Advertisement promotion demand prediction method and system based on deep learning

Abstract

The application relates to the technical field of advertisement delivery, and discloses an advertisement promotion demand prediction method and system based on deep learning, wherein the advertisement promotion demand prediction method based on deep learning projects cross-platform delivery behavior data of an advertisement owner to a unified semantic space through a cross-platform semantic alignment network to generate an intention matching degree score, carries out supply side feature migration based on advertisement position similarity to generate a cold start baseline demand prediction value, filters cross-platform behavior signals with real causal effects by using a causal inference module, combines an intention-driven demand estimation value and a supply side baseline prediction value through an adaptive fusion network, and carries out online adjustment based on actual bidding data to output a smooth transition demand prediction result.

Description

Technical Field

[0001] This invention relates to the field of advertising delivery technology, and more specifically, to a method and system for predicting advertising promotion demand based on deep learning. Background Technology

[0002] In the field of advertising technology, when an advertising platform opens a new ad slot, this ad slot has no historical bidding data. However, the platform needs to provide demand forecasts from the initial opening stage to set the bidding floor and recommend the ad slot to advertisers. In existing technologies, cold start methods construct a baseline for demand forecasting by calculating the multi-dimensional similarity between the new ad slot and existing ad slots and migrating virtual historical bidding data from highly similar existing ad slots. Cross-platform intent prediction methods, on the other hand, obtain summaries of advertisers' advertising behavior on other platforms through cross-platform data interfaces, and predict advertisers' advertising intent on this platform after semantic alignment and attention encoding.

[0003] However, cold start methods based purely on the migration of similar ad placement features only utilize the supply-side similarity of ad placements and lack predictive basis at the demand-side advertiser intent level. Advertisers bidding on similar ad placements may not necessarily be willing to place ads on new ad placements. Cross-platform intent prediction methods can perceive advertiser intent, but cannot determine whether these intents will point to a new ad placement without historical data. There is semantic and temporal heterogeneity between cross-platform behavioral data and local platform data, and direct fusion will introduce noise. There is an uncertainty bias between virtual historical data and real data, which may lead to inaccurate predictions. Summary of the Invention

[0004] This invention provides a method and system for predicting advertising demand based on deep learning, which solves the technical problems in related technologies, such as inaccurate demand prediction due to the lack of historical data for new advertising slots, inability to effectively identify potential advertisers, and large prediction deviations during the cold start phase.

[0005] This invention discloses a method for predicting advertising promotion demand based on deep learning, comprising the following steps: Obtain metadata features and traffic sampling data for the new ad placement, and encode them into a comprehensive feature vector for the new ad placement; obtain summary data of each advertiser's placement behavior on other ad platforms through cross-platform data interfaces, and generate a cross-platform placement behavior summary sequence for each advertiser; The cross-platform campaign behavior summary sequence of each advertiser is standardized and encoded, and then processed by a cross-platform semantic alignment network to generate a unified cross-platform behavior embedding sequence. The intent matching score between the projection vector of each advertiser and the projection vector of the new ad space is calculated, and advertisers with intent matching scores exceeding an adaptive threshold are selected to generate a list of potential demand advertisers. Calculate the multi-dimensional similarity between the comprehensive feature vector of the new ad slot and the comprehensive feature vector of each existing ad slot. Select the existing ad slots with the highest similarity ranking as the source ad slot set. Combine the historical bidding features of each source ad slot in the source ad slot set according to the similarity weight to synthesize virtual historical bidding features and generate the cold start baseline demand prediction value. Based on the intent matching score, the historical delivery intensity features of each potential demand advertiser are extracted and scaled to generate an expected delivery intensity estimate. The expected delivery intensity estimates are then aggregated to obtain the intent-driven demand estimate. An adaptive fusion network is used to merge the intent-driven demand estimate with the cold start baseline demand prediction to output the new ad space demand prediction. The cross-platform semantic alignment network is a shared projection layer that embeds and projects the behaviors of different platforms into a unified semantic space.

[0006] Furthermore, the metadata features include ad placement type, media app category, page location code, and expected visibility index; the traffic sampling data includes user profile distribution vector and traffic time period distribution curve; the cross-platform advertising behavior summary data includes changes in ad activity status, advertising category change records, and advertising scale level changes on each platform.

[0007] Furthermore, the training of the cross-platform semantic alignment network adopts a supervised learning method, using advertisers known to have placed ads on this platform as positive samples and advertisers not placed ads on this platform as negative samples. The input is a concatenated vector of cross-platform placement behavior summary sequences of each advertiser after processing by the embedding layer. The binary classification prediction result of the placement intent is output through a fully connected layer, and the loss function adopts the binary cross-entropy loss function.

[0008] Furthermore, the calculation of the multi-dimensional similarity includes: using category matching scores for the ad placement type and media category dimensions; using the negative value of Jensen-Shannon divergence for the user profile distribution dimension; using the dynamic time warping algorithm to calculate the similarity for the traffic time period distribution dimension; and weighting and fusing the similarities of each dimension through the learned dimension weights.

[0009] Furthermore, it also includes: introducing an uncertainty label vector to the virtual historical bidding features, wherein the values ​​of each dimension of the uncertainty label vector are determined according to the variance of the corresponding feature dimension values ​​in the source ad slot set; after processing by the uncertainty perception attention layer, the attention weight is calculated according to the uncertainty variance of each feature dimension, and the feature dimensions with high uncertainty are assigned low weights.

[0010] Furthermore, before generating the expected delivery intensity estimate, the process includes: using a causal inference module to estimate the causal effect of cross-platform behavioral signals of each potential demand advertiser on the delivery intent of this platform; calculating the difference in intent prediction under the two conditions of having cross-platform behavioral signals and not having cross-platform behavioral signals through counterfactual reasoning; marking cross-platform behavioral signals with significant causal effects as strong causal signals and assigning them enhanced weights; and reducing the weights of cross-platform behavioral signals with insignificant causal effects to generate a causal enhanced intent representation vector for each potential demand advertiser.

[0011] Furthermore, the causal inference module adopts a dual-head neural network structure, including a shared feature extraction layer and two independent prediction heads. The first prediction head takes into input a concatenated vector of the shared feature representation vector and the cross-platform behavior signal and outputs the intention prediction value under the condition of having cross-platform behavior signal. The second prediction head takes into input a concatenated vector of the shared feature representation vector and the zero vector and outputs the intention prediction value under the condition of not having cross-platform behavior signal.

[0012] Furthermore, it also includes: embedding the causal enhancement intent representation vector of potential demand advertisers and cross-platform behavior into a sequence input heterogeneous event cross-attention encoder to predict the expected time distribution of each potential demand advertiser's start of advertising, and generating an advertising intent maturity curve; multiplying the advertising intent maturity curve of each potential demand advertiser with the expected advertising intensity estimate and summing them by the time dimension to generate a time dimension demand influx prediction curve for new ad slots; the heterogeneous event cross-attention encoder learns an independent attention weight matrix for the combination of event types on this platform and cross-platform while retaining a continuous time decay mechanism.

[0013] Furthermore, it also includes: as new ad slots accumulate actual bidding data, by utilizing the hit rate of the actual bidder identities and potential demand advertiser lists, and the deviation between the actual bidding volume and the predicted demand value of the new ad slots, the adaptive threshold of intent matching score and the weight of the adaptive fusion network are adjusted through online learning methods. The weight of virtual historical bidding features and intent-driven demand estimates in feature calculation is gradually reduced according to a preset decay coefficient until they are completely replaced by actual bidding data, and the demand prediction results are output, smoothly transitioning from the cold start stage to the regular prediction stage.

[0014] This invention provides a deep learning-based advertising demand prediction system, comprising: The feature acquisition module is used to acquire metadata features and traffic sampling data of the new ad slot and encode them into a comprehensive feature vector of the new ad slot, as well as to acquire the cross-platform delivery behavior summary sequence of each advertiser through the cross-platform data interface; The intent matching module is used to standardize and encode the cross-platform advertising behavior summary sequence, calculate the intent matching score after processing by the cross-platform semantic alignment network, and filter and generate a list of potential advertisers. The supply-side migration module is used to calculate the multi-dimensional similarity between new ad slots and existing ad slots, select the source ad slot set and synthesize virtual historical bidding features to generate cold start baseline demand forecast values; The demand fusion module generates an estimate of the expected delivery intensity based on the intent matching score and aggregates them to obtain an intent-driven demand estimate. An adaptive fusion network then merges the intent-driven demand estimate with the cold-start baseline demand forecast to output the new ad placement demand forecast. The process involves: acquiring metadata features and traffic sampling data for the new ad placement, encoding them into a comprehensive feature vector; obtaining summary data of each advertiser's campaign behavior on other ad platforms through a cross-platform data interface, generating a cross-platform campaign behavior summary sequence for each advertiser; standardizing and encoding the cross-platform campaign behavior summary sequences of each advertiser, processing them through a cross-platform semantic alignment network to generate a unified cross-platform behavior embedding sequence, calculating the intent matching score between each advertiser's projection vector and the new ad placement's projection vector, and selecting advertisers with intent matching scores exceeding an adaptive threshold to generate a list of potential demand advertisers; calculating the multi-dimensional similarity between the comprehensive feature vector of the new ad placement and the existing comprehensive feature vectors of each ad placement, and selecting... The existing ad slots with the highest similarity ranking are selected as the source ad slot set. The historical bidding features of each source ad slot in the source ad slot set are weighted according to similarity weights to synthesize virtual historical bidding features, generating a cold start baseline demand prediction value. Based on the intent matching score, the historical delivery intensity features of each potential demand advertiser are extracted and scaled to generate an expected delivery intensity estimate. The expected delivery intensity estimates are aggregated to obtain the intent-driven demand estimate. The intent-driven demand estimate is merged with the cold start baseline demand prediction value through an adaptive fusion network to output the new ad slot demand prediction value. The cross-platform semantic alignment network is a shared projection layer that embeds the behavior of different platforms into a unified semantic space.

[0015] Furthermore, the metadata features include ad placement type, media app category, page location code, and expected visibility index; the traffic sampling data includes user profile distribution vector and traffic time period distribution curve; the cross-platform delivery behavior summary data includes changes in ad activity status, delivery category change records, and delivery scale level changes on each platform.

[0016] Furthermore, the training of the cross-platform semantic alignment network adopts a supervised learning method, using advertisers known to have placed ads on this platform as positive samples and advertisers not placed ads on this platform as negative samples. The input is a concatenated vector of cross-platform placement behavior summary sequences of each advertiser after processing by the embedding layer. The binary classification prediction result of the placement intent is output through a fully connected layer, and the loss function adopts the binary cross-entropy loss function.

[0017] Furthermore, the calculation of the multi-dimensional similarity includes: using category matching scores for the ad placement type and media category dimensions; using the negative value of Jensen-Shannon divergence for the user profile distribution dimension; using the dynamic time warping algorithm to calculate the similarity for the traffic time period distribution dimension; and weighting and fusing the similarities of each dimension through the learned dimension weights.

[0018] Furthermore, it also includes: introducing an uncertainty label vector to the virtual historical bidding features, wherein the values ​​of each dimension of the uncertainty label vector are determined based on the variance of the corresponding feature dimension values ​​in the source ad slot set; after processing by the uncertainty perception attention layer, the attention weight is calculated based on the uncertainty variance of each feature dimension, and lower weights are assigned to feature dimensions with higher uncertainty.

[0019] Furthermore, before generating the expected delivery intensity estimate, the process includes: using a causal inference module to estimate the causal effect of cross-platform behavioral signals of each potential demand advertiser on the delivery intent of this platform; calculating the difference in intent prediction under the two conditions of having cross-platform behavioral signals and not having cross-platform behavioral signals through counterfactual reasoning; marking cross-platform behavioral signals with significant causal effects as strong causal signals and assigning them enhanced weights; reducing the weights of cross-platform behavioral signals with insignificant causal effects; and generating a causal enhanced intent representation vector for each potential demand advertiser.

[0020] Furthermore, the causal inference module adopts a dual-head neural network structure, including a shared feature extraction layer and two independent prediction heads. The first prediction head takes into input a concatenated vector of the shared feature representation vector and the cross-platform behavior signal and outputs the intention prediction value under the condition of having cross-platform behavior signal. The second prediction head takes into input a concatenated vector of the shared feature representation vector and the zero vector and outputs the intention prediction value under the condition of not having cross-platform behavior signal.

[0021] Furthermore, it also includes: embedding the causal enhancement intent representation vector of potential demand advertisers and cross-platform behavior into a sequence input heterogeneous event cross-attention encoder to predict the expected time distribution of each potential demand advertiser's start of advertising, and generating an advertising intent maturity curve; multiplying the advertising intent maturity curve of each potential demand advertiser with the expected advertising intensity estimate and summing them by time dimension to generate a time dimension demand influx prediction curve for new ad slots; the heterogeneous event cross-attention encoder learns an independent attention weight matrix for the combination of local and cross-platform event types and retains a continuous time decay mechanism.

[0022] Furthermore, it also includes: as new ad slots accumulate actual bidding data, by utilizing the hit rate of the actual bidder identities and potential demand advertiser lists, and the deviation between the actual bidding volume and the predicted demand value of the new ad slots, the adaptive threshold of the intent matching score and the weight of the adaptive fusion network are adjusted through online learning methods. The weight of virtual historical bidding features and intent-driven demand estimates in feature calculation is gradually reduced according to a preset decay coefficient until they are completely replaced by actual bidding data, and the demand prediction results are output, smoothly transitioning from the cold start stage to the regular prediction stage.

[0023] This invention discloses a deep learning-based advertising demand prediction system, comprising: a feature acquisition module, used to acquire metadata features and traffic sampling data of new ad slots and encode them into a comprehensive feature vector of the new ad slots, and to acquire cross-platform delivery behavior summary sequences of each advertiser through a cross-platform data interface; an intent matching module, used to standardize and encode the cross-platform delivery behavior summary sequences, calculate the intent matching score after processing by a cross-platform semantic alignment network, and filter to generate a list of potential demand advertisers; a supply-side migration module, used to calculate the multi-dimensional similarity between the new ad slots and existing ad slots, select a set of source ad slots and synthesize virtual historical bidding features to generate a cold start baseline demand prediction value; and a demand fusion module, used to generate an expected delivery intensity estimate based on the intent matching score and summarize it to obtain an intent-driven demand estimate, and to merge the intent-driven demand estimate with the cold start baseline demand prediction value through an adaptive fusion network to output the new ad slot demand prediction value.

[0024] This invention eliminates semantic heterogeneity between behavioral data from different platforms through a cross-platform semantic alignment network, accurately calculating the matching degree between advertiser intent and new ad space features in a unified semantic space, thus supplementing the demand-side advertiser intent signals. A causal inference module distinguishes between cross-platform behavioral signals that truly influence intent and noise signals, effectively suppressing noise interference introduced during direct fusion of heterogeneous cross-platform data. An uncertainty-aware attention layer quantifies the reliability of each dimension of virtual historical bidding features, reducing the excessive impact of uncertainty on predicted values. An adaptive fusion network dynamically adjusts the contribution ratio of intent-driven demand estimates to supply-side baseline predictions based on the sufficiency of intent signals, overcoming the incomplete information dimensions of single methods. A heterogeneous event cross-attention encoder captures heterogeneous correlation patterns between cross-platform behavioral events and placement timing, improving the accuracy of the time-dimensional demand influx prediction curve. An online learning method achieves a smooth transition from virtual historical bidding features to actual bidding data, ensuring the continuity of demand prediction from the cold start phase to the regular prediction phase, solving the technical problem of insufficient demand prediction accuracy during the cold start phase of new ad spaces. Attached Figure Description

[0025] Figure 1This is a flowchart of the deep learning-based advertising demand prediction method provided in an embodiment of the present invention; Figure 2 This is a sub-flowchart of the deep learning-based advertising promotion demand prediction method provided in the embodiments of the present invention, which integrates intent-driven demand estimation and supply-side baseline. Figure 3 This is a dimensional distribution diagram of the comprehensive feature vector of the new advertising space provided in this embodiment of the invention; Figure 4 This is an advertising idea image matching score distribution diagram provided in an embodiment of the present invention; Figure 5 This is a multi-dimensional similarity comparison chart of source ad slots provided in an embodiment of the present invention; Figure 6 This is a cross-platform behavioral signal causality evaluation diagram provided in an embodiment of the present invention; Figure 7 This is an advertisement intent maturity curve provided in an embodiment of the present invention; Figure 8 This is a time-dimensional demand inflow prediction curve provided in an embodiment of the present invention; Figure 9 This is a diagram illustrating the online parameter adjustment and dynamic optimization process provided in this embodiment of the invention. Detailed Implementation

[0026] In the field of advertising technology, when an advertising platform opens a new ad slot, this slot lacks any historical bidding data. However, the platform needs to provide demand forecasts from the outset to set the minimum bidding price and recommend the ad slot to advertisers. Existing technologies employ cold-start methods, which calculate multi-dimensional similarity between the new ad slot and existing ad slots, migrating virtual historical bidding data from highly similar existing ad slots to construct a baseline for demand forecasting. Cross-platform intent prediction methods, on the other hand, obtain summaries of advertisers' campaign behavior on other platforms through cross-platform data interfaces, and predict advertisers' campaign intent on this platform after semantic alignment and attention encoding.

[0027] However, the aforementioned existing technologies have the following technical problems: Cold start methods based purely on the feature migration of similar ad placements only utilize the supply-side similarity at the ad placement level, lacking predictive basis at the demand-side advertiser intent level. Advertisers bidding on similar ad placements may not necessarily be willing to place orders for new ad placements, as their placement decisions also depend on their overall cross-platform placement strategy and budget allocation plan. Furthermore, while cross-platform intent prediction methods can perceive advertiser intent, they cannot determine whether these intents will point to a new ad placement with no historical data. In addition, there is semantic and temporal heterogeneity between cross-platform behavioral data and local platform data; direct fusion will introduce noise. Uncertainty and deviation exist between virtual historical data and real data, potentially leading to inaccurate predictions.

[0028] At least one embodiment of the present invention discloses a method for predicting advertising promotion demand based on deep learning, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the comprehensive feature vector of the new ad slot and the summary sequence of cross-platform advertising behavior of each advertiser; Obtain metadata features and traffic sampling data from the initial opening phase of the new ad slot, and encode the metadata features and traffic sampling data into a comprehensive feature vector for the new ad slot; obtain publicly observable ad placement behavior summary data of each advertiser on other ad platforms through cross-platform data interfaces authorized by advertisers, and generate cross-platform ad placement behavior summary sequences for each advertiser.

[0029] It should be noted that the aforementioned metadata features include ad placement type, media app category, page location code, and expected visibility metric. The aforementioned traffic sampling data includes user profile distribution vectors and traffic time-period distribution curves. The user profile distribution vectors represent the distribution characteristics of the user group visiting the ad placement across dimensions such as age, gender, region, and interests. The traffic time-period distribution curves represent the traffic variation patterns of the ad placement throughout the day.

[0030] It should be noted that the above cross-platform advertising behavior summary data includes changes in ad activity status, ad category change records, and ad scale level changes on each platform. Changes in ad activity status refer to events such as an advertiser going from never advertising on a platform to starting advertising, or from active advertising to paused advertising; ad category change records refer to events where the category of goods or services advertised by the advertiser changes; and ad scale level changes refer to events where the advertiser's average daily spending changes from one level to another.

[0031] It should be noted that before encoding into the new ad space comprehensive feature vector, the categorical data in the metadata features is converted into a numerical vector using one-hot encoding. The numerical features are then standardized using Z-score to eliminate the dimensional differences between different features. Finally, the standardized features of each dimension are concatenated to form the new ad space comprehensive feature vector.

[0032] Step 2: Perform semantic alignment processing on the cross-platform advertising behavior summary sequence to generate a list of potential advertisers and their intent matching scores; Each event in the cross-platform campaign summary sequence of each advertiser is standardized and encoded. The platform identifier, behavior type, category tag, and scale level are mapped through independent embedding layers and then concatenated. After processing by a cross-platform semantic alignment network, a unified embedding sequence of cross-platform behaviors for each advertiser is generated. The comprehensive feature vector of the new ad placement is also projected onto the same semantic space through the cross-platform semantic alignment network. The cosine similarity algorithm is used to calculate the similarity between the projection vector of each advertiser and the projection vector of the new ad placement as the intent matching score. Advertisers whose intent matching scores exceed an adaptive threshold are selected as the set of potential demand advertisers, generating a list of potential demand advertisers and their intent matching scores.

[0033] It should be noted that the aforementioned cross-platform semantic alignment network is a shared projection layer, which embeds the behaviors of different platforms into a unified semantic space.

[0034] Furthermore, the aforementioned shared projection layer performs a linear transformation on the behavioral embeddings of each platform through a projection matrix and adds a bias vector, mapping the behavioral embeddings of each platform into a semantic space of a unified dimension, so that behavioral features from different platforms can be compared and computed in the same vector space.

[0035] Furthermore, the training of the aforementioned cross-platform semantic alignment network adopts a supervised learning approach, using advertisers known to have placed ads on this platform as positive samples and advertisers not placed ads on this platform as negative samples. The input is a concatenated vector of the cross-platform placement behavior summary sequence of each advertiser after processing by the embedding layer. The fully connected layer outputs the binary classification prediction result of the advertiser's placement intention on this platform, and the loss function adopts the binary cross-entropy loss function.

[0036] It should be noted that the above intent matching score is calculated as follows: the advertiser's cross-platform behavior is uniformly embedded into a sequence and averaged to obtain the advertiser's projection vector. The projection vector of the new ad placement is recorded as the feature vector of the new ad placement. The cosine similarity algorithm is used to calculate the intent matching score.

[0037] It should be noted that the above adaptive threshold is dynamically determined based on the distribution characteristics of intent matching scores in the set of potential demand advertisers.

[0038] Furthermore, the aforementioned adaptive threshold can be the mean plus a standard deviation of the intent matching score distribution, or the [missing value] of the intent matching score distribution. Percentiles are used as adaptive thresholds, where These are the preset percentile parameters.

[0039] Step 3: Perform supply-side feature transfer based on ad placement similarity to generate cold start baseline demand forecast values; Calculate the multi-dimensional similarity between the comprehensive feature vector of the new ad placement and the comprehensive feature vectors of all existing ad placements. Then, select the top-ranked similarity by weighting and fusing the learned dimensional weights. The existing ad slots are used as the source ad slot set, among which The number of source ad slots is preset; the historical bidding characteristics of each source ad slot in the source ad slot set are weighted and averaged according to similarity weights to synthesize the virtual historical bidding characteristics of the new ad slots, generating the cold start baseline demand forecast value for supply-side migration.

[0040] It should be noted that the above multi-dimensional similarity calculation includes the following dimensions: the ad placement type and media category dimensions use category matching scores. When the new ad placement is the same as the source ad placement in terms of type or category, the category matching score is 1; otherwise, the category matching score is 0 or a partial matching score is calculated based on the category hierarchy. The user profile distribution dimension uses the negative value of Jensen-Shannon divergence. The traffic time period distribution dimension uses the dynamic time warping algorithm to calculate similarity.

[0041] Furthermore, the Jensen-Shannon divergence mentioned above is used to measure the similarity between two probability distributions. By calculating the information difference between the user profile distributions of the new ad placement and the source ad placement, the negative value of the Jensen-Shannon divergence is taken as the similarity score. The smaller the Jensen-Shannon divergence value, the more similar the two distributions are, and the larger the negative value, the higher the similarity.

[0042] It should be noted that the above-mentioned virtual historical bidding features are synthesized as follows: the historical bidding feature vectors of each source ad in the source ad set are normalized and weighted according to the multi-dimensional similarity between the source ad and the new ad. The source ad with higher similarity contributes more weight in the synthesis.

[0043] In this embodiment of the application, in order to address the reliability deviation problem of virtual historical data, the following steps are also included: the demand substitution elasticity coefficients between each source ad slot and other ad slots in the source ad slot set are weighted and averaged according to similarity weights to synthesize the virtual substitution elasticity coefficient of the new ad slot; an uncertainty label vector is introduced into the virtual historical bidding features; and after processing by the uncertainty perception attention layer, the feature dimensions with higher uncertainty are assigned lower weights.

[0044] Furthermore, the aforementioned demand substitution elasticity coefficient characterizes the degree of impact on the demand for other ad slots when the price of a certain ad slot changes. By weighted averaging the demand substitution elasticity coefficients of each source ad slot in the source ad slot set, the competitive substitution relationship between the new ad slot and other ad slots can be estimated.

[0045] Furthermore, the values ​​of each dimension of the aforementioned uncertainty marker vector are determined based on the variance of the corresponding feature dimension values ​​in the source ad space set; the larger the variance, the higher the uncertainty.

[0046] Furthermore, the aforementioned uncertainty-perceiving attention layer calculates attention weights based on the uncertainty variance of each feature dimension. Dimensions with higher uncertainty are assigned lower attention weights, thereby reducing the impact of unreliable information during feature weighting. The attention weights are obtained by taking the negative exponent of the uncertainty variance and normalizing it after adjusting for temperature parameters.

[0047] Furthermore, the output of the aforementioned uncertainty-aware attention layer is a virtual historical bidding feature vector after attention weighting. The adjusted feature representation is obtained by multiplying the original values ​​of each feature dimension by the corresponding attention weight and summing them by dimension.

[0048] Step 4: Integrate intent-driven demand estimation with supply-side baseline to output new ad space demand forecast; Based on intent matching scores, historical ad intensity characteristics of each potential advertiser on similar ad slots on other platforms are extracted. Scaling adjustments are made using the ratio of traffic volume between the new ad slot and the source platform ad slots, and the dispersion of user profile distribution, to generate estimated expected ad intensity values ​​for each potential advertiser on the new ad slot. These estimated expected ad intensity values ​​are then aggregated according to competitive game logic. The bidding process is simulated using the expected bid levels of each potential advertiser and the expected inventory of the new ad slot, estimating the total demand under equilibrium conditions as the intent-driven demand estimate. Finally, the intent-driven demand estimate is merged with the cold-start baseline demand forecast from the supply-side migration through an adaptive fusion network to output the new ad slot demand forecast value, which integrates information from both sides.

[0049] It should be noted that the above-mentioned historical ad placement intensity characteristics include average daily consumption, bid level, and ad placement frequency.

[0050] Furthermore, the above scaling adjustment linearly scales the historical ad intensity characteristics based on the traffic scale ratio between the new ad placement and the original platform ad placement, while exponentially decaying the adjustment based on the user profile distribution dispersion. The larger the user profile distribution dispersion, the greater the difference in the user group and the lower the reliability of the ad intensity migration. Therefore, the expected ad intensity estimate is reduced through the exponential decay function.

[0051] Furthermore, the above competitive game logic is based on Nash equilibrium theory. It assumes that the bidding strategies of each potential demand advertiser on the new ad slot influence each other and eventually reach an equilibrium state. In the equilibrium state, the bidding strategy of each potential demand advertiser is the optimal response to the bidding strategies of other potential demand advertisers. By iteratively solving the equilibrium bid and winning probability of each potential demand advertiser, the total demand is obtained by weighting and summing the estimated expected ad intensity of all potential demand advertisers according to their winning probabilities.

[0052] Furthermore, the simulated bidding process described above sorts potential demand advertisers according to their expected bid levels, determines the set of potential demand advertisers who can obtain display opportunities by combining the expected inventory of new ad slots, and sums up the estimated expected ad intensity of potential demand advertisers who can obtain display opportunities to obtain the total demand estimate.

[0053] It should be noted that the weights of the aforementioned adaptive fusion network are jointly determined by the size of the set of potential demand advertisers and the concentration of intent matching degree distribution.

[0054] Furthermore, the concentration of the aforementioned intent matching degree distribution is represented by the negative value of the entropy of the intent matching degree score. The smaller the entropy value, the more concentrated the intent matching degree score, the stronger the consistency of the intent signal, and the higher the concentration of the intent matching degree distribution.

[0055] Furthermore, the weights of the aforementioned adaptive fusion network are mapped to a linear combination of the logarithm of the potential demand advertiser set size and the concentration of intent matching degree distribution to a range of 0 to 1 using the sigmoid function. When the number of potential demand advertisers is large and the intent matching degree distribution is relatively concentrated, the weights of the adaptive fusion network tend to drive the demand estimate based on intent; conversely, the weights of the adaptive fusion network tend to retain the cold start baseline demand forecast value of the supply-side migration.

[0056] Furthermore, the aforementioned adaptive fusion network is a single-layer fully connected neural network. Its input consists of a concatenated vector of intent-driven demand estimates, cold-start baseline demand predictions from supply-side migration, the size of the potential demand advertiser set, and the concentration of intent matching degree distribution. The fully connected layer outputs a new ad space demand prediction that integrates information from both sides. The adaptive fusion network is trained using supervised learning, with the actual demand for the already opened ad spaces during the cold-start phase used as the label. The loss function is the mean squared error loss function.

[0057] In this embodiment of the application, in order to filter cross-platform signals with real causal effects and reduce the weight of noise signals, the following steps are included before generating the expected delivery intensity estimate: Step 4-1: Estimate the causal effect of cross-platform behavioral signals and generate a causal enhancement intent representation vector; For each potential advertiser in the potential demand advertiser set, the causal inference module is used to estimate the causal effect of each potential advertiser's cross-platform behavioral signals on the platform's advertising intention. The difference in intention prediction under the two conditions of having cross-platform behavioral signals and not having cross-platform behavioral signals is calculated through counterfactual inference. Cross-platform behavioral signals with significant causal effects are marked as strong causal signals and given enhanced weights, while cross-platform behavioral signals with insignificant causal effects are deweighted, generating a causal enhanced intention representation vector for each potential advertiser.

[0058] Furthermore, the counterfactual reasoning described above compares the intention prediction value of the causal inference module under the condition that cross-platform behavioral signals are observed with the intention prediction value under the condition that cross-platform behavioral signals are not observed, and calculates the difference between the two as the causal effect of cross-platform behavioral signals. The larger the absolute value of the difference, the stronger the causal influence of cross-platform behavioral signals on intention prediction.

[0059] Furthermore, when the absolute value of the causal effect is greater than the preset causal effect threshold, the cross-platform behavior signal is marked as a strong causal signal, and the weight of the strong causal signal is linearly enhanced according to the magnitude of the causal effect; otherwise, the weight of the cross-platform behavior signal is reduced to the preset weight reduction coefficient.

[0060] Furthermore, the aforementioned causal inference module adopts a dual-head neural network structure, including a shared feature extraction layer and two independent prediction heads. The shared feature extraction layer is a multi-layer fully connected neural network. The input of the shared feature extraction layer is the cross-platform behavior embedding vector of the advertiser, and the output of the shared feature extraction layer is a shared feature representation vector. The first prediction head takes the concatenation vector of the shared feature representation vector and the cross-platform behavior signal as input and outputs the intention prediction value under the condition of having cross-platform behavior signal. The second prediction head takes the concatenation vector of the shared feature representation vector and the zero vector as input and outputs the intention prediction value under the condition of not having cross-platform behavior signal. The training of the causal inference module adopts a supervised learning method, using the known actual placement intention of the advertiser on this platform as the label, and the loss function adopts the mean squared error loss function.

[0061] Step 5: Generate a time-dimensional demand inflow forecast curve for new ad slots; For potential advertisers whose cross-platform advertising behavior summary sequence shows an active advertising adjustment state, the causal enhancement intent representation vector of the potential advertisers and the cross-platform behavior are uniformly embedded into the sequence input heterogeneous event cross-attention encoder to predict the expected time distribution of each potential advertiser starting to advertise in the new ad slot, generating an advertising intent maturity curve; after multiplying the advertising intent maturity curve of each potential advertiser with the estimated expected advertising intensity, the curves are summarized by the time dimension to generate the time dimension demand influx prediction curve for the new ad slot.

[0062] It should be noted that the above heterogeneous event cross-attention encoder learns independent attention weight matrices for the combination of event types on this platform and cross-platform, while retaining the continuous time decay mechanism.

[0063] Furthermore, the heterogeneous event cross-attention encoder comprises the following components: an event type combination embedding layer, which encodes the combination of event types from this platform and cross-platform event types into a combination embedding vector; an independent attention weight matrix set, which learns attention weight matrices for different event type combinations; a time decay module, which calculates decay weights based on the time difference between the event occurrence time and the current time; and a cross-attention calculation unit, which calculates the final cross-attention output based on the above components.

[0064] Furthermore, the aforementioned set of independent attention weight matrices maintains independent query, key, and value transformation matrices for each event type combination, enabling different types of event combinations to learn their own specific attention patterns, thereby capturing the differentiated relationships between heterogeneous events.

[0065] Furthermore, the aforementioned time decay module uses an exponential decay function to weight the impact of historical events over time. The further away an event is from the current moment, the smaller the weight of the historical event's impact on the current intention prediction. The decay rate is controlled by the decay rate parameter.

[0066] Furthermore, the data transmission relationship of the heterogeneous event cross-attention encoder is as follows: the causal enhanced intent representation vector and the cross-platform behavior unified embedding sequence are first input into the event type combination embedding layer to generate a combination embedding vector sequence. The combination embedding vector sequence is then input into the corresponding independent attention weight matrix set according to the event type combination index for querying, key-value transformation, and then the attention score is calculated and normalized. Next, the attention score is multiplied element-wise with the decay weight output by the time decay module to obtain the time decay attention score. Finally, the value vector is weighted and summed through the cross-attention calculation unit to obtain the encoded representation vector. The outputs of all event type combinations are concatenated and then passed through the linear mapping layer to output the final encoded representation vector.

[0067] Furthermore, the output of the heterogeneous event cross-attention encoder is processed by the time distribution prediction layer, which is a multi-layer fully connected neural network. The input of the time distribution prediction layer is the encoded representation vector. The delivery probability distribution at each time point is output through the fully connected layer. The training of the time distribution prediction layer adopts a supervised learning method, using the actual delivery start time of the known advertiser in the historical ad slot as the label, and the loss function adopts the cross-entropy loss function.

[0068] Furthermore, the output of the time distribution prediction layer is normalized by softmax and converted into a probability distribution form, which serves as a directly usable deployment intention maturity curve.

[0069] It should be noted that the above-mentioned advertising intention maturity curve represents the probability distribution of potential demand advertisers starting to advertise at various points in the future. The horizontal axis represents time, and the vertical axis represents the advertising probability density or cumulative probability at that point in time.

[0070] Furthermore, the above-mentioned time-dimensional demand inflow prediction curve is generated as follows: for each future time point, the probability of all potential demand advertisers placing orders at that time point is multiplied by their estimated expected order intensity, and then the sum is obtained for all potential demand advertisers to obtain the expected total demand at that time point. The expected total demand at all time points is then connected in chronological order to form the time-dimensional demand inflow prediction curve.

[0071] Step 6: Make online adjustments based on actual bidding data and output demand forecast results with a smooth transition; As new ad slots accumulate actual bidding data, the system utilizes the hit rate of the actual bidders' identities and the potential demand list, along with the deviation between the actual bidding volume and the predicted demand value of the new ad slots, which is obtained by integrating information from both sides. Through online learning methods, the system adjusts the adaptive threshold of the intent matching score and the weights of the adaptive fusion network. The system gradually reduces the weights of virtual historical bidding features and intent-driven demand estimates in feature calculations according to a preset decay coefficient until they are completely replaced by actual bidding data. The system outputs the new ad slot demand prediction results that smoothly transition from the cold start stage to the regular prediction stage.

[0072] It should be noted that the above hit rate is calculated as follows: count how many of the advertisers who actually participated in the bidding were predicted to be in the list of potential demand advertisers, and calculate the proportion of the number of advertisers who were predicted to be hit to the total number of advertisers who actually participated in the bidding.

[0073] Furthermore, the above deviation is calculated as follows: calculate the absolute value of the difference between the actual bidding volume and the predicted demand value of the new ad slots that integrates information from both sides, and then divide it by the actual bidding volume to obtain the relative deviation percentage.

[0074] It should be noted that the above decay coefficient is applied as follows: as the opening time of new ad slots increases, the weights of virtual historical bidding features and intent-driven demand estimates are gradually reduced through an exponential decay function, while the weights of actual bidding data are increased accordingly. When the opening time is long enough, the weights of virtual historical bidding features and intent-driven demand estimates approach zero, and demand forecasting depends entirely on actual bidding data.

[0075] Furthermore, the above online learning method employs the gradient descent algorithm to construct a loss function based on the hit rate and bias. It updates the adaptive threshold of the intent matching score and the parameters of the adaptive fusion network through backpropagation, so that the predicted value of the new ad space demand, which integrates information from both sides, gradually approaches the actual observed value.

[0076] Furthermore, in the above-mentioned smooth transition process, the proportion of the currently accumulated actual bidding data to the preset sufficient data threshold is calculated in each time window as the actual data maturity. The actual data maturity is used as the benchmark value of the actual bidding data weight, and the weights of virtual historical bidding features and intention-driven demand estimates are set as the complement of the actual data maturity, ensuring that the sum of the weights of the two types of data is always 1, thereby achieving a smooth transition of weights.

[0077] This implementation uses a cross-platform semantic alignment network to project heterogeneous multi-platform advertising behavior data into a unified semantic space and calculate intent matching scores. This supplements the cold start prediction of new ad slots with demand-side advertiser intent signals. Since the cross-platform semantic alignment network uses a shared projection layer to eliminate the semantic heterogeneity between behavioral data from different platforms, it can accurately calculate the degree of matching between advertiser intent and new ad slot features in a unified semantic space. This overcomes the problem that pure supply-side feature transfer methods lack predictive basis at the advertiser demand level.

[0078] This implementation estimates the causal effect of cross-platform behavioral signals on the platform's delivery intention through a causal inference module. Cross-platform behavioral signals with real causal effects are given enhanced weights, while cross-platform behavioral signals with insignificant causal effects are deweighted. Since the causal inference module distinguishes between cross-platform behavioral signals that truly affect delivery intention and noise signals that only have correlation through counterfactual reasoning, it can effectively suppress noise interference introduced when cross-platform heterogeneous data is directly fused.

[0079] This implementation uses an uncertainty-aware attention layer to process virtual historical bidding features, assigning lower weights to feature dimensions with higher uncertainty. Since the uncertainty label vector is determined based on the variance of the corresponding feature dimension values ​​in the source ad set, it can quantify the reliability of each dimension of virtual historical bidding features. Therefore, it avoids the excessive influence of the uncertainty of virtual historical bidding features on the cold start baseline demand forecast value of supply-side migration.

[0080] This implementation method organically combines the intent-driven demand estimate with the cold-start baseline demand forecast of supply-side migration through an adaptive fusion network. The weights of the adaptive fusion network are jointly determined by the size of the potential demand advertiser set and the concentration of intent matching degree distribution. Since the adaptive fusion network can dynamically adjust the contribution ratio of the two estimates according to the sufficiency of the intent signal, it overcomes the defect of incomplete information dimension when using either method alone.

[0081] This implementation generates a campaign intent maturity curve through a heterogeneous event cross-attention encoder, predicting the expected time distribution when advertisers with potential demand will begin campaigning. Since the heterogeneous event cross-attention encoder learns independent attention weight matrices for different event types and retains a time decay mechanism, it can capture heterogeneous correlation patterns between cross-platform behavioral events and campaign timing, thus improving the accuracy of the time dimension demand influx prediction curve during the cold start phase.

[0082] This implementation method uses an online learning approach to adjust the adaptive threshold of intent matching score and the weights of the adaptive fusion network based on actual bidding data. The weights of virtual historical bidding features are gradually reduced according to the decay coefficient until they are completely replaced by actual bidding data. Since this gradual replacement mechanism avoids abrupt switching from virtual historical bidding features to actual bidding data, it ensures the continuity of demand forecasting from the cold start stage to the regular forecasting stage.

[0083] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

[0084] On March 15, 20XX, a major video platform opened a new advertising slot at the top of the short video feed (ad slot number). This ad placement is located on the homepage of the platform's newly launched "Lifestyle" channel and has the following initial characteristics: the ad placement type is a native in-feed ad, the media app category is a short video social platform, and the page location code is [missing information]. The expected visibility index is 0.89. Traffic sampling during the initial 72 hours of operation showed that the main user group was aged 25-35 (42%), female users accounted for 63%, and users from first- and second-tier cities accounted for 71%. User interest tags were concentrated in three categories: beauty, fashion, and home decor. Traffic distribution showed a typical bimodal pattern, with peak traffic occurring between 12:00-13:30 and 20:00-22:30.

[0085] The platform needs to predict demand for the next 7 days to set a reasonable bid floor for the ad slot, given that there is no historical bidding data available for that ad slot. It also needs to recommend the ad slot to potential advertisers. Through a data cooperation agreement, the platform obtained summary data on the publicly available advertising behavior of authorized advertisers on three other mainstream advertising platforms (e-commerce platforms, news platforms, and social media platforms), including changes in these advertisers' activity levels, product categories, and campaign scale over the past 30 days.

[0086] The platform's data collection system first extracts new ad slots. The metadata features were analyzed, and traffic data within a 72-hour sampling period were statistically analyzed. Categorical features in the metadata (ad placement type, media category, page location) were transformed through unique encoding, while numerical features (visibility metrics) underwent Z-score standardization. User profile distribution vectors were generated by statistically analyzing the proportion of users across various dimensions, and traffic time-period distribution curves were calculated using hourly average traffic values.

[0087] Table 1. Original data of new ad slot metadata characteristics:

[0088] Table 2: Results of generating comprehensive feature vectors for new ad slots:

[0089] Meanwhile, the platform obtained summaries of the advertising behavior of 1,847 authorized advertisers on other platforms through cross-platform data interfaces. This information was then used to analyze the advertising behavior of one typical advertiser. (A certain beauty brand) The system collected its cross-platform behavioral events over the past 30 days.

[0090] Table 3: Sample sequence of advertiser cross-platform campaign behavior summary:

[0091] The system will advertisers Each event in the cross-platform behavior summary sequence is standardized and encoded. The platform identifier, behavior type, category tag, and scale level are each mapped to a 64-dimensional vector through an independent embedding layer. The four embedding vectors are concatenated to form a 256-dimensional event embedding vector. The sequence of the four event embedding vectors is then input into the shared projection layer of the cross-platform semantic alignment network for processing.

[0092] Shared projection layer via projection matrix and bias vector The event embedding vectors are mapped from 256 dimensions to a 128-dimensional unified semantic space:

[0093] For advertisers The four events are projected to obtain four 128-dimensional aligned vectors, and then average pooling is used to obtain the advertiser projection vector. The new ad placement's comprehensive feature vector (93 dimensions) is also generated using a dedicated projection matrix. Projecting onto a 128-dimensional semantic space yields the new ad placement projection vector. .

[0094] Table 4. Intent matching results after semantic alignment (partial advertiser sample):

[0095] The formula for calculating the intent match score is:

[0096] For advertisers Substitute the numerical values ​​into the calculation:

[0097] The system statistically analyzes the intent matching score distribution of all 1847 advertisers and calculates the mean. Standard deviation The adaptive threshold is set to A total of 327 advertisers with an intent match score higher than 0.728 were selected, forming a list of potential advertisers.

[0098] The system calculates the number of ad slots from the platform's existing 1,523 ad slots and the new ad slots. Multi-dimensional similarity. For a typical source ad placement... (The advertisement is located in the information feed of the food channel), and the system calculates the similarity of each dimension.

[0099] Table 5. Multi-dimensional similarity calculation between the new ad placement and the original ad placement:

[0100] User profile distribution similarity was calculated using Jensen-Shannon divergence for the source ad placement. :

[0101] in Calculations yielded The negative value is used as the similarity score. After normalization and mapping to the 0-1 interval, it becomes 0.858.

[0102] The system selects the top-ranked systems based on overall similarity. The source ad slots are taken as a set of source ad slots, and the historical bidding characteristics of these source ad slots are weighted and averaged according to similarity weights.

[0103] Table 6. Composite data of virtual historical bidding features:

[0104] Synthetic calculation of virtual historical bidding features:

[0105] For the average number of daily bids:

[0106] The system further calculates the virtual substitution elasticity coefficient, source ad placement. The average elasticity of substitution coefficient with other ad placements is 0.34, and the weighted average virtual elasticity of substitution coefficient for the new ad placement is 0.31. The uncertainty label vector is calculated based on the variance of the feature dimension values ​​corresponding to the 10 source ad placements, and the variance of the average daily number of bids is... The variance of the average bid is The larger the variance, the higher the uncertainty.

[0107] The uncertainty-aware attention layer calculates attention weights for each feature dimension, with the temperature parameter set to... :

[0108] After attention weighting, the effective weight of the average number of daily bids decreased from 1.0 to 0.73, and the weight of the average bid decreased from 1.0 to 0.89, generating an adjusted supply-side migration baseline demand forecast of 6523 times / day.

[0109] Table 7. Estimated Intensity of Potential Demand Advertisers' Expected Spending (Partial Sample):

[0110] The system targets potential advertisers. Extract its data from e-commerce platforms Historical ad placement intensity characteristics of similar information flow ad slots: average daily expenditure of 2,850 yuan, average bid of 3.62 yuan, and continuous daily placement frequency.

[0111] New ad space The expected daily average traffic is 128,000 impressions for e-commerce platforms. The corresponding source ad placement has an average daily traffic of 185,000 impressions, and the traffic scale ratio is:

[0112] The user profile distribution divergence (JS divergence) is calculated to be 0.167, and the exponential decay adjustment factor is:

[0113] Calculation of adjustments to the estimated expected intensity of investment:

[0114] The system simulates the bidding process based on each advertiser's expected bid level (migrated from historical data) and the expected inventory of new ad slots (128,000 impressions per day). Assuming the ad slots use a second-price sealed-bid auction mechanism, the total demand is calculated by summing the expected ad intensity of the first 128,000 impressions after sorting by expected bids. Through Nash equilibrium iteration and considering the competitive game among advertisers, the final intention-driven demand estimate is 8923 impressions per day.

[0115] Adaptive fusion network based on the size of the potential demand advertiser set The concentration of the intent match distribution is used to calculate the fusion weights. The entropy of the intent match score is calculated as follows:

[0116] The concentration is 0.68 after normalizing the negative value of entropy. The fusion weights are calculated using the sigmoid function.

[0117] Final demand forecast for new ad placements integrating information from both sides:

[0118] The system targets potential advertisers. Causal effect estimation was performed on cross-platform behavioral signals. The advertiser's "e-commerce platform scale and level upgrade" event on March 5, 20XX was identified as a key signal. The dual-head neural network of the causal inference module calculated the intent prediction value under the conditions of having this signal and not having this signal.

[0119] Table 8. Estimation results of causal effects of cross-platform behavioral signals:

[0120] The shared feature extraction layer maps the advertiser's cross-platform behavior embedding vector (256-dimensional) to a shared feature representation vector (128-dimensional). The first prediction head takes as input the concatenated vector (192-dimensional) of the shared feature representation vector and the scale-level boosting event embedding (64-dimensional), and outputs an intent prediction value of 0.823 under signal conditions. The second prediction head takes as input the concatenated vector of the shared feature representation vector and the zero vector (64-dimensional), and outputs an intent prediction value of 0.615 under no-signal conditions.

[0121] Causal effect calculation:

[0122] The causal effect threshold is set to 0.15. This signal's causal effect of 0.208 exceeds the threshold and is therefore labeled as a strong causal signal. The weighting enhancement coefficients are linearly mapped according to the magnitude of the causal effect.

[0123] For another event, "Category Change of Information Platform", the calculated causal effect was 0.087, which did not exceed the threshold, and the weight was reduced to the preset weight reduction coefficient of 0.6.

[0124] The causal enhanced intent representation vector is generated by re-averaging and pooling the event vectors of the original cross-platform behavior unified embedding sequence by multiplying them by their corresponding causal weights. The causal enhancement intention representation vector improved the values ​​of each dimension by about 52% compared to the original vector in the dimension corresponding to the strong causal signal.

[0125] The system targets potential advertisers in an active campaign adjustment state (189 out of 327 are selected) and embeds their causal enhanced intent representation vectors and cross-platform behaviors into a sequence input heterogeneous event cross-attention encoder.

[0126] For advertisers Its most recent cross-platform event was on March 10, 20XX, on a social media platform. "Start Execution" – The event type combination embedding layer encodes the events (new ad slot opening on this platform, cross-platform ad promotion start execution based on deep learning) into a combined embedding vector. The corresponding independent attention weight matrix is... Transform the query, key, and value vectors respectively.

[0127] The time decay module calculates that the time difference between this event and the current time (March 15, 20XX) is 5 days, and the decay rate parameter is... The attenuation weight is:

[0128] The cross-attention calculation unit performs a weighted summation of the outputs of all event type combinations, which is then passed through a linear mapping layer to obtain an encoded representation vector (128 dimensions). The temporal distribution prediction layer (a three-layer fully connected network) maps the encoded representation vector to the delivery probability distribution for the next 14 days.

[0129] Table 9. Prediction Results of Advertiser Intent Maturity Curve (Advertisers) ):

[0130] The system multiplies the campaign intent maturity curves of all 189 active advertisers by their estimated expected campaign intensity, and then summarizes the results by time. For March 19, 20XX (4 days in the future), the calculation method is as follows:

[0131]

[0132] Table 10: Time-based demand inflow forecast curves:

[0133] The generated time-dimensional demand inflow prediction curve shows that the demand for new ad slots peaks on the 4th-5th day after opening, and then gradually stabilizes, which is consistent with the typical inflow pattern of ad slot cold start.

[0134] New ad space On the third day after the opening (March 18, 20XX), actual bidding data was accumulated. A total of 78 advertisers participated in the bidding that day, of which 62 were predicted from the list of potential demand advertisers. Hit rate calculation:

[0135] The actual number of bids on that day was 7456, while the demand forecast based on both sides of the information was 8276, with a relative deviation of:

[0136] Table 11 Online parameter adjustment update process:

[0137] The system adjusts parameters through online learning methods, and the loss function is constructed by combining hit rate and bias:

[0138] The adaptive threshold for updating the intent matching score using the gradient descent algorithm was adjusted from 0.728 to 0.715, and the weight of the intent-driven branch in the adaptive fusion network was adjusted from 0.73 to 0.69.

[0139] The open period has been extended to the 3rd day, with an actual bidding data volume of 21,834 times (cumulative over 3 days). The preset sufficient data volume threshold is 100,000 times. Actual data maturity:

[0140] Virtual historical bidding feature weights are calculated using a decay coefficient. Decreasing daily, the weight on day 3 is:

[0141] The weight of actual bidding data will be increased accordingly. The demand forecast has been updated to:

[0142] With the opening period extended to the 14th day, the actual data maturity reached 0.856, and the weight of the virtual historical bidding feature decayed to 0.065. The system achieved a smooth transition from the cold start stage to the regular prediction stage.

Claims

1. A method for predicting advertising promotion demand based on deep learning, characterized in that, Includes the following steps: Obtain metadata features and traffic sampling data for the new ad placement, and encode them into a comprehensive feature vector for the new ad placement; obtain summary data of each advertiser's placement behavior on other ad platforms through cross-platform data interfaces, and generate a cross-platform placement behavior summary sequence for each advertiser; The cross-platform campaign behavior summary sequence of each advertiser is standardized and encoded, and then processed by a cross-platform semantic alignment network to generate a unified cross-platform behavior embedding sequence. The intent matching score between the projection vector of each advertiser and the projection vector of the new ad space is calculated, and advertisers with intent matching scores exceeding an adaptive threshold are selected to generate a list of potential demand advertisers. Calculate the multi-dimensional similarity between the comprehensive feature vector of the new ad slot and the comprehensive feature vector of each existing ad slot. Select the existing ad slots with the highest similarity ranking as the source ad slot set. Combine the historical bidding features of each source ad slot in the source ad slot set according to the similarity weight to synthesize virtual historical bidding features and generate the cold start baseline demand prediction value. Based on the intent matching score, the historical delivery intensity features of each potential demand advertiser are extracted and scaled to generate an expected delivery intensity estimate. The expected delivery intensity estimates are then aggregated to obtain the intent-driven demand estimate. An adaptive fusion network is used to merge the intent-driven demand estimate with the cold start baseline demand prediction to output the new ad space demand prediction. The cross-platform semantic alignment network is a shared projection layer that embeds and projects the behaviors of different platforms into a unified semantic space.

2. The method according to claim 1, characterized in that, The metadata features include ad placement type, media app category, page location code, and expected visibility index; the traffic sampling data includes user profile distribution vector and traffic time period distribution curve; the cross-platform advertising behavior summary data includes changes in ad activity status, advertising category change records, and advertising scale level changes on each platform.

3. The method according to claim 1, characterized in that, The training of the cross-platform semantic alignment network adopts a supervised learning approach, using advertisers known to have placed ads on the platform as positive samples and advertisers not placed ads on the platform as negative samples. The input is a concatenated vector of cross-platform placement behavior summary sequences of each advertiser after processing by the embedding layer. The output is a binary classification prediction result of the placement intent through a fully connected layer, and the loss function adopts the binary cross-entropy loss function.

4. The method according to claim 1, characterized in that, The calculation of the multi-dimensional similarity includes: category matching score for ad type and media category dimensions; negative Jensen-Shannon divergence for user profile distribution dimension; dynamic time warping algorithm for traffic time period distribution dimension; and weighted fusion of the similarity of each dimension through the learned dimension weights.

5. The method according to claim 1, characterized in that, Also includes: An uncertainty label vector is introduced into the virtual historical bidding features. The values ​​of each dimension of the uncertainty label vector are determined based on the variance of the corresponding feature dimension values ​​in the source ad slot set. After processing by the uncertainty perception attention layer, the attention weight is calculated based on the uncertainty variance of each feature dimension, and the feature dimensions with high uncertainty are assigned low weights.

6. The method according to claim 1, characterized in that, Before generating the expected delivery intensity estimate, the process also includes: using the causal inference module to estimate the causal effect of cross-platform behavioral signals of each potential demand advertiser on the delivery intent of this platform; calculating the difference in intent prediction under the two conditions of having cross-platform behavioral signals and not having cross-platform behavioral signals through counterfactual reasoning; marking cross-platform behavioral signals with significant causal effects as strong causal signals and assigning them enhanced weights; reducing the weights of cross-platform behavioral signals with insignificant causal effects; and generating a causal enhanced intent representation vector for each potential demand advertiser.

7. The method according to claim 6, characterized in that, The causal inference module adopts a dual-head neural network structure, including a shared feature extraction layer and two independent prediction heads. The first prediction head takes into input a concatenated vector of the shared feature representation vector and the cross-platform behavior signal and outputs the intention prediction value under the condition of having cross-platform behavior signal. The second prediction head takes into input a concatenated vector of the shared feature representation vector and the zero vector and outputs the intention prediction value under the condition of not having cross-platform behavior signal.

8. The method according to claim 6, characterized in that, Also includes: The causal enhancement intent representation vector of potential demand advertisers and cross-platform behavior are uniformly embedded into the sequence input heterogeneous event cross-attention encoder to predict the expected time distribution of each potential demand advertiser's start of advertising, generating an advertising intent maturity curve; the advertising intent maturity curve of each potential demand advertiser is multiplied by the estimated value of expected advertising intensity and then summarized by time dimension to generate a time dimension demand influx prediction curve for new ad slots; the heterogeneous event cross-attention encoder learns an independent attention weight matrix by combining event types of this platform and cross-platform and retains a continuous time decay mechanism.

9. The method according to claim 1, characterized in that, Also includes: As new ad slots accumulate actual bidding data, the system utilizes the hit rate of the actual bidders' identities and the potential demand list, as well as the deviation between the actual bidding volume and the predicted demand for the new ad slots. Through online learning methods, it adjusts the adaptive threshold of intent matching scores and the weights of the adaptive fusion network. The system gradually reduces the weights of virtual historical bidding features and intent-driven demand estimates in feature calculations according to a preset decay coefficient until they are completely replaced by actual bidding data. This outputs demand prediction results that smoothly transition from the cold start phase to the regular prediction phase.

10. A deep learning-based advertising demand prediction system, used to execute the method according to any one of claims 1-9, characterized in that, include: The feature acquisition module is used to acquire metadata features and traffic sampling data of the new ad slot and encode them into a comprehensive feature vector of the new ad slot, as well as to acquire the cross-platform delivery behavior summary sequence of each advertiser through the cross-platform data interface; The intent matching module is used to standardize and encode the cross-platform advertising behavior summary sequence, calculate the intent matching score after processing by the cross-platform semantic alignment network, and filter and generate a list of potential advertisers. The supply-side migration module is used to calculate the multi-dimensional similarity between new ad slots and existing ad slots, select the source ad slot set and synthesize virtual historical bidding features to generate cold start baseline demand forecast values; The demand fusion module is used to generate an estimated value of expected delivery intensity based on the intent matching score and to summarize the intent-driven demand estimate. The intent-driven demand estimate is then merged with the cold start baseline demand prediction value through an adaptive fusion network to output the new ad space demand prediction value.

Images