator fuses the association fatigue representations and the individual exposure memory representations to generate candidate ads. Regarding the exposure frequency control results of the target audience, this embodiment addresses the issue that the ad exposure frequency status cannot simultaneously express cross-ad object association and user temporal decay. It transforms the multi-source interaction records in the ad request stream into graph structure representation and sequence representation. The graph structure representation is used to carry the propagable fatigue relationship between different ad objects, creatives, scenarios, and users. The sequence representation is used to carry the time interval, object similarity, and interaction feedback changes of the same user during continuous exposure. The two are fused in the same feature space within the gating layer, so that the frequency control result of the candidate ad object is transformed from isolated counting judgment to interval discrimination result constrained by associated fatigue and individual memory. The advantage of this embodiment is that it can cover scenarios with incomplete frequency control status such as ad object splitting, creative reuse, dense requests, and sparse historical samples, and provides a unified data generation path for allowed exposure, deweighted exposure, and paused exposure.

[0020] In this embodiment, the sample preprocessor reads user identifier, ad object identifier, ad creative identifier, ad creative text, landing page summary, material tags, campaign theme, campaign scenario, request time, exposure receipt, click receipt, conversion receipt, block receipt, and skip receipt from the ad request log. The user identifier and ad object identifier are used to form a discrete index using hash mapping. The creative text, landing page summary, material tags, and campaign theme are semantically embedded through word segmentation, removal of invalid symbols, word vector mapping, and sentence vector convergence. The campaign scenario is embedded using media type, page position, content theme, and request time period. The request time is divided into time slices according to a scrolling time window. Exposure results and interaction feedback are categorized by whether they occurred. The occurrence time and feedback type generate state codes. Normalization coding uses discrete index coding, continuous numerical scaling, and category embedding to form a unified sample vector. Missing interactive feedback is not directly filled as a negative sample, but is recorded as a no-return feedback marker to avoid mistakenly writing feedback delay into the frequency state. The preprocessed samples are stored using a composite index based on user identifier, advertising object identifier, and time slice, providing the same reference basis for subsequent construction of time-series heterogeneous graphs and exposure sequence organization. The advantage of this embodiment is that it transforms the count data, semantic data, scenario data, and feedback data in the advertising request log into an input structure that can be processed by deep learning networks, reducing semantic loss caused by directly splicing different data forms.

[0021] In this embodiment, the temporal heterogeneous graph builder uses a scrolling time window as the boundary for generating graph instances. Each graph instance includes user nodes, ad object nodes, creative nodes, scene nodes, and time window nodes within the current window. User nodes store recent exposure statistics, interaction feedback statistics, and historical active time slice codes for users. Ad object nodes store the creative semantic embedding, scene distribution embedding, and interaction state within the window for ad objects. Creative nodes store the fused embedding of ad creative text, landing page summary, material tags, and campaign theme. Scene nodes store the embeddings corresponding to the campaign media, page position, and content theme. Time window nodes store the start and end positions of the window and the index of adjacent windows. The relation-aware graph neural network... Different types of connection edges are assigned independent edge type codes, and the contribution of adjacent nodes is calculated separately according to edge type during message aggregation. The user's exposure, clicks, conversions, and negative feedback to candidate ad objects are no longer mixed into the same count value, but are respectively entered into the fatigue component and response component under edge type constraints. The content ownership edge between the ad object and the creative node is used to transmit creative similarity, and the scene association edge between the ad object and the scene node is used to transmit the similarity of the delivery environment. The time window node is used to embed the state of the adjacent window into the current window. The advantage of this embodiment is that it enables the contact traces of the same user on different similar ad objects to have a graph propagation path, and retains the different sources of influence of different interaction types on frequency state.

[0022] In a graph message aggregation implementation, the sub-relationship messages of candidate ad object nodes within a time window can be represented by the following formula: ; in, This indicates that node v is within the time window. The message vector obtained from the edge type r is the inner edge. This represents the set of adjacent nodes connected to node v via edge type r. This represents the normalized weight when adjacent node u passes a message to node v. This represents the linear transformation matrix corresponding to edge type r. This indicates that the adjacent node u is within the time window. For example, when the edge type r is a negative feedback edge, the candidate ad object node connects two user nodes with two normalized weights of 0.60 and 0.40 respectively. The two adjacent nodes pass through... When the transformed one-dimensional example values ​​are 0.80 and 0.30 respectively, the message example value is 0.60×0.80+0.40×0.30=0.60, and this example value is written into the negative feedback message component of the corresponding node.

[0023] In this embodiment, the exposure memory network reads continuous exposure sequences using the user identifier as the primary key. The continuous exposure sequence includes exposure records of candidate ad objects, exposure records of creative neighboring ad objects, exposure records of non-neighboring ad objects, exposure occurrence time, post-exposure interaction feedback, and request context. The exposure sequences are arranged from earliest to latest according to the request time, and the time interval between adjacent exposures is encoded as a decay position vector. Exposure records with shorter time intervals retain a stronger near-temporal marker in the decay position vector, while those with longer time intervals retain a weaker near-temporal marker. The temporal attention encoder receives the ad object embedding, interaction feedback embedding, and decay position vector, calculates the correlation between the current candidate ad object and historical exposure records, and outputs an individual exposure memory representation. This individual exposure memory representation includes the user's different historical contact states with the candidate ad object itself, creative neighboring ad objects, and non-neighboring ad objects. The advantage of this embodiment is that the exposure state of the same user is no longer represented solely by the cumulative number of exposures within the window. Short-cycle dense contact and long-cycle scattered contact are mapped to different sequence memory structures, providing a time decay basis for subsequent gating discrimination.

[0024] In an exposure sequence attention implementation, the weighting of time decay of historical exposure records can be expressed by the following formula: ; in, This indicates the participation weight of the i-th historical exposure record in the memory encoding of the current candidate ad object. The query vector representing the candidate ad object. This represents the key vector of the i-th historical exposure record. This represents the vector inner product between the candidate ad object and its historical exposure records. Indicates the time decay coefficient. This represents the time interval between the i-th historical exposure record and the current request. This represents the number of historical exposure records. For example, when there are 2 historical exposure records, the inner products of the two vectors are 1.20 and 1.00 respectively, the time decay coefficient is 0.10, and the two time intervals are 1 and 5 respectively, the two unnormalized exponential inputs are 1.10 and 0.50 respectively, and the corresponding weights are approximately 0.65 and 0.35. The more recent exposure record has a higher participation weight in the memory encoding.

[0025] In this embodiment, the frequency control gating generator receives the correlation fatigue representation output by the relation-aware graph neural network and the individual exposure memory representation output by the exposure memory network. The representation alignment layer projects the two onto the same dimension and the same numerical range. The fusion gating layer generates participation coefficients based on the sample density of candidate ad objects, the number of historical exposures of users, and the number of creative nearest neighbors. The frequency interval discrimination layer outputs allowed exposure intervals, deweighted exposure intervals, or suspended exposure intervals based on the fused frequency risk representations. Candidate ad objects corresponding to allowed exposure intervals continue to participate in ranking, candidate ad objects corresponding to deweighted exposure intervals participate in subsequent decisions with lower ranking weights, and candidate ad objects corresponding to suspended exposure intervals are no longer displayed in the current scrolling window. The gating results are simultaneously written back to the state caches of user nodes and ad object nodes and used as historical inputs for graph construction and sequence organization when the next request arrives. The advantage of this embodiment is that graph-side cross-object fatigue propagation and sequence-side individual decay memory converge in the same gating space, and the gating results can update the frequency state in reverse, so that the frequency control judgment in continuous requests forms a closed loop.

[0026] In one implementation of fusion gating, the frequency risk representation after fusion can be expressed as the following formula: ; in, Indicates user u within the time window The fusion frequency risk characterization of the corresponding candidate advertisement object a. This represents the association fatigue representation output by the relation-aware graph neural network. This represents the individual's exposure memory representation output by the exposure memory network. The graph represents the participation coefficient. The sequence represents the participation coefficient; for example, when When the single-dimensional example value of the associated fatigue representation is 0.80 and the single-dimensional example value of the individual exposure memory representation is 0.40, the fusion example value is 0.70×0.80+(1-0.70)×0.40=0.68. This fusion value is used as one of the input components of the frequency interval discrimination layer.

[0027] In a preferred embodiment, refer to the appendix. Figure 2 The temporal heterogeneous graph builder includes a node mapper and an edge relationship generator. The node mapper maps users, ad objects, ad creatives, placement scenarios, and time slices appearing within the same scrolling time window as graph nodes with type identifiers. The edge relationship generator establishes exposure edges, click edges, conversion edges, and negative feedback edges between user nodes and ad object nodes, content ownership edges between ad object nodes and creative nodes, and scenario association edges between ad object nodes and scenario nodes. It also writes time slice tags and interaction result tags for each connection edge. Specifically, the node mapper assigns a type number and a local index to each type of node. Entities maintain stable entity indices within adjacent scrolling windows. Ad objects that do not appear in the window but have creative proximity relationships with candidate ad objects can be introduced as candidate adjacent nodes. The edge relationship generator generates multiple types of connection edges based on exposure receipts and interaction receipts. When the same user has multiple exposure records for the same ad object within the same time slice, the edge relationship generator compresses the exposure count, most recent exposure time, and feedback status into an edge attribute vector to avoid excessive duplicate edges causing aggregation skew. The advantage of this embodiment is that both nodes and edges have clear generation criteria, the graph structure can remain updatable within the scrolling window, and different interaction results are preserved at the edge attribute level. Table 1 below shows the relationship edge encoding table.

[0028] Table 1 Relationship Edge Encoding Table

[0029] In this embodiment, Table 1 illustrates the edge relationship generation rules in the temporal heterogeneous graph builder. Exposure edges, click edges, conversion edges, and negative feedback edges all connect user nodes and ad object nodes, but with different write tags. Content ownership edges connect ad object nodes and creative nodes, while scene association edges connect ad object nodes and scene nodes. The relationship-aware graph neural network reads the relationship edge name to determine the edge type encoding, reads the write tag to determine the edge attribute vector, and reads the participation processing field to determine whether the edge enters a fatigue state input, response state input, creative relationship input, or scene relationship input. The time slice tag is used for cross-window state updates, and the interaction result tag is used to distinguish between no feedback, positive feedback, and negative feedback after exposure. The node mapper and edge relationship generator use a common index to enable the content ownership edges, scene association edges, and user interaction edges of the same ad object to be jointly processed in the same graph instance. The advantage of this embodiment is that it provides a clear data structure for the graph neural network, avoiding the compression of ad exposure times, creative similarity, and scene co-occurrence relationships into a single feature field.

[0030] In a further preferred embodiment, the relationship-aware graph neural network includes an edge type encoding layer, a sub-relationship message aggregation layer, and a cross-window state update layer. The edge type encoding layer generates edge type vectors for exposure edges, click edges, conversion edges, negative feedback edges, content attribution edges, and scene association edges, respectively. The sub-relationship message aggregation layer calculates adjacent node messages according to the edge type vectors. The cross-window state update layer performs gating updates on the node messages of the current scrolling time window and the node embeddings of adjacent historical scrolling time windows to generate an associated fatigue representation that includes the similarity influence relationship of advertising objects and the user fatigue migration relationship. Specifically, the edge type encoding layer maintains a gating system for each type of relationship edge. Independently encoded vectors are used, and the edge attribute vectors and edge type vectors are concatenated and input into the sub-relationship message aggregation layer. The sub-relationship message aggregation layer performs adjacency sampling, weight normalization, and message transformation according to the relationship edge category. Exposure edges and negative feedback edges enter the fatigue aggregation channel, click edges and conversion edges enter the response aggregation channel, and content belonging edges and scene association edges enter the object relationship aggregation channel. The cross-window state update layer reads the embedding of the same node in the previous scrolling window and determines the retention ratio of historical embeddings and current messages through a gating coefficient. The advantage of this embodiment is that different relationship edges will not be mixed at the aggregation entry point, and the historical state of the same node will not be directly discarded due to window switching.

[0031] In this embodiment, the cross-window state update layer can synthesize the aggregated messages and historical node embeddings in the current window using a gating method. The gating coefficient is generated by the current window message intensity, historical window activity, and time interval. When the candidate ad object has few samples in the current window but many interaction records in the historical window, the historical node embedding retains a high level of participation in the update. When the candidate ad object has dense exposure and clear negative feedback in the current window, the current window message has a high level of participation in the update. The association fatigue representation output by the relation-aware graph neural network includes individual fatigue information transmitted from user nodes to candidate ad object nodes, creative neighbor information transmitted from creative nodes to ad object nodes, environmental similarity information transmitted from scene nodes to ad object nodes, and state continuation information transmitted from historical windows to the current window. The sub-relationship message aggregation layer can also set a higher fatigue component channel participation weight for negative feedback edges, so that user blocking, skipping, and other behaviors are not covered by click and conversion signals. The advantage of this embodiment is that repeated touches between similar creative ad objects can enter the candidate ad object state through content affiliation and neighbor connections, and low-sample ad objects can obtain supplementary representations from the graph neighborhood.

[0032] In a further preferred embodiment, refer to the appendix Figure 3 The exposure memory network includes an exposure sequence organizer, a time interval encoder, and a temporal attention encoder. The exposure sequence organizer extracts exposure records of candidate ad objects and their similar ad objects in multiple time slices according to user identifiers, and forms an exposure sequence sorted by request time. The time interval encoder maps the time interval between adjacent exposure records to a decay position vector. The temporal attention encoder jointly encodes the exposure sequence and the decay position vector, outputting an individual exposure memory representation that distinguishes between short-cycle dense exposure and long-cycle scattered exposure. Specifically, the exposure sequence organizer uses user identifiers and the current request time as retrieval conditions to extract exposure records in the current scrolling window and adjacent historical windows from the state cache, and writes the exposure records of similar ad objects of the candidate ad object into the sequence according to the creative nearest neighbor edge. The time interval encoder does not use a fixed window count to replace the time difference, but instead maps the actual interval between adjacent exposure records to a continuous decay position vector. The temporal attention encoder uses the candidate ad object as the query object and calculates the object relevance and time relevance of each exposure record in the historical sequence. The advantage of this embodiment is that the exposure sequence includes the candidate ad object itself and its nearest neighbor objects, which can reflect the user contact trajectory when similar ads appear repeatedly.

[0033] In this embodiment, the exposure sequence organizer performs deduplication and hierarchical retention of historical exposure records. Repeated exposure records of the same ad object within the same time slice are merged into a single record containing the number of exposures and the most recent exposure time. Different creative neighboring ad objects retain independent records, and non-neighboring ad objects are marked with a low participation flag according to their thematic distance from the candidate ad object. The time interval encoder calculates the exposure time by subtracting the historical exposure time from the current request time. The time interval, whether it is the same advertising object, whether it is a creative neighbor, and the interaction feedback type are concatenated into a sequence position input. The temporal attention encoder can use a gated recurrent unit or a self-attention structure to process the sequence input. When a self-attention structure is used, the candidate advertising object embedding is used as the query vector, the historical exposure record embedding is used as the key vector and value vector, and the decay position vector is used as the attention bias in the calculation. The output individual exposure memory representation is written into the frequency control gate generator. The advantage of this embodiment is that it avoids misprocessing long-period scattered exposures into short-period dense exposures and preserves the state order in the continuous request stream for a single user dimension.

[0034] In a preferred embodiment, the temporal heterogeneous graph builder further includes a creative semantic consistency processor. The creative semantic consistency processor extracts semantic embeddings from the ad creative text, landing page summary, creative tags, and ad placement theme. Based on the proximity relationships between the semantic embeddings, it generates creative nearest neighbor edges between ad object nodes and performs joint constraint encoding on the creative nearest neighbor edges and content attribution edges. This ensures that ad object nodes with the same type of creative share a fatigue propagation path in the relation-aware graph neural network, while ad object nodes with different themes retain independent sets of connection edges. Specifically, the creative semantic consistency processor inputs the ad creative text and landing page summary into a text encoding network, respectively. The process involves inputting material tags and campaign themes into a tag embedding layer, then weighting and aggregating various semantic vectors to form an ad object semantic embedding. Candidate nearest neighbor sets are generated through vector proximity retrieval among the ad object semantic embeddings. Creative nearest neighbor edges are established between ad object nodes in the nearest neighbor set. The edge weights of these creative nearest neighbor edges are determined by semantic embedding proximity, material tag overlap, and campaign theme consistency. Content attribution edges are used to confirm the basic correspondence between ad objects and creative nodes, while creative nearest neighbor edges express the similarity influence between different ad objects. The advantage of this embodiment is that it can preserve fatigue propagation paths when ad object numbers are changed or similar materials are split for deployment. The creative semantic consistency input is shown in Table 2 below.

[0035] Table 2 Creative Semantic Consistency Input Table

[0036] In this embodiment, Table 2 is used to illustrate the input fields, encoding methods, and writing locations of the creative semantic consistency processor. The ad creative text and landing page summary are entered into the ad object semantic embedding through text vector encoding. The material tags and the campaign theme are entered into the creative nearest neighbor weights through tag embedding encoding. Semantic proximity is used to express the similarity of creative content, and theme proximity is used to express the similarity of landing page content. Tag overlap and theme consistency are used to correct the reliability of the nearest neighbor edges. When two ad objects are similar in text content but have significantly different campaign themes, the creative nearest neighbor weights are reduced. When two ad objects have different numbers but their creative text, material tags, and campaign themes are similar, the creative nearest neighbor weights are retained. The relation-aware graph neural network reads the weights as message passing weights when processing creative nearest neighbor edges. The advantage of this embodiment is that similar creative relationships are not generated by manual grouping, but are formed by reproducible semantic encoding and weight calculation, and can directly participate in fatigue propagation.

[0037] In a further preferred embodiment, the relational message aggregation layer includes a negative feedback suppression channel and a positive interaction retention channel; the negative feedback suppression channel performs joint encoding on the negative feedback edge and its connected adjacent exposure edge, and writes the joint encoding result into the fatigue component of the ad object node; the positive interaction retention channel separates and aggregates the click edge and conversion edge, and writes the aggregation result into the response component of the ad object node; the cross-window state update layer generates the relational frequency state of the candidate ad object in the current request time slice based on the fatigue component and the response component. Specifically, the negative feedback suppression channel reads the user node... The negative feedback edge between the advertising object node is retrieved, and the adjacent exposure edge of the same user before the negative feedback occurs is retrieved. The time of negative feedback occurrence, the time interval between adjacent exposures, and the number of exposures are jointly encoded as fatigue input. The positive interaction retention channel establishes aggregation sub-channels for click edge and conversion edge respectively. The aggregation result of click edge retains the short link response state, and the aggregation result of conversion edge retains the long link response state. The fatigue component and response component are synthesized into a relational frequency state in the cross-window state update layer through gating. The advantage of this embodiment is that the negative feedback will not be canceled out by the positive interaction, and clicks and conversions will not be directly covered by the negative feedback channel.

[0038] In this embodiment, the negative feedback suppression channel writes feedback such as shielding, skipping, and closing into a unified negative feedback vector, and co-encodes the negative feedback vector with the time slice marker of the corresponding exposure edge. If the negative feedback occurs after a short period of dense exposure, the negative feedback vector and the adjacent exposure edge jointly form a higher fatigue component. If the interval between the negative feedback and the exposure is long, the near-temporal participation of the fatigue component is reduced. The positive interaction retention channel calculates response messages for the click edge and the conversion edge respectively. The click message is used to retain the immediate interaction, and the conversion message is used to retain the delayed interaction. The two are not added at the input layer, but are stored in the response component of the ad object node in different sub-vector positions. The cross-window state update layer generates a relational frequency state based on the fatigue component, the response component, and the historical node embedding. The relational frequency state then enters the fusion gating layer. The advantage of this embodiment is that multiple feedback signals have different storage locations and transmission paths inside the graph neural network, which can reduce the frequency discrimination shift caused by the mixing of heterogeneous feedback.

[0039] In a further preferred embodiment, the temporal attention encoder includes an ad object attention head and a time decay attention head; the ad object attention head assigns attention weights to candidate ad objects, creative neighbor ad objects, and non-neighbor ad objects respectively; the time decay attention head adjusts the participation weights of different historical exposure records in the user exposure sequence according to the decay position vector; the temporal attention encoder gates and concatenates the outputs of the two attention heads to form an individual exposure memory representation corresponding to the candidate ad object. Specifically, the ad object attention head sets the candidate ad object exposure records to the same object category, and assigns attention weights to the creative neighbor ad objects and non-neighbor ad objects respectively. The ad object exposure records with nearest neighbor edges are set as nearest neighbor object categories, while ad object exposure records without creative nearest neighbor edges are set as non-nearest neighbor object categories. Attention weights within categories and gating weights between categories are calculated separately. The time decay attention head reads the decay position vector and assigns higher near-temporal participation to records closer to the current request, while retaining lower participation for records farther from the current request. The outputs of the two attention heads are not directly added, but are used to generate a sequence memory in the gating splicing layer based on object similarity and time decay. The advantage of this embodiment is that the participation level of historical exposure records is constrained by both object relationships and time distance.

[0040] In this embodiment, the same-object output, nearest-object output, and non-nearest-object output generated by the ad object attention head are each written into three sub-vectors. The near-temporal output and far-temporal output generated by the time decay attention head are each written into two sub-vectors. The gating splicing layer reads five sub-vectors and generates splicing gating based on the semantic embedding of the candidate ad object, the current request time slice, and the user's recent feedback status. If the user has multiple exposures to the creative nearest-object ad recently and the feedback is weak, the nearest-object output and near-temporal output maintain a high level of participation in splicing. If the user only has a small number of non-nearest-object ad exposures in an earlier history window, the non-nearest-object output and far-temporal output maintain a low level of participation in splicing. The spliced ​​individual exposure memory representation and relational frequency status enter the representation alignment layer. The advantage of this embodiment is that the sequence memory does not process all historical exposure records in the same way, which can reduce the interference of irrelevant ad history on the frequency control judgment of candidate ads.

[0041] In a further preferred embodiment, the frequency control gating generator includes a representation alignment layer, a fusion gating layer, and a frequency interval discrimination layer; the representation alignment layer projects the relational frequency state and the individual exposure memory representation to the same feature space; the fusion gating layer generates a graph representation participation coefficient and a sequence representation participation coefficient based on the historical sample density of the candidate advertising object, and performs weighted fusion on the projected representations; the frequency interval discrimination layer outputs an allowed exposure interval, a reduced-weight exposure interval, or a suspended exposure interval based on the fused frequency risk representation. Specifically, the representation alignment layer performs linear operations on the relational frequency state and the individual exposure memory representation respectively. Projection, normalization, and activation processing ensure that the two have the same dimension and comparable numerical range. The fusion gating layer reads the historical exposure count of candidate ad objects, the historical exposure count of users, the number of creative nearest neighbors, the number of scene-related edges, and the recent feedback completeness to calculate the participation coefficient. The frequency interval discrimination layer maps the fusion frequency risk representation to the discrimination scores of three intervals and selects the interval with the highest score as the frequency control result of the current request. The advantage of this embodiment is that the participation relationship between graph representation and sequence representation changes according to the sample state, avoiding over-reliance on a single user sequence in a low sample state and ignoring the continuous exposure trajectory of individuals in a high sample state.

[0042] In a frequency interval discrimination implementation, the interval discrimination score can be expressed as the following formula: ; in, Indicates user u within the time window Let U be the interval probability vector corresponding to candidate ad object a, and b be the interval discrimination matrix. This represents the set of intervals consisting of the allowed exposure interval, the reduced-weight exposure interval, and the paused exposure interval. This represents the unnormalized score of interval c. For example, when the unnormalized scores of the three intervals are 1.00, 1.50 and 0.50 respectively, the normalized probabilities of the three intervals are approximately 0.31, 0.51 and 0.18. The frequency interval discrimination layer outputs the reduced-weight exposure interval as the frequency control result of the current candidate advertising object.

[0043] In a further preferred embodiment, the fusion gating layer further includes a sparse state recognizer; the sparse state recognizer generates sparse state codes based on the number of historical exposures of users, the number of historical exposures of candidate ad objects, the number of nearest neighbor edges of creatives, and the number of scene-related edges; the fusion gating layer increases the graph representation participation coefficient and limits the variation range of the sequence representation participation coefficient when the sparse state code represents low-sample states, and increases the sequence representation participation coefficient based on the continuous exposure sequence of users when the sparse state code represents high-sample states. Specifically, the sparse state recognizer reads the number of historical exposures of users and the number of historical exposures of candidate ad objects from the state cache and the temporal heterogeneous graph. The number of light bars, the number of creative nearest neighbors, and the number of scene-related edges are counted. These counts are logarithmically scaled and then concatenated with the feedback completeness code to generate sparse state codes for low, medium, or high samples. In the low sample state, the gating layer reads more of the association fatigue representation generated by the relation-aware graph neural network and suppresses the fluctuations in individual exposure memory representation caused by the short sequence. In the high sample state, the gating layer reads more of the individual exposure memory representation formed by the near-temporal records in the user's continuous exposure sequence. The advantage of this embodiment is that the frequency control gating has a stable input source when the sample size changes, and can supplement the frequency state with graph neighborhood in new advertising objects or new user scenarios.

[0044] In this embodiment, sparse state coding is not directly used as the frequency control interval output, but as an intermediate input for the fusion gating layer to generate participation coefficients. The number of historical exposures of users reflects the length of user sequences, the number of historical exposures of candidate ad objects reflects the sample density of the ad objects themselves, the number of creative nearest neighbor edges reflects the scale of information that can be borrowed from the graph neighborhood, and the number of scene-related edges reflects the coverage of the delivery environment relationship. When the number of historical exposures of users is small and the number of creative nearest neighbor edges is large, the fusion gating layer increases the graph representation participation coefficient. When the number of historical exposures of users is large and the number of historical exposures of candidate ad objects is large, the fusion gating layer increases the sequence representation participation coefficient. When the states of the two types of samples are close, the fusion gating layer maintains the intermediate participation ratio and the frequency interval discrimination layer completes the discrimination. Sparse state coding is also written into the cross-object frequency control consistency checker for use in correcting the gating values ​​between nearest neighbor ad objects. The advantage of this embodiment is that the sparsity of samples, the availability of graph neighborhoods, and the availability of user sequences are uniformly incorporated into the fusion control, avoiding frequency control deviations caused by a fixed fusion ratio.

[0045] In a further preferred embodiment, the frequency interval discrimination layer further includes a cross-object frequency control consistency checker; the cross-object frequency control consistency checker calls the creative nearest neighbor edge, the relational frequency state, and the sparse state encoding to perform consistency checks on the frequency intervals of the same user on candidate ad objects and their creative nearest neighbor ad objects; when the frequency interval of the candidate ad object is inconsistent with the frequency interval of the creative nearest neighbor ad object, the cross-object frequency control consistency checker re-determines the frequency control gating value of the candidate ad object based on the nearest neighbor edge weight and the graph representation participation coefficient. Specifically, after the frequency interval discrimination layer outputs the initial result, the cross-object frequency control consistency checker reads and... The candidate ad object has a set of ad objects with creative nearest neighbors. The verifier reads the most recent frequency control interval, relational frequency state, and feedback state of the same user on these nearest ad objects. If the candidate ad object is determined to be allowed to be exposed while multiple nearest ad objects are determined to be suspended, the verifier calculates and corrects the gate value according to the creative nearest neighbor edge weight and the graph representation participation coefficient. If the candidate ad object and the nearest ad object have a large difference in theme or the sparse state encoding indicates that the graph neighborhood is not reliable enough, the verifier reduces the correction strength of the nearest neighbor result to the candidate result. The advantage of this embodiment is that the frequency control results between similar ad objects can maintain the same source verification, avoiding the frequency control state fragmentation caused by object splitting.

[0046] In a cross-object consistency check implementation, the corrected gating value can be expressed as the following formula: ; in, This represents the corrected gating value for candidate ad targets. This represents the initial gate value for candidate ad objects. This represents the set of creative nearest neighbor ad objects for candidate ad object a. This represents the creative nearest neighbor edge weight between candidate ad object a and its nearest neighbor ad object b. The graph represents the participation coefficient. This indicates that neighboring ad object b is targeting the same user u within the same time window. For example, when the initial gate value of a candidate ad object is 0.40, there are two neighboring ad objects with edge weights of 0.80 and 0.50 respectively, the graph representation participation coefficient is 0.60, and the gate values ​​of the two neighbors are 0.90 and 0.70 respectively, the corrected gate value is (0.40+0.80×0.60×0.90+0.50×0.60×0.70)÷(1+0.80×0.60+0.50×0.60)≈0.64. This corrected value then enters the interval mapping process.

[0047] In this embodiment, the cross-object frequency control consistency checker performs nearest neighbor result verification before mapping the gate value to the interval. The verification input includes the initial gate value of the candidate ad object, the creative nearest neighbor edge weight, the graph representation participation coefficient, the sparse state code, and the nearest gate value of the nearest ad object. If the sparse state code shows that there are few candidate ad object samples but a sufficient number of creative nearest neighbor edges, the nearest ad object gate value maintains a high participation in the correction. If the sparse state code shows that the nearest neighbor relationship is insufficient or the theme difference is large, the correction process mainly retains the initial gate value of the candidate ad object. The checker also writes the difference in gate values ​​before and after correction into the state cache for the cross-window state update layer of the next scrolling window to read. After obtaining the corrected gate value, the frequency interval discrimination layer remaps it to the allowed exposure interval, the reduced exposure interval, or the paused exposure interval. The advantage of this embodiment is that the frequency control judgment between creative nearest ad objects can constrain each other at the gate value level, while retaining the control of the constraint strength by the sparse state, avoiding unnecessary frequency control entanglement between unrelated ad objects.

[0048] In this embodiment, refer to the appendix Figure 1 The complete process consists of ad request arrival, sample preprocessing, graph instance updating, exposure sequence organization, graph neural network inference, exposure memory inference, fusion gating, interval discrimination, consistency verification, and state write-back. After the ad request arrives, the sample preprocessor encodes the current request field and reads the historical state of the corresponding user and candidate ad object. The temporal heterogeneous graph builder updates nodes and connecting edges according to the current scrolling window. The relation-aware graph neural network outputs the association fatigue representation. The exposure memory network synchronously outputs the individual exposure memory representation. The frequency control gating generator adjusts the participation ratio of graph representation and sequence representation according to the sparse state encoding and outputs the initial frequency control interval. The cross-object frequency control consistency checker corrects the initial result according to the creative nearest neighbor edge. The corrected result enters the final interval output. After the exposure result and subsequent interaction feedback are returned, they are written back to the user node, ad object node, related connecting edge, and exposure sequence cache. The advantage of this embodiment is that each processing link shares the same entity index, time window index, and edge relationship index. A traceable data closed loop is formed between the graph structure, sequence structure, and gating result, which can support the stable operation of deep learning-based ad exposure frequency control in a continuous request stream.

Claims

1. A deep learning-based intelligent optimization system for ad exposure frequency, characterized in that, Includes a sample preprocessor, a temporal heterogeneous graph builder, a relation-aware graph neural network, an exposure memory network, and a frequency control gating generator; The sample preprocessor normalizes and encodes the user identifier, ad object identifier, creative features, delivery scenario features, request time slice, exposure results, and interaction feedback in the ad request log; The temporal heterogeneous graph builder generates a heterogeneous graph containing user nodes, advertising object nodes, creative nodes, scene nodes, and time window nodes according to a rolling time window. The relation-aware graph neural network outputs an association fatigue representation of candidate advertising objects based on different types of connection edges; The exposure memory network outputs an individual exposure memory representation based on the user's continuous exposure sequence; The frequency control gating generator integrates the associated fatigue characterization and the individual exposure memory characterization to generate the exposure frequency control result of the candidate advertising object.

2. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 1, characterized in that, The temporal heterogeneous graph builder includes a node mapper and an edge relation generator; The node mapper maps users, advertising objects, advertising creatives, delivery scenarios, and time segments that appear within the same scrolling time window to graph nodes with type identifiers. The edge relationship generator establishes exposure edges, click edges, conversion edges, and negative feedback edges between user nodes and advertising object nodes, establishes content ownership edges between advertising object nodes and creative nodes, establishes scene association edges between advertising object nodes and scene nodes, and writes time segmentation tags and interaction result tags for each connection edge.

3. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 2, characterized in that, The relation-aware graph neural network includes an edge type encoding layer, a relation-specific message aggregation layer, and a cross-window state update layer; The edge type encoding layer generates edge type vectors for exposure edges, click edges, conversion edges, negative feedback edges, content ownership edges, and scene association edges, respectively. The sub-relationship message aggregation layer calculates the adjacent node messages according to the edge type vector; The cross-window state update layer performs gated updates on the node messages of the current scrolling time window and the node embeddings of adjacent historical scrolling time windows, generating an associated fatigue representation that includes the similarity influence relationship of advertising objects and the user fatigue migration relationship.

4. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 3, characterized in that, The exposure memory network includes an exposure sequence organizer, a time interval encoder, and a temporal attention encoder; The exposure sequence organizer extracts the exposure records of candidate advertising objects and their similar advertising objects in multiple time segments according to the user identifier, and forms an exposure sequence sorted by the requested time. The time interval encoder maps the time interval between adjacent exposure records into a decay position vector; The temporal attention encoder jointly encodes the exposure sequence and the decay position vector, outputting an individual exposure memory representation that distinguishes between short-period dense exposures and long-period dispersed exposures.

5. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 2, characterized in that, The temporal heterogeneous graph builder also includes a creative semantic consistency processor; The creative semantic consistency processor extracts the semantic embeddings of the advertising creative text, landing page summary, material tags, and campaign theme. Based on the adjacency relationship between the semantic embeddings, it generates creative nearest neighbor edges between advertising object nodes and performs joint constraint encoding on the creative nearest neighbor edges and the content attribution edges. This enables advertising object nodes with the same type of creative to share the fatigue propagation path in the relation-aware graph neural network, while advertising object nodes with different themes retain independent sets of connection edges.

6. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 3, characterized in that, The sub-relationship message aggregation layer includes a negative feedback suppression channel and a positive interaction retention channel; The negative feedback suppression channel performs joint encoding on the negative feedback edge and its connected adjacent exposure edge, and writes the joint encoding result into the fatigue component of the advertising object node; The positive interaction retention channel separates and aggregates the click edge and the conversion edge, and writes the aggregation result into the response component of the ad object node; The cross-window state update layer generates a relational frequency state of candidate advertising objects in the current request time slice based on the fatigue component and the response component.

7. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 4, characterized in that, The temporal attention encoder includes an advertising object attention head and a time decay attention head; The advertising object attention head assigns attention weights to candidate advertising objects, creative neighboring advertising objects, and non-neighboring advertising objects respectively; The time decay attention head adjusts the participation weight of different historical exposure records in the user's exposure sequence according to the decay position vector; The temporal attention encoder gates and splices the outputs of the two attention heads to form an individual exposure memory representation corresponding to the candidate advertising object.

8. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 6, characterized in that, The frequency control gating generator includes a characterization alignment layer, a fusion gating layer, and a frequency interval discrimination layer; The representation alignment layer projects the relational frequency state and the individual exposure memory representation to the same feature space; The fusion gating layer generates a graph representation participation coefficient and a sequence representation participation coefficient based on the historical sample density of the candidate advertising objects, and performs weighted fusion on the projected representations. The frequency interval discrimination layer outputs the allowed exposure interval, the reduced-weight exposure interval, or the suspended exposure interval based on the fused frequency risk characterization.

9. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 8, characterized in that, The fusion gating layer also includes a sparse state recognizer; The sparse state recognizer generates sparse state codes based on the number of historical exposures of the user, the number of historical exposures of the candidate ad object, the number of creative nearest neighbor edges, and the number of scene-related edges. The fusion gating layer increases the graph representation participation coefficient when sparse state encoding represents low-sample states and limits the variation range of the sequence representation participation coefficient. When sparse state encoding represents high-sample states, it increases the sequence representation participation coefficient based on the user's continuous exposure sequence.

10. The intelligent optimization system for advertising exposure frequency based on deep learning according to claim 9, characterized in that, The frequency interval discrimination layer also includes a cross-object frequency control consistency checker; The cross-object frequency control consistency checker calls the creative nearest neighbor edge, the relational frequency state and the sparse state code to perform consistency check on the frequency interval of the same user on the candidate ad object and its creative nearest neighbor ad object; When the frequency range of a candidate ad object is inconsistent with the frequency range of a creative neighbor ad object, the cross-object frequency control consistency checker re-determines the frequency control gating value of the candidate ad object based on the neighbor edge weight and the graph representation participation coefficient.

Images